identification and classification of coating defects using...
TRANSCRIPT
IDENTIFICATION AND CLASSIFICATION OF
COATING DEFECTS USING MACHINE VISION
A THESIS
submitted by
RIBY ABRAHAM BOBY
for the award of the degree
of
MASTER OF SCIENCE
(By Research)
MANUFACTURING ENGINEERING DEPARTMENT OF MECHANICAL ENGINEERING INDIAN INSTITUTE OF TECHNOLOGY MADRAS
OCTOBER 2010
ii
And God said, "Let there be LIGHT," and there was LIGHT.” Genesis 1:3
iii
Dedicated to the one, who loved me and endured the cross even when I was an enemy to him.
iv
THESIS CERTIFICATE
This is to certify that the thesis entitled “IDENTIFICATION AND
CLASSIFICATION OF COATING DEFECTS USING MACHINE VISION”
submitted by RIBY ABRAHAM BOBY to the Indian Institute of Technology
Madras for the award of the degree of Master of Science (By Research) is a bonafide
record of research work carried out by him under our supervision. The contents of this
thesis, in full or in parts, have not been submitted to any other Institute or University
for award of any degree or diploma.
Chennai-600 036 Research Guides
Date:
Prof. B. RAMAMOORTHY
Prof. M. SINGAPERUMAL
v
ACKNOWLEDGEMENT
Life and research in particular had never been a bed of roses and surviving this ordeal
is due to the bountiful grace of the Lord and Savior Jesus Christ who is closer to me
than a friend and a brother. The times spent reading the scripture had helped to
refocus and reinstate the fact that ‘hope is a good thing, perhaps the best of things’.
The guidance and faith of Prof. B Ramamoorthy had kept the ship sailing even in the
toughest of storms. He has given me ample liberty thus motivating me to give the best
in all situations. In spite of his varied responsibilities, it was never impossible to break
in and stir up a discussion with him. Prof. M. Singaperumal could even use a few
moments at the parking lot to give a great idea that is worth hours of arduous
searching of literature. Here I could see excellence and experience personified. I am
truly indebted to them for the guidance and mentoring.
The laboratory was a family away from home and especially Srikanth and
Rahamatullah always had a friendly air about them. Srinivas, Arunachalam, Vinoth
Jose, Mohan, Denis Ashok, Murugarajan, Giridhar, Periasamay, Bala, Kanthababu
were not just research colleagues but fellow comrades of a great ‘team’ who even
ventured outdoor with some bouts of cricket on weekends. Srikanth, Sekhar Reddy
and Janardhan have given me significant suggestions regarding research and career.
Thippesamy and Rajsekhar, have been a source of a lot of activity and life and gave a
breather during our monotonous laboratory schedule. Santhosh, Uppu Srinivas Rao,
Deepak Lawrence and Binu helped me with their invaluable suggestions regarding
research and life at IIT.I also thank Praveen Kumar and Prashant S. Sonakar had
helped me significantly during the early days of my research.
IITCF was another big family I had in IIT and we spent a lot of good times singing
songs and worshipping God. It supplemented as a place to learn about helping one
vi
another and thus living loving God and people. Srikar, Revanth, Srinivas, Praveen
Kumar, Eldhose, Balamurugan, Oliver has all been source of so much warmth. The
members of IITCF including Chethan, Srinath, Ebenezer, Christopher, Ronald are
people whom one will like to stay close to till the very end. Prof. David Koilpillai,
Prof. G.L.Samuel, Prof. Koti, Prof. George, Kokila akka, Anu akka and other elders in
IITCF had been real mentors who showed the way to live, bearing fruits which will be
useful to society as a whole. I also feel indebted to the friends and pastor at NLAG
Chennai for the great time we had.
The experience in Munich was as varied as the -200C of Bavarian winter and the 420C
of Chennai summer. It was divine grace that I could meet Thorbjörn Buck, Ana Peréz,
Mathias Müller, Thomas Bodendorfer, Daniel Dorigo, Florian Hirth, Benjamin and
Markin Jakobi. Their company always inspired me to get back to work after the
slumber during Bavarian weekends. I had really nice time singing Deutsche Gospel
hymns with Xi, Daniel Baumann, Korina, Sedoni, Jana and Prisca and would like to
relive those moments.
It was during the last year that a beautiful damsel made entry into my life and life was
never the same after that. Life itself has found a new course of sailing once she has
come into my life. Also I am indebted to my father mother and brother who were so
close even understanding my thoughts even before I spoke about them. They were
always there in all situations and were a constant source of encouragement and
comfort. I convey my gratitude to the former HOD Prof. M. S. Shanmugam and
current HOD Prof. S. P. Venkateshan for their inspiration and assistance. I thank Mr P
Sreenivasulu, Mr Sridhar, Mrs J E Malarvizhi Alice, Mrs A Vasantha Kumari and
Mrs B. Sundari for their assistance in all administration related issues. I extend my
warmest thanks to Mrs. Thamara Dineshan, Mrs. Saraswathi, Mr. Muthiyal, Mr.
Paranjyothi, Mr. Radhakrishnan, and Mrs. Raniamma for helping in all the
vii
proceedings in the laboratory. I also thank Prof. K. Balakrishnan of Machine design,
Prof. Chandrasekhar of Computer science, Prof. Kothiyal of Applied optics, Prof.
Baburaj of Applied Mechanics, Prof. N Ramesh Babu, Prof. Shanmugam and Prof. L
Vijayaraghavan for their assistance at various stages of my work. I also express
special gratitude to Lakshmi Machine Works Coimbatore India for supplying the ring
samples for inspection and for their hospitality. Though space, time constraint renders
personal expression of gratitude impossible, there have been innumerable people who
have contributed to my life during stay at IIT Madras and TU München and I am
really happy that I could meet them all and I thank each of them for leaving a mark in
my life.
Riby A. B.
viii
ABSTRACT
Keywords: machine vision, inspection, coating defects, classification, dark field, Gray
Level Co-occurence Matrix, image series
Textile spinning uses spindle ring machines in which large number of chrome coated
ring components is used. The rings are made of Bearing Steel and then Chrome
coating is done to improve the life of the rings while in use. After coating, the rings
are to be inspected for coating defects which are visually done in the current set up.
The automation of the process has been attempted but with no significant success.
Given the reflective nature of the surface and the inner intricate profiles, it is really
challenging to build a set up that can successfully detect all the coating defects. To
solve the inspection problem, first an attempt was made using bright field illumination
set up based on which two different algorithms for defect detection and region
segmentation have been attempted. One was based on Fast Fourier Transform (FFT)
filtering, which was previously used for defect detection on textured surfaces. To
augment the performance, an auto-median operation based on morphology was
appended to the FFT filtering algorithm. The new algorithm was more accurate than
the one which is currently used in the industry. From the segmented defective region
in the image, many geometrical parameters can be identified. The geometrical
parameters of defect that are useful for the classification have been identified. But, as
many non defective samples were triggering false alarms, it was necessary to image
the surface with greater contrast.
Dark field illumination is a technique of getting very high contrast images by using
illumination with light falling at grazing angles to the surface under inspection. This
method of illumination helped in imaging the defective area with higher contrast. The
histogram had a bimodal distribution, which implies that lesser computation will be
ix
sufficient for segmenting out the defects. To detect the presence of defects, a single
step thresholding method was used. Later, a better thresholding was done using the
Fractional Change in Derivative (FCD) method which is proposed in this work. The
dark field illumination was found to be more accurate and faster than the bright field
illumination based approaches. It is thus possible to detect even smaller sized defects
on the ring surface using higher magnification.
The dark field approach which is based on scattering of light might lead to loss of
some information about the defective regions in the image. As a result, a new
approach based on different illumination sources was used. This helped in avoiding
the loss of information through the scattering of light at the defective surface and has
been used to highlight discontinuities. A series of images have been captured and
were used to find the Gray Level Co-occurrence Matrix (GLCM) parameters for each
of the pixels under consideration but across the series. A window was transposed over
the pixel and the GLCM matrix for the same was obtained. Later the thresholding
operation was done using the GLCM parameters and the thresholded regions were
used to train a Bayesian classifier. All these steps used an information fusion
approach, where results from different approaches and sources are combined together
to obtain the final conclusion for classification of defects. The pattern recognition
machine developed in this manner can classify each of the pixels of the input image
series as a particular defect class. Thus it is possible to detect the presence of defects
and then to classify the rings according to the type of defects. Therefore the proposed
inspection set up described in this work consists of a machine vision system using
dark field illumination technique coupled with multiple illumination sources. The
associated algorithms developed are aimed at detecting the presence of defects and the
subsequent classification of the rings with the speed and accuracy required by the
industry.
x
TABLE OF CONTENTS
Title Page No. ACKNOWLEDGEMENTS……………………………………………...... v
ABSTRACT……………………………………………………………....... viii
LIST OF TABLES…………………………………………………….…... xiii
LIST OF FIGURES…………………………………………………….…. xiv
ABBREVIATIONS…………………………………………………….….. xviii
MATHEMATICAL NOTATIONS............................................................. xix
CHAPTER 1 INTRODUCTION
1.1 Quality control……………………………………………….…. 1
1.2 Automated visual inspection…………………………………… 2
1.2.1 Basic steps to build a machine vision set up…………………… 3
1.3 Pattern recognition using machine vision………......................... 4
1.3.1 Sensing…………………………………………………………. 4
1.3.2 Segmentation and grouping…………………………………….. 5
1.3.3 Feature extraction………………………………………………. 5
1.3.4 Classification…………………………………………………… 6
1.3.5 Post processing…………………………………………………. 6
1.4 Objectives and scope of the present work……………………… 7
1.5 Organization of the thesis…………………………………....…. 7
CHAPTER 2 LITERATURE REVIEW
2.1 Introduction................……………………………….................. 9
2.1.1 Inspection of reflective and curved surfaces……….................... 9
2.1.2 Set up for inspection of reflective and curved surfaces................ 11
xi
TABLE OF CONTENTS (Contd...) Page No. 2.1.3 Illumination set up…………………………………................... 13
2.1.4 Defect classification…………………………………................ 19
2.2 Summary……………………………………….……................ 20
CHAPTER 3 PROBLEM DEFINITION
3.1 Introduction………………………………………..................... 22
3.2 Ring spinning.............................................................................. 22
3.3 Ring specifications………………………………….................. 24
3.4 Ring inspection system………………………………................ 25
3.5 Motivation for the image processing algorithms………............. 28
3.6 Summary .........................…………………………................... 30
CHAPTER 4 IMAGING AND ILLUMINATION SYSTEM
4.1 Introduction................................................................................. 32
4.2 Bright field illumination………………………..................…… 33
4.2.1 Axial illumination system…………………………................... 36
4.3 Dark field illumination………………………………................ 41
4.3.1 Implementation of dark field illumination.................................. 50
4.4 Summary…………………………………………..................... 52
CHAPTER 5 DEFECT IDENTIFICATION
5.1 Introduction................................................................................. 54
5.2 Bright field illumination…………………………….................. 54
5.2.1 FFT filtering based on critical radius……………….................. 55
5.2.2 Auto-median based approach in image segmentation................. 58
5.3 Dark field illumination………………………………................ 59
5.3.1 Fractional change in derivative (FCD) method for finding optimum threshold value.............................................................
62
5.4 Defect classifier using bright field illumination images............. 63
xii
TABLE OF CONTENTS (Contd..) Page No.
5.4.1 Classification parameter selection…………………................... 64
5.5 Results and discussions……………………………................... 65
5.5.1 Image processing algorithms for bright field images.................. 65
5.5.2 Image processing algorithm for dark field images….................. 70
5.6 Summary…………………………………………..................... 78
CHAPTER 6 DEFECT CLASSIFICATION USING MULTIPLE IMAGES
6.1 Introduction 80
6.2 Multiple imaging set up…………………………................…... 82
6.3 Gray Level Co-occurrence Matrix (GLCM) approach for multiple images...........................................................................
84
6.4 Image masks for each region of the ring…………................…. 92
6.5 Image thresholding operation…………………................…….. 93
6.6 Defect classifier………………………………................……... 98
6.6.1 Bayesian classifier…………………………................………... 100
6.7 Results and discussions……………………................………... 107
6.8 Summary…………………………………................………..... 109
CHAPTER 7 CONCLUSION AND SCOPE FOR FUTURE WORK
7.1 Conclusion………….………………………................……….. 111
7.2 Scope for future work………………………..………................ 112
REFERENCES.……………………………………...…………………… 117
LIST OF PUBLICATIONS……………………………………………… 121
xiii
LIST OF TABLES
Table No. Table caption Page No.
5.1 Performance comparison of algorithms.......................................... 72
5.2 Comparative performance with respect to FFT approach on bright field images.................................................................................... 78
5.3 Time of inspection per ring............................................................ 78
6.1 Classification based on discriminant (g(x)) values.......................... 109
6.2 Decision making by scoring for different parameters corresponding to pitting defect..............................................................................
105
xiv
LIST OF FIGURES
Fig. No. Title Page No.
1.1 Standard steps in the pattern recognition system............................. 4
1.2 Steps in classifier design.................................................................. 6
2.1 Variation in application of illumination techniques according to the surface reflectivity and shape................................................. 18
3.1 Textile ring component............................................................... 23
3.2 Details of ring component........................................................... 26
3.3 Images obtained using a machine vision system representing the various defects on ring components............................................ 26
3.4 Camera positions in automated defect detection set up used in industry........................................................................................ 28
3.5 Non-homogenous texture of Chrome coated surfaces seen at 100x magnification................................................................................... 29
3.6
Steps involved in the adopted methodology................................... 31
4.1 Lambertian surface showing the incident light being scattered in all possible directions....................................................................... 33
4.2 Reflection of light on specular surface............................................ 33
4.3 Experimental setup used for imaging of components using a vision system.................................................................................... 37
4.4 Images showing non defective surfaces under bright field illumination.................................................................................. 37
4.5 Images showing deep line under bright field illumination............. 38
4.6 Images showing yellow stain under bright field illumination......... 38
4.7 Images showing blister under bright field illumination.................. 39
4.8 Images showing built up under bright field illumination............... 39
4.9 Images showing damage under bright field illumination................ 40
xv
List of Figures (Contd...) Page No.
4.10 Images showing peel off under bright field illumination................ 40
4.11 Images showing pitting under bright field illumination.................. 41
4.12 Images showing rough finish under bright field illumination......... 41
4.13 Principle of dark field illumination.............................................. 42
4.14 Setup for inspection of ring outer diameter using dark field illumination...................................................................................... 45
4.15 Images showing non defective surface under dark field illumination................................................................................. 46
4.16 Images showing blister under dark field illumination..................... 46
4.17 Images showing deep line under dark field illumination................. 47
4.18 Images showing damage under dark field illumination................... 47
4.19 Images showing rough finish under dark field illumination............ 47
4.20 Images showing pitting under dark field illumination..................... 48
4.21 Images showing built up under dark field illumination.................. 48
4.22 Images showing peel off under dark field illumination................... 49
4.23 Images showing yellow stain under dark field illumination............ 49
5.1 FFT filtering................................................................................... 57
5.2 Image characteristics for non defective surface............................ 60
5.3 Image characteristics for defective surface..................................... 60
5.4 Effect of using different threshold values.................................... 61
5.5 Conditional probability density functions for blister, built up and damage classes............................................................................. 65
5.6 Image segmentation using FFT approach to highlight the defective region in bright field image........................................... 67
5.7 Failure of the FFT method on bright field images of non defective surface.......................................................................................... 67
5.8 Image segmentation using auto-median approach on bright field image............................................................................................ 68
xvi
List of Figures (Contd...) Page No.
5.9 Performance of FFT and Auto-median approach on images of defective surface of ring............................................................... 69
5.10 Performance of FFT and auto-median approach on images of non defective surface of ring............................................................... 70
5.11 Image of blister in different illumination systems........................... 72
5.12 Image of peel off in different illumination systems......................... 73
5.13 Single step thresholding for images of blister............................... 73
5.14 Single step thresholding for images of deep line.......................... 73
5.15 Single step thresholding for images of heavy blister.................... 74
5.16 Single step thresholding for images of built up............................ 74
5.17 Single step thresholding for images of rough finish..................... 74
5.18 Single step thresholding for images of pitting............................. 75
5.19 Single step thresholding for images of damage........................... 75
5.20 Single step thresholding for images of non defective surface...... 75
6.1 Experimental set up for illumination with 3 illumination sources.. 83
6.2 Effect of variation in illumination position..................................... 83
6.3 GLCM computation..................................................................... 84
6.4 Image fusion for calculation of GLCM parameters......................... 90
6.5 Gray scale representations of GLCM parameters............................ 91
6.6 Masks used for different regions of the image of ring surface........ 93
6.7 Thresholding operation.................................................................... 95
6.8 Threshold images considering different parameters........................ 96
6.9 Information fusion for thresholding operation............................. 97
6.10 Plot for Sum average showing the separation of defects using normal distribution values........................................................... 98
6.11 Plots showing the sum of squares against sum average with the clusters of defects well separated out........................................... 99
xvii
List of Figures (Contd...) Page No.
6.12 Working of a classifier.................................................................... 101
6.13 Bayesian decision boundaries for defect classes in the scatter plot of sum average against difference variance.................................. 104
6.14 Illustration of information fusion in pixel classification............... 105
6.15 Thresholded images with defective areas colored........................ 106
6.16 Flowchart for the defect classification approach............................. 109
xviii
ABBREVIATIONS
CCD Charge Coupled Device
DFT Discrete Fourier Transform
FFT Fast Fourier Transform
GLCM Gray Level Co-occurence Matrix
ID Inner diameter
LED Light emitting diode
OD Outer diameter
RAM Random Access Memory
xix
MATHEMATICAL NOTATIONS
English Symbols
c Number of classes considered for classification
CΔ(g1,g2) The entry at (g1,g2) of the Gray Level Co-occurrence Matrix with
neighbor condition defined as Δ=( Δx, Δy, Δθ).
d Dimension of x
F(u,v) Fast Fourier Transform of f(x,y)
f(x,y) A function in x and y. Used here as representing pixel values for the
index i=x and j=y.
g Gray scale value of the image
g(x) Discriminant function
g(x,y,θ) Gray scale value at index (x,y) in the image taken with illumination
position θ
h(g) One dimensional array representing the histogram values of the
image
I Two dimensional matrix representing image gray scale values
j √ 1
Li Represents the instance of the ith illumination being switched on
ln(x) Natural logarithm of x
M Two dimensional matrix representing structuring element in image
morphology
N(x) Normal distribution function
nCm Combinations of selecting m out of n samples. nCm!
!
Ng Number of distinct gray scale levels. It can be a maximum of 256.
xx
Nx Number of pixels in horizontal direction
Ny Number of pixels in vertical direction
p(i,j) Entry at (i,j)th address of Gray Level Co-occurence Matrix
P(u,v) Power spectrum of F(u,v)
p(x) Probability density function for x
p(x, ωi) Probability density function in two variables x and ωi
p(x/ωi) Class conditional probability density function of x given class ωi.
Also called likelihood.
P(ωi) Prior probability for event ωi to occur
P(ωi/x) Posterior probability
px(i) Array formed by addition of all the column elements of p(i,j)
py(j) Array formed by addition of all the row elements of p(i,j)
x Parameter vector given as input to the classifier defined as
[x1,x2....xd]T
Greek Symbols
δ Kronecker symbol. Becomes equal to one when a=b else 0
µ Mean
σ Standard deviation
Σ Covariance matrix for x given by Σ(i,j)=cov(xi,xj)=E[(xi- µi)(xj- µj)]
where µi=E(xi) is the expected value of i th element of x.
ωi ith category of defect
CHAPTER 1
INTRODUCTION
1.1 QUALITY CONTROL
Quality is pivotal to the successful functioning of an enterprise as it can be explained
as the ability of the products or services to satisfy the requirement of the customers,
rather than just conforming to the standards. Quality planning, quality control and
quality improvement are suggested by Juran et al. (1998) as the trilogy which ensures
quality in the modern day manufacturing centres. For an effective decision at all these
stages (in line with Garbage In Garbage Out-GIGO- concept), there needs to be a
measure of the current status of the quality in an accurate and quick way. A viable
option is to use sensors for such measurements of the critical parameters of the
finished product. There is another important decision to be made as to the degree of
inspection that is required for any product. Given the situation of progressive
improvement of the manufacturing process, the inspection has to be flexible enough
to accommodate newer quality standards.
When the process capability is to be improved due to sudden changes in the quality
specification, there is a requirement for 100% inspection. 100% inspection has its
disadvantages, since the cost involved is higher and additionally the human operators
also have to be educated to make the right decision. But the availability of
inexpensive sensors, higher computer memory storage capabilities and calculation
speeds makes it economical to implement automated inspection methodology in the
current production environment.
2
Though sensors are available for automating many of the inspection tasks, there are
still many situations where manual inspection is the best possible solution. When it
comes to visual inspection of varied kinds of defects in complex surfaces which are
sometimes hidden to the view completely, manual inspection still may be considered
as a suitable option.
Accounting for all shortcomings, it has been found that manual inspection has only
80% accuracy in making correct inspections (Juran et al., 1994). As a result there is
always an attempt to augment the senses of the inspector using instruments like sound
amplifier, optical magnifier etc. The development of an application specific inspection
system is the ultimate form of this. Manual inspection can be replaced by automatic
inspection to reduce labour fatigue, and save money and valuable man hours wherever
possible. This helps in quality improvement, cost reduction, increased volume, and
shorter cycle time.
1.2 AUTOMATED VISUAL INSPECTION
Currently in industries, inspection of components is automated to a very large extent.
In this context, optical inspection always has been a tricky problem to solve since
each situation needs a unique approach. Machine vision is a form of non-contact
inspection, which can be used for a wide range of electro optical sensing techniques
from triangulation, profiling, 3D object recognition to bin picking based on image
processing routines. In addition, it is being used for simple detection and measuring
tasks to robotic control. The most common applications are measuring critical
dimensions, detecting flaws, counting/sorting, assembly verification, position
3
analysis, character or bar-code reading/verification, and determination of
presence/absence of features on small parts. The major advantage is avoidance of
human error. The process involves capturing an image of the object to be inspected
using vision sensors and then processing it using image processing algorithms to
measure the features of interest. One very important advantage of machine vision
systems is that 100% inspection can be achieved.
1.2.1 Basic Steps to Build a Machine Vision Set Up
The machine vision approach to solve optical inspection problems can be done in
three stages. The first is to set up a camera and proper illumination to capture images
with very good quality required for any specific application. Generally the usage of
vision inspection is dictated by the application and it should be always aimed to make
it simple. Complex image processing algorithms that are used in many automatic
inspection systems could create a burden on the machine memory and delay the
detection process. By careful selection of the light source, illumination intensity and
its positioning, the effectiveness of inspection can be ensured. It is more cost effective
and better engineering practice to capture a good image at the source, rather than to
pour resources into cleaning it up or simplifying it later. The second stage is the pre
processing of images to gather the necessary quality for the application desired and to
obtain the image data. Later, the data has to be processed and then the decision
regarding the inspection status has to be made. The available information has to be
used to make a decision regarding the type of defects, since no approach aimed at
defect detection will be complete without quantifying the decision. The final step is an
4
elaborate process in itself because decision making is a case of pattern recognition
and involves a lot of preliminary steps to ensure correct decisions.
1.3 PATTERN RECOGNITION USING MACHINE VISION
The defects identification and classification problem using machine vision application
can be modelled as a pattern recognition problem, since the underlying aim is to make
classification based on the input image. The steps involved are shown in Figure 1.1
(Duda et al., 2001).
Figure 1.1 Standard steps in the pattern recognition system (Duda et al., 2001)
1.3.1 Sensing
The inputs to the pattern recognition system are the images of the component to be
inspected. For this purpose, a Charge Coupled Device (CCD) camera coupled to an
image grabbing system is generally used. A suitable lighting system and fixtures to
hold the ring is also a part of the set up. The orientation of the camera and lighting
system are made in an optimum fashion such that the images are clear, better quality
and of high contrast. Also the optics should have sufficient magnification and
Sensing Segmentation Feature extraction Classification Post processing
Input
Analysis & decision
5
working distance. If intrinsic surfaces are to be imaged, then the orientation should be
proper to enable easy imaging of each surface.
1.3.2 Segmentation and Grouping
The output from the sensor is digital images, which consist of small picture elements
or pixels having a particular range of gray scale values. This value is proportional to
the amount of light that is incident on the physical pixel element on the machine
vision sensor after reflection of the incident light from the surface or object under
inspection. The final image is a matrix having indices of the pixels and their gray
scale values. If the image is in colour, the matrix will have the third dimension as the
different gray scale values corresponding to Red, Green and Blue (RGB). But the
feature to be inspected is among other information and cannot be easily separated out.
For this purpose, the region of interest has to be separated out from the rest of the
region where measurements are not made and this process is image segmentation. The
algorithms for image segmentation are highly dependent on the type of images taken.
It varies from one situation to another but can be a set of independent image
processing operations which are done one after another to get the required results. In
most of the defect detection set ups, the algorithm should reveal the defective area of
the surface imaged, if they exist.
1.3.3 Feature Extraction
The results from segmentation will have the gray scale values and indices of the
defective area of the image and the details are supplied to the feature extraction
6
algorithm. The features may be many, depending upon either the gray scale values or
a combination of both or even just the index values. Many statistical and geometrical
features can be defined for the area that is under consideration. The best choices are
the ones which will be useful in making a decision at the next step, which is the
classification. While designing the pattern recognition system, there should always be
a feedback from the higher levels to the lower levels of the system.
1.3.4 Classification
The feature vectors should initially be used to find the presence of defects, which can
also be done at the segmentation stage. The defect classes where the probable
defective component may belong to, have to be clearly defined. Also other
information regarding the peculiarities or statistics of occurrence of defects has to be
procured. The probability of the defect belonging to a particular defect class is found
out and the one with the greatest probability is considered as the final defect type.
Figure 1.2 presents an idea about the typical classifier design.
Figure 1.2 Steps in classifier design (Duda et al., 2001).
1.3.5 Post Processing
The ultimate aim is to separate out the defective products into respective bins. The
output from the classifier can be supplied to mechanical actuators, for example, to
Collect data Choose features Choose model Train classifier Evaluate classifier
Start End
7
make the relevant sorting. Also at this stage, the success of the whole system can be
evaluated and used for subsequent improvements in the classifier.
1.4 OBJECTIVES AND SCOPE OF THE PRESENT WORK
To identify and to classify the various coating defects present in textile ring
components using a machine vision system. The methodology followed to achieve
this is as follows
1. To use multiple cameras at various locations and at specified angles.
2. To use different methodologies of illumination namely bright field, axial and
dark field. Then to choose the best one for the specific application.
3. To develop image processing algorithms which are fast and accurate for every
kind of illumination set up used and compare their performances.
1.5 ORGANIZATION OF THE THESIS
The thesis is organized as follows. The need for automated visual inspection and the
methodology for solving pattern recognition problem using machine vision has
already been presented in Chapter 1. Chapter 2 sketches a brief literature survey
describing the different applications where machine vision has been used for defect
detection on curved and reflective surfaces. Different methodology for classification
of defects using images also has been discussed in this chapter. The description and
basic information regarding the ring inspection problem is given in Chapter 3. In
Chapter 4, different kinds of illumination systems that are used for capturing the
images in the present work have been explained. Different algorithms to enable fast
8
and accurate defect region segmentation on the images captured using the imaging set
up is discussed in Chapter 5. In Chapter 6 an approach to classify defects based on
information from multiple images is presented. Also explained in this chapter are the
GLCM approach of parameter calculation for image series and the thresholding
operation using series of images. The summary of work done and the conclusion and
scope for future work are presented in the final Chapter.
CHAPTER 2
LITERATURE REVIEW
2.1 INTRODUCTION
There have been many attempts to inspect manufactured components for surface
defects using machine vision. Most of these works have been on the area of
development of image processing algorithms since the decision about the component
quality is made on the basis of the image and not directly based on the real object.
Many of them were based on FFT based filtering since it is an easy and fool proof
method of separating out defective regions from images of the component to be
inspected. Illumination has been studied by industrial users of machine vision and the
knowledge has been put to good use in quick and accurate inspection. Pattern
recognition algorithms are useful in the final classification of defects and form the
final stage of the solution for defect identification and classification problems.
2.1.1 Inspection of Reflective and Curved Surfaces
Fourier domain operations are a useful way of visualizing frequency variation in
signals. Two dimensional Discrete Fourier Transform (2D DFT) has been
successfully applied in the separation of lay patterns of surface textures in many
applications. Bayerer and León (1997, 1998) have used a similar method for the
separation of background and lay pattern of surface images. León (2002) has
successfully applied it in industry for the detection of honing angle in cylinder
engines. The approach involved the separation of background and lay pattern. The lay
10
pattern is visible as straight lines in the intensity plot of the 2D DFT of the image. The
angle between these lines which is the actual angle of honing in the cylinders can be
found out by Hough transform. The problem is similar in the sense that the inner
surface of the cylinder is reflective.
Tsai and Huang (1999, 2003) used Fourier image reconstruction for the detection of
defects on engineered surfaces. They had dealt with statistical textures like that of
sandpaper and have clearly demonstrated the selection of cut-off radius for such
surfaces. Later the same algorithm was used on flat sputtered surfaces which are non
homogenous and non-statistical texture. The inspection problem has a prominent
resemblance to the one discussed in this thesis, albeit curvature of the inspection
surface.
Zhang et al. (2006) attempted defect detection on ground and polished surfaces. The
approach involved extraction of the image features which were then used to classify
the defects into the respective classes. The defect classes were well defined and the
accuracy of the system to classify these defects has been discussed.
Wiltschi et al. (2000) devised a method of finding steel quality from the microscopic
images of etched and polished steel specimens. The process was automated with a
very significant accuracy. The steel carbide distribution images were taken and then
the carbide regions were thresholded based on the second derivative of the histogram
values. The shape and size of the segmented surface area was determined and the
most valid parameters were selected as input to the classifier. Finally classification
was done using a minimum Euclidean distance classifier. The process has been
11
implemented in the steel manufacturing industry and is currently used for routine
inspection.
2.1.2 Set Up for Inspection of Reflective and Curved Surfaces
Rosati et al. (2009) have attempted real time defect detection on curved reflective
surfaces using a set of mirrors. A special mirror was designed to direct the
illumination at specific direction and the presence of defects could be observed by
analysing the images using standard image processing software. Bright field
illumination approach was used, where the machine vision sensor was in the path of
specularly reflected light rays. A similar set up can be used for inspection of
automobile parts with highly reflective surfaces. But the high intensity reflections
might saturate the machine vision sensor resulting in loss of information and this was
not discussed in detail. To avoid saturation, reduction in the intensity of incident light
might cause lesser contrast in the defects that are imaged. This is a particular anomaly
while using bright field imaging systems.
Sun et al. (2005) applied X-ray imaging for real time detection of welding defects in
steel tubes. Though X-ray was used, the signals were captured in digital form just like
visible light and the image was used for subsequent analysis. The use of X-rays and
image processing paradigms could easily automate the inspection process which was
undertaken by human inspectors till then. The feed back after the implementation of
the set up was that it is more efficient than manual inspection of video streams. This is
an example of capturing higher contrast image information by use of X-rays.
12
Moreover, in this particular case, choices for inspection set ups were limited since the
defects were subsurface.
Luo and Liou (1998) used multiple cameras to measure the surface conditions of a
cylinder and used the images to measure deviations. The methodology was based on
binocular vision. The images captured by the two imaging devices have been
correlated with each other to determine the diameter of the sample to be inspected.
Both the imaging devices were used to capture the same surface under inspection. Use
of two cameras helped in the evaluation of depth information, though there were
anomalies due to the curvature of the surface. The explanation is particularly useful
for determining the diameter of cylindrical components, rather than surface
inspection.
Microstructure surfaces were reconstructed from SEM images by Samak et al. (2007).
Unlike the previously discussed approach, the sample to be inspected was tilted and
the change in position of the point under observation was converted into height
information using the standard triangulation algorithm used for most of the depth
detection problems. Later De launey’s triangulation has been done using this
information. The images were microscopic and the CAD model of many broken
materials was successfully created such that fracture analysis can be done using this
information.
Abrahamovich et al. (2005) described the calibration and use of multiple cameras for
inspection. Quite interestingly, the cameras used were commercially available
webcams, array of cameras such that imaging of a large area can be covered. Many of
13
the inspection scenarios like wooden sheet inspection, sheet metal inspection demand
multiple cameras and the performance, as stated by the author, is really convincing for
an industrial user since the web camera costs 1/50th of an industrial camera. The
methodology of imaging different areas of the same surface and then stitching all of it
together will enhance the speed of inspection provided the illumination source for all
the surfaces also remains the same. But the approach is really beneficial when the
surface for inspection is planar (which the author has specifically stated) and can be
imaged at once.
2.1.3 Illumination Set Up
Microscopic defect detection on mirror surfaces using laser beams has been
investigated by Porteus et al. (1986). The mirror surfaces that get damaged due to
laser illumination were inspected for defects based on the scatter from the surface of a
beam of laser. The system was developed for analyzing a very small surface area.
Though the technique of illumination was for detecting presence of laser initiated
defects, a similar scattering technique can be used with a machine vision sensor for
inspection of a larger surface area of highly reflective surfaces. Additionally, the use
of polarized beams can reveal the presence of defects and enable easier separation of
background light from saturating the machine vision sensor. The approach reveals that
optical methods also will aid in the detection of defective surfaces and this
observation was used again in the detection of laser damage in mirrors by Marrs and
Porteus (1985). In this case, a source of pulsed laser was used to illuminate the
surface to be inspected and the scattered light was received into a video microscope
system. The illumination direction was normal to the surface to be inspected and a
14
polarizer was used in front of the video microscope to enable reception of more
details regarding the light scattered by defective surfaces. The set up was used to
obtain the physical images of the surfaces with defects shown at higher contrast.
Khalili and Webb (2007) used multiple wavelength illumination technique for defect
detection on hemispherical surfaces. The surface to be reflected was rivet surface
which is spherical and metallic, thus reflective. Such surfaces always make a standard
illumination difficult since the light is received at varying illumination angles at the
different surfaces of the spherical rivet. Added to it, the problem was an online
inspection system, further compounding the difficulty in deciding upon a single
illumination system. The solution was use of multiple illumination sources, but that
meant, usage of a particular image for a given illumination source. Later two lights of
different wavelength were used to illuminate the surface in unison while using a
colour camera to process different intensities of the corresponding imaged surface
from single image, which enabled quicker inspection. They used comparison of the
two images to get the final image, free from specular reflection and shadows. Later a
correction was done to accommodate for linear variation of light because of the off
centric position of the imaging source. Image of a rivet can be captured using a single
image and the defects can be identified to enable automatic inspection. This is a very
good example where illumination variation in position has been used to successfully
detect the presence of defects with higher speed in an online inspection environment.
Nayar et al. (1991) have discussed, in detail, the geometrical reflectance model which
does not use the electromagnetic theory of propagation of light and the physical
reflectance models which consider the electromagnetic theory of propagation of light.
15
When the physical features under observation have dimensions comparable to the
wavelength of light, physical optics are valid. Geometrical optics is a generalized case
of physical reflectance model when the wavelength is much lesser in comparison to
the surface dimensional features. On the basis of Beckmann-Spizzochino (Beckmann
et al., 1963) and Torrence-Sparrow (Torrance et al., 1967) models, any illuminated
surface will produce a specular lobe, specular peak and a diffuse lobe. The incident
luminous intensity on the sensor, thus, can be written as a sum of specular reflection
lobe component, diffuse reflection lobe component and specular reflection peak
component. Prior knowledge about the surface under inspection will help in finalizing
the reflectance model to be implemented, because in any given application one of the
reflecting natures is supposed to dominate. According to the surface characteristics of
reflection, the best approach can be selected and used to solve the machine vision
problem.
Later, a similar approach was used to image dielectric substances, but the method
used was polarisation and colour separation (Nayar et al., 1997). The highlights due to
specular reflection were masked out using polarizing filters. Colour intensities of the
image were also considered to capture images with better details. The approach
worked well for many applications involving non metallic surfaces. In the case of
non-metallic surfaces, all the reflections from the inspected object will be completely
masked out by a polarising filter, thus leading to loss of information.
Geometrical measurements and object verification can be done very efficiently if
images are of very high contrast. This was studied by Yi et al. (1995) by varying the
position of the sensor and the illumination source. The best contrast images will have
16
the least variance across the edges in the image and have the maximum possibility of
being detected. It is obvious that the capture of high contrast images will avoid the
requirement of complex image processing routines, which are an imposition on the
speed, accuracy and development cost of the machine vision system. Software has
been developed to find the optimum position of sensor and the illumination source,
given the geometry of the object, photometric property of the object, object edges
involved in inspection, measurement type and light source. Torrence Sparrow model
of reflection has been used to develop the aforesaid approach. Though the approach
sheds significant amount of light into the positioning of the sensors and illumination
source in the pursuit of images of high quality, it was suitable for simple edge
detection problems where the edge could be easily represented by simple geometrical
curves. The experimental data on dimensional measurement of cubic shaped
specimens also reinstates the feasibility of the approach in getting more accurate
information about the imaged region.
For practical applications, there are a wide variety of illumination systems that are
dependent on the condition of surface under inspection and feature to be inspected.
Since most of the vision inspection systems attempt to undertake the inspection
process in minimum time and with maximum accuracy, the best type of illumination
and the best position for illumination has to be determined. Connolly (2002) has given
a description of the application specific illumination systems and explained the
concept of dark field illumination, which is used to highlight scratches and other
surface irregularities in metals. The light source is kept at an angle in such a way that
there will be a low angle of incidence to the object surface to be inspected. If the
object is flat, the light emerges in the dark field angular region and thus misses the
17
sensor giving a dark appearance to the object surface. This is not the case if there are
some scratches. The specular light reflects back into the sensor region from the angled
side of the scratch, thus appearing bright. Dark field illumination can be used behind
and around small transparent objects to highlight their edges against a bright
background and this is used for inspection of syringes and in microscopy to examine
cells in medical applications. In short, the dark field illumination enhances the
contrast between the specular and non specular areas of the object, and this can be
exploited to a very good extent in the inspection of metallic surfaces.
Dark field illumination is predominantly used in many applications, especially
microscopy. In a perfectly clear, transparent medium, the field of vision will remain
black, but in the presence of refractive boundary or scattering surface or particles,
light is reflected into the optical axis and is visible, just as dust is visible in sunbeam.
The key advantage is the avoidance of flooding of the focal plane by specular
reflections. It is also used for macro specimen photography like revealing the fine
structure in the cataract of the eyes, lenses, jellyfish, amoeba etc. (David, 2010).
For machine vision applications the lighting condition has been the subject of
considerable importance as it has wide and varied inspection applications. The
lighting systems have a major influence in achieving the final result/goal for machine
vision applications. Martin (2007) explained the commercially available and more
popular lighting systems in detail. The illustration regarding the application of
different kinds of illumination is given in Figure 2.1. An interesting observation is
that when the surface becomes reflective and curved, the inspection problem becomes
more complex. Dome illumination or diffuse illumination, which is the recommended
18
methodology, is possible for easy to illuminate surfaces like spherical surfaces and
not for internal features.
Figure 2.1 Variation in application of illumination techniques according to the surface reflectivity and shape (Martin, 2007)
Images with dark field features have been used in many applications and one of it is in
the process of lace cutting (Bamforth et al., 2007). The illumination using ring light
gave a higher contrast image of the transparent lace boundary. The approach could
reveal more details about the lace boundary and thus enabled precision laser cutting of
the lace pattern. Miyoshia et al. (2001) has used dark field illumination systems in
microscopy of semiconductor surfaces. Noda et al. (2008) has boasted of sub-
Matte
Mixed
Mirror Specular
Flat Uneven Topography
Curved
Bright Field
Dark Field
Diffuse Dome
Geometry independent area
19
nanometre measurements on biological specimens, using dark field imaging set up
utilizing axicons. Piper (2009) has attempted to create dark field images using
scattered light but directed normally rather than oblique illumination. The aim was to
increase the contrast of the specimen images in microscopy. Metallography is another
application of oblique illumination using laser source but in this case with the
production of speckles. The investigation by Povolo (1997) compared the results with
that from partial dark field illumination. All these applications reiterate the fact that
the contrast enhancement by oblique illumination can be used for inspection of
metallic surfaces.
2.1.4 Defect Classification
The development of efficient defect detection systems will not be successful without
an accompanying classification system. The classification problem, which is the final
stage of defect detection, involves the selection of the most relevant parameters as
well as the associated classifying algorithm.
Iivarinen et al. (1998) explained the selection of features for classification from the
segmented images of defects and used Gray Level Co-occurrence Matrix (GLCM)
and the related parameters from it for giving input to the classifier.
Guise et al. (2002) discussed the defect classification for semiconductor surfaces. The
classification was done using the shape of the defects using the shape based
parameters or geometrical parameters. The shape description parameter, especially
perimeter by area ratio (P/A), has been used in cartography by Salas et al. (2003).
20
Most of the development in the area of textural studies was promoted by the need for
remote sensing and cartography. Pioneering work was done by Haralick et al. (1973),
in which the Gray Level Co-occurrence Matrix was introduced and fourteen texture
parameters based on it were used for quantification. These parameters were used in
many applications involving surface characterization, though the primary intent
behind the introduction of them was to use them for analyzing remote sensing images.
These parameters have been used in the classification of directional textures
(Gadelmawla, 2004).
2.2 SUMMARY
There have been several attempts previously to detect defects on metallic surfaces.
But the approach of each researcher was unique, given the situation. Most of the
applications involved microscopy, which is not necessarily 100% inspection and thus
slower. Also, the components under inspection were either specially prepared surfaces
or else planar with minimum reflection. 100% inspection on metallic surfaces with
curvatures demanded special set ups, which are normally custom made according to
the applications. Most of them cannot be generalized to other applications but the
approach used gives a faint idea of how to solve a similar inspection problem. In this
work, machine vision system has been custom made according to the inspection
requirement which is to sense the ring defects according to the need at industry.
The lighting system, which is an inherent and important part of the machine vision
system, has been given the least significance for many years. This casts the entire
21
burden on the algorithms for defect detection. But many studies have proved beyond
doubt, that an optimum positioning of the sensor and the illumination system will
enable capture of very high contrast images and thus simplify the image processing
procedures by quickening it and minimizing the storage requirements.
Any defect detection algorithm has to incorporate a defect classification system.
There are many parameters based on geometry and texture that can be used but the
selection of the best one will enable accurate and quicker defect classification. The
steps involved have been briefly illustrated in Figure 1.2. The information regarding
the defect classes has to be collected. Defect class features to be used for
classification needs to be selected. After this, most suitable model of classifier has to
be chosen. Later the classifier is trained and then used for testing the data at the later
stages in the industry. Any shortfall in performance requires the redesigning of
intermediate steps until the industrial requirement is satisfied.
Based on these observations the following methodology has been adopted in this
work. Initially the maximum contrast image is captured by utilizing different
illumination systems. The relevant algorithms to process the images captured using
each of these illuminations have been developed. The results of the algorithms were
analyzed for accuracy and speed. Later the classification algorithm was developed
using image series based on an image fusion approach. A GLCM based approach has
been used to process the images for defect classification.
CHAPTER 3
PROBLEM DEFINITION
3.1 INTRODUCTION
The most important parts of a textile spinning machine are the spindles and rings
which are used in large numbers in each of the machines. As a result, thousands of
rings are manufactured at the industry to cater to the needs of the newer machines and
for the maintenance of older ones. The set up for ring spinning consist of a traveller
running on a ring and it is described in the following section.
3.2 RING SPINNING
There are four main methods of textile spinning, namely, ring spinning, rotor
spinning, air-jet spinning and friction spinning (Jayavarthanavelu, 2010). Of these,
ring spinning has the advantage of providing higher quality yarn that can be used in
many types of textile end products. Though the other approaches are new
developments which allow higher rates of production, the yarn that is produced
through these methodologies is of lower quality. This restricts their use and is the
reason behind the widespread use of ring spinning. The only technological limitation
for ring spinning lies in the quality of the rings that are used. This makes the ring as
the most vital part of the machinery.
The spinning process involves converting the un-spun fibres to yarn. In ring spinning,
the fibres are passed through rollers to compact them into a single yarn. Later, it
passes through the traveller and then, it is wound to the cob. The cobs are attached to
23
spindles which run at a maximum speed of 20,000 rpm. The traveller revolves around
the ring slower than this, but still the speed is considerable enough to cause significant
wear in the traveller. The traveller wears out more frequently than the ring and the
amount of wear is dependent on the overall quality and in particular the surface finish
of the ring surface. As a result, the rings are made with high carbon high chromium
ball bearing steel and are given a low friction coating. The low friction coating allows
the traveller to run at 3 m/s higher speed than the uncoated rings. This enhances the
productivity and at the same time, reduces the expenditure due to the wear in the
travellers. The spinning machine is called ring frame and has multiple (at least
600/machine) spindle ring combination system on both sides. It is clear from the name
itself that the ring traveller is a principal component of the whole machinery. The ring
assembly is shown in the Figure 3.1b.
(a) (b)
Figure 3.1 Textile ring component (a) CAD drawing for textile ring component, (b) Ring and traveller assembly in ring frame
24
These rings are produced at a very large scale owing to the fact that the ring is
subjected to constant wear and needs replacement frequently. Also, the functional
region of the ring has to be maintained at a high level of quality.
3.3 RING SPECIFICATIONS
A typical photograph of a ring component and its cross section details are shown in
Figure 3.2. Due to their critical role, they are subjected to inspection for various
possible defects in order to maintain a high quality, as explained later in this section.
The rings are manufactured in a CNC machine and then the chrome coating is done.
Hard chromium coating is highly reflective. It is an example of a non homogenous
texture. The thickness of coating varies from 5 μm to 10 μm. The outside diameter of
ring is around 45 mm and the inside diameter 38 mm (Figure 3.2c). Later, the rings
are subjected to an inspection process normally using conventional equipments, such
as profile projector, tool makers microscope etc., to identify the coating defects. The
frictional wear is dependent on the surface finish of the ring. For non defective
surfaces, the surface finish value should be approximately 0.1 μm for minimum
friction and frictional wear. It has been observed that the defective surfaces will have
higher surface finish values of the order of 0.5 μm. Such regions are visible under the
microscope on careful observation. Defects can be considered as those kinds of
features which deem the component to be unusable by the customers. The types of
possible defects (Figure 3.3) on the component surface are given below.
1. Pitting - Small pits in the coating leading to corrosion.
2. Damage - Dents of different shapes and sizes.
25
3. Deep line - Marks similar to feed mark which continue around the whole ring.
4. Built up - Bright elevated deposits. This is a more common defect. It can be
reprocessed before use to salvage the work piece.
5. Peel off - Small patches or spots showing small discontinuities, which are
normally visible only on careful manual inspection and are likely to be present
both on inner and outer surfaces.
6. Rough finish - Elevated features occurring as clusters, which are similar to
peel off, usually found at the inner and outer edges.
7. Blisters - Small porous surfaces on the coating, usually larger in spread as
compared to other defects.
3.4 RING INSPECTION SYSTEM
There have been efforts in the industry to attain lesser or even nil inspection
requirement which is not generally practical. Normally, for high volume applications,
inspection methodology consists of a sampling plan. This involves selective
inspection and is suitable for applications where the process is really capable of
producing components having consistent quality. This is however used at an advanced
stage of quality improvement process by which the manufacturing process has been
improved. This is normally done based on careful evaluation of process capability
based on the output from the inspection process over a long period. So, the quality
control department interacts with the production departments in the industry, making
a closed cycle of interaction and improvement of the process. But, any change in the
26
product specification will warrant a full repetition of the whole process from the
starting point.
Figure 3.2 Details of ring component: (a) Ring (b) Cross section of the ring showing defect prone area (c) Cross section details of the ring with dimensions
(d) (e) (f)
Figure 3.3 Images obtained using a machine vision system representing the various defects on ring components: (a) Pitting (b) Damages and deep lines (c) Built up (d) Peel off (e) Blister (f) Rough finish
In the present inspection problem, inspection criteria have been strengthened for
increasing the quality of the end product supplied to customers. This is a normal
development in any given industry due to a host of parameters like competitor pricing,
27
competitor quality, customer feedback etc. As a result, initially for a long duration,
the inspection was carried out manually by inspectors using magnification 10X
microscopes. Since the volume of production is very high (of the order of a few
hundreds of thousands of components every month), it presents a huge challenge and
a considerable amount of strain on the quality control department. There is also much
incompetence arising from using human inspectors for a long duration as discussed in
Chapter 1. The immediate solution was to switch over to an automated inspection
system.
An automated machine vision inspection system exists in the industry for this
purpose. The set up included five different cameras oriented at five different
directions to image and to analyze each of the defect prone areas of the ring (Figure
3.4). A gripper was used to bring the ring to be inspected into the inspection area and
images were taken after rotating the ring. The images were analyzed using LabView
software and based on the results; the rings were sorted into respective bins. The type
of illumination used was bright field and it was implemented using two fluorescent
ring lamps inside the enclosure of the machine vision system. The inspection time
required for one ring was 13 seconds. After switching over to an automated vision
inspection, it was observed that the accuracy of detection was a maximum of 60%
(i.e. If 100 defective samples classified as defective by the human inspector was
supplied to the system, only 60 were correctly detected as defective) and was
inconsistent too (i.e. accuracy used to be slightly higher if all non defective rings were
supplied).
In the light of the present situation, the industry wanted to completely automate the
system making it still faster and foolproof, in addition to avoiding human involvement
28
in the whole process of inspection. The rings have to be imaged using a machine
vision system and the corresponding algorithms need to be developed, tested and
analyzed for all the types of defects mentioned in the earlier section.
Figure 3.4 Camera positions in automated defect detection set up used in industry
3.5 MOTIVATION FOR THE IMAGE PROCESSING ALGORITHMS
The coating defects are really difficult to detect using a microscope. Only an
experienced technician can correctly identify the defects, which is really strenuous in
the long run. The major challenges in the inspection of these components are the
highly reflective nature of Chrome coated surfaces (Rosati et al., 2009; Khalili and
Webb, 2007; Pfeifer and Wiegers, 1998), the curvature of the surface to be inspected
(Rosati et al., 2009; Lee et al., 2000; Aluze et al., 2002) and also the inaccessibility of
some inner surfaces. Unlike directional textured surfaces, homogenous statistical
features pose considerable challenge when it comes to the detection of defects. If
there is a presence of a definite lay pattern, there can be a standard procedure to
segment out the particular pattern and then analyze the continuity of it. Discontinuity
will deem the sample to be defective. Most of the machining processes like turning
Camera 1
Camera 2
Camera 3
Camera 4
Camera 5
29
and milling produce such patterns of definite lay pattern. In a homogenous statistical
texture like that of sandpaper, the statistical features are more consistent all through
and this characteristic can be utilized to successfully ascertain the quality of the
surface (though the approach may not be fully foolproof using the digital images). It
is not so in the case of non homogenous features. Most of the coating processes
produce non homogenous texture (Figure 3.5).
(a) (b) (c)
Figure 3.5 Non-homogenous texture of Chrome coated surfaces seen at 100x magnification
The curvature of the surface might give bands of illumination which might interfere
with the image details and often gives an impression of false defects. This problem is
compounded by the fact that the surface is highly reflective owing to the low friction
surface coating. Inner fillets with varying direction of curvature make it really
difficult to image larger surface area, avoiding the specular reflections to the machine
vision sensor. The requirement to inspect the inner diameter surface makes it
necessary to overcome the space constraint for imaging and illumination source. As a
result, it is very difficult to image the surface with the defects in higher contrast. It
was impossible to use a single illumination source for the whole ring since the
reflective and curved surfaces used to reflect the light in different directions, thus
making inter-reflections on the surface. Though many standard software modules are
30
available to inspect defects on flat surfaces and non-reflective surfaces, none of them
are useful for inspecting internal, curved and highly reflective surfaces. Also, a less
computationally intensive algorithm is required to speed up the entire inspection
process.
3.6 SUMMARY
The inspection problem is difficult to solve, owing to the condition of surfaces to be
inspected particularly that of the ring components considered in this study. Special
nature of curved and reflective surfaces warrants a newer approach for achieving
accurate and speedy inspection. The adopted approach in this work is briefly given
below.
The first effort was made to image the surface using a standard illumination source,
which was a ring fluorescent lamp. The resulting images are called bright field
images. The images were then segmented using different algorithms. A Fourier
transform method has been implemented in an attempt to achieve this (Tsai et al.,
1999). To increase the accuracy of defect detection, a method combining a Fourier
filtering method and an auto-median operation has been proposed in the second
algorithm. Since the algorithms were slow and not consistently accurate, a different
illumination methodology was attempted.
The illumination source was used to direct illumination at a glazing angle to the
surface under inspection to achieve the dark field illumination. Another approach
which involves single step thresholding of the dark field image helps in increasing the
speed and accuracy of detection. This approach is based on an optimum positioning of
31
camera and illumination source (Pfeifer and Weigers, 1999). All the algorithms were
designed so that they will work for images taken from any of the four camera
directions, given the kind of illumination. Another major deciding factor is that the
inspection of a single ring has to be completed in 8 seconds with accuracy as per the
conditions set by the company. So, the algorithms have to be optimized for speed. All
the defects stated earlier in Section 3.3 have to be detected using the algorithms.
The classification of rings into categories of defects was attempted using multiple
illumination sources. Gray Level Co-occurrence Matrix (GLCM) was used to find the
parameters relevant for accurate classification and then an algorithm for thresholding
and classifying the defective region in the image was developed. It was finally
possible to detect the presence of defects and then to classify them, enabling sorting
of the defective rings with considerable accuracy and speed as required by the
industry. To achieve this result, various steps and procedures followed in this work
are shown in Figure 3.6. More details and explanations of the same are presented in
the following chapter.
Figure 3.6 Steps involved in the adopted methodology
MACHINE VISION SET
UP Camera set up
Illumination set up
•Bright Field
•Axial
•Dark Field
•Comparison
Optics
Frame grabber
Software
IMAGE PROCESSING
ALG
ORITHMS Bright Field
•FFT approach
Automedian
Dark Field
•Single step thresholding
•FCD method
Comparison
DEFECT CLA
SSIFICATION Image fusion
GLCM for image series
Thresholding algorithm
Bayes classifier
CHAPTER 4
IMAGING AND ILLUMINATION SYSTEM
4.1 INTRODUCTION
In a machine vision system, the sensor receives light rays reflected off the surface
being imaged using the camera system. The primary objective is to obtain maximum
details of all the defects that are to be inspected. If defective areas are visible in the
image with very high contrast a better analysis of the inspected surface is possible.
This is dependent on the machine vision set up including the camera, illumination,
lenses etc. There are different kinds of illumination that are possible and the decision
about illumination is mostly governed by the nature of the surface being inspected.
The usually considered model for illuminating surfaces is the Lambertian surface,
which causes diffuse reflection and where the brightness of the illuminated surface is
unchanged depending on the direction of observation. In day to day life, this is the
most widely encountered type of surface and many reflection models can be
developed on the basis of this model. But in the case of metallic objects, the reflection
conditions are based on the specular lobe of reflection. All the rays of incident light
are scattered or reflected predominantly in a single direction, which is usually referred
to as specular lobe of reflection (León, 2002). The reflection characteristics of the
different surfaces are shown in Figures 4.1 and 4.2. The ring surface under
consideration in this work is an example of specular object and produces directional
reflection on being illuminated.
Figu
Figu
4.2
This
ima
obje
with
ure 4.1
ure 4.2 R
BRIGH
s type of ill
age that reve
ect boundar
h regards to
Lambertianpossible dir
Reflection o
HT FIELD
lumination i
eals a lot o
ries are visib
o different ty
n surface shrections
of light on s
ILLUMIN
is well suite
of informatio
ble in the im
ypes of illum
33
howing the
specular sur
NATION
ed for inspe
on about it.
mage. A det
mination is
incident li
rface
ecting diffus
. More info
tailed explan
presented l
ight being
se objects a
ormation im
nation of th
later. The so
scattered in
and produce
mplies that m
his phenome
ource of lig
n all
es an
more
enon
ght is
34
a ring shaped fluorescent lamp. The light is incident as shown in Figure 4.3a. The
characteristic name ‘bright field illumination’ comes from the nature of the image,
which is the whole image area looking very bright. This is one of the most commonly
used illumination systems.
The basic set up for the inspection remains the same for different illumination systems
except the change in position of illumination. The camera used is PULNIX TMC 6
analog camera. The camera is connected to Matrox Corona frame grabber which
facilitate conversion of analog signal from camera into digital images in the PC to be
used for further processing. Matrox Imaging Library is used to capture and to store
the image. Later the images have been processed using Labview or Matlab software.
Optics which is a convex lens attachment is used to form the image in the machine
vision sensor with the required magnification. An aperture can be used to control the
amount of light flooding the machine vision sensor axis. Here 4X objective is used in
microscope and it is attached using a custom made adapter for C-mount lens.
According to the requirement, higher magnification objectives can be used. The
lighting for bright field illumination was made using a circular fluorescent bulb. For
other schemes of illumination, halogen or LED based illumination sources can be
used. Also the direction of illumination changes according to the type of illumination.
For bright field illumination set up the circular fluorescent bulb is concentric with the
camera axis. For using axial illumination which is a special case, a beam splitter was
used. Beam splitter reflects the incident light into the direction of the axis of the
camera and at the same time it transmits the reflected light from the object under
inspection into the camera through the lens system. The ring to be inspected is kept on
an indexing mechanism which can be used to change its position to enable 360°
inspection, step by step.
35
Four camera positions are required to image the critical regions of the ring. Camera
position-4 is used to image two surfaces that are at 103° angle. So the point of
intersection and a small part of both the surfaces can be imaged with camera aligned
bisecting the angle thus giving 52° orientation with respect to vertical for camera-4.
Camera position-1 is used to image the surfaces that are at 112° angle, the bisector of
which is 56°. It has been observed that the imaging is not much affected with ±4°
variation in camera angular position. Angular position of 52° may give an option to
image the ring surface corresponding to camera position-4, in case one camera is to be
avoided (using one camera for both camera position-1 and camera position-4, by
imaging the current surface and the surface corresponding to camera position-4 by
suitably translating the ring or camera since the surfaces are diametrically opposite).
Camera position-2 is for imaging the top surface which is at 6° and thus camera
position can be 6±4° and an angular position of 8° have been decided and used for the
current application according to convenience in installation. Camera position-3 is
used to image the curved surface. If a tangent is drawn to the curved surface, the
normal to the tangent is at 52° with respect to vertical and hence the angular position
value of 52° for that camera.
The calibration of the set up has been done by capturing image of a physical object of
known dimension. The dimension divided by the number of pixels between the edges
in the image will give the calibration value. This has been done for both x direction
and y direction.
The set up discussed in Chapter 6 was developed at Messsystem und Sensor Technik,
Technical Universität München, Munich, Germany. The imaging system consists of a
36
Firewire (IEEE 1394) camera. This is a digital camera which means the image can be
directly captured and stored in the computer system in the form of a digital image
without using an intermediate frame grabber. The Apple Macintosh system can be
directly connected to the imaging system through the IEEE firewire port which is
available in all Apple systems. The images are captured using the Macintosh based
image capturing software and then processed using the standard software as discussed
before. The optics consisted of a convex lens of focal length 6cm and a variable
aperture to control the amount of incident light on the machine vision sensor.
OSRAM LED light sources were connected to a 2.4 volt 700mA DC power source
through a junction box facilitating selective lighting of each of them. The whole set
up was enclosed in a dark light isolating case to prevent disturbance from external
light sources.
4.2.1 Axial Illumination System
Axial illumination system is a special case of bright field illumination. The axial
illumination set up has the direction of illumination and the axis of the camera
coinciding. This is attained by using a beam splitter at the camera axis which allows
illumination of the specimen from a direction perpendicular to the axis of the camera.
The reflected beam from the specimen reaches back through the beam splitter to the
camera. Figure 4.3b shows the system which can use either the axial illumination or
the circular illumination using two separate electrical switches. The images of the ring
taken under bright field illumination are presented in Figures 4.4 -4.12.
37
Figure 4.3 Experimental setup used for imaging of components using a vision
system: (a) Schematic representation of the setup used, showing only camera-2 position, as in Figure 4.3c (b) Photograph of the setup (c) The camera positions for imaging different surfaces.
(a) (b) (c)
(d) (e) (f)
(g) (h) Figure 4.4 Images showing non defective surfaces under bright field illumination
Figu
Figu
(a)
(d
ure 4.5 I
ure 4.6 I
d)
Images sho
(a)
Images sho
wing deep l
wing yellow
38
(b)
(e)
line under b
(b)
(d)
w stain unde
bright field i
er bright fie
(c)
illumination
(c)
eld illumina
n
ation
Figu
Figu
(a)
(d
ure 4.7 I
(a)
(d) ure 4.8 I
)
d)
Images sho
)
Images sho
wing blister
wing built u
39
(b)
(e)
r under brig
(b)
(e) up under br
ght field illu
right field ill
(c)
(f)
umination
(
lumination
)
(c)
Figu
Figu
(a
ure 4.9 I
(a)
(d)ure 4.10 I
a)
(d)
Images sho
)
) Images sho
wing damag
wing peel o
40
(b)
(e)
ge under br
(b)
(e) off under br
right field il
ight field ill
(c
lumination
(
(f)lumination
)
(c)
f)
Figu
Figu 4.3
In m
ima
chal
ima
insp
stud
refle
(a)
ure 4.11 I
(a)
(ure 4.12 I
DARK
microscopy
age and fo
llenging. Th
ages by usin
pected surfa
died using
ection and
)
(d) Images sho
(d) Images sho
FIELD IL
application
or highly
he dark fie
ng an illum
ace (Figure
this strateg
thus receiv
wing pitting
wing rough
LLUMINAT
ns, attempts
reflective
eld approach
mination so
4.13). Typi
gy. The cam
ves only li
41
(b)
(e)g under brig
(b)
(e) h finish unde
TION
s are always
or transpa
h is a meth
ource that s
cally, non-L
mera is ke
ight rays th
ght field illu
er bright fie
s made to im
arent objec
hod of imp
sends rays
Lambertian
ept away fr
hat have be
(c
(f) umination
(c
eld illumina
mprove the
cts, this b
proving the
at a grazin
surfaces ca
from the sp
een scattere
c)
c)
(f) ation
e contrast of
becomes re
contrast of
ng angle to
an be image
pecular lob
ed by the n
f the
eally
f the
o the
ed or
e of
non-
unif
met
Figu
Dar
prep
and
be p
the
back
obje
dark
cont
In t
syst
illum
form defect
thod of illum
ure 4.13
rk field illum
pared surfac
reducing im
predominan
name. Th
kground. T
ects and in
k field illum
trast and thu
the case of
tems are us
mination w
ive surface.
mination.
Principle odefective sa
mination is
ces. It has th
mage proce
ntly dark, if
e presence
This is a co
the inspecti
mination w
us required
f most of th
sed (Figure
would result
. Chrome co
of dark fielample.
applied in
he advantag
essing steps
f there are n
of abnorm
ommonly us
ion of highl
were free fro
less proces
he machine
2.1). Howe
t in very b
42
oated ring su
ld illumina
microscopy
ge of highlig
required fo
no abnormal
malities wi
sed procedu
ly reflective
om the effe
ssing steps f
e vision ap
ever, if the
bright regio
urface is we
ation; (a) d
y for transp
ghting the f
or certain ap
lities in the
ill figure a
ure in the m
e surfaces. T
ect of spec
for segment
plications,
surface is
ns in the i
ell suited fo
defective sa
parent, plana
features wit
pplications.
imaged sur
as bright re
microscopy
The images
ular reflect
tation of def
axial surfa
reflective a
image getti
or this partic
ample; (b)
ar and spec
th good con
The image
rface and he
egions in
y of transpa
s captured u
tions, had g
fects.
ace illumina
and curved,
ng highligh
cular
non
ially
ntrast
will
ence
dark
arent
using
good
ation
this
hted,
43
which suppress the surface details and get detected as boundaries when standard
image processing algorithms are used. Whereas in the case of dark field illumination,
the light rays strike the component at an angle (Figure 4.13) and as a result, flooding
of the machine vision sensor from the on-axis light from the main source of
illumination is reduced (Martin, 2007; Connolly, 2002). The presence of some foreign
particles will reflect the light back into the optical axis (Figure 4.13a). The non
defective surface sends rays out of the machine vision sensor (Figure 4.13b). The
schematic and laboratory setup for dark field imaging is shown in Figure 4.14a and
Figure 4.14b respectively.
Images of defective and non defective regions of different ring specimens have been
captured using bright field and dark field illumination and are shown in Figures 4.4 –
4.12 and Figures 4.15 – 4.23 respectively. Figure 4.15 shows a non defective ring
surface. The whole image is dark and this implies that there are no defective regions.
Non defective surfaces imaged under bright field illumination have been shown in
Figure 4.4 for comparison. Figure 4.16 shows a surface image with blister defect. It is
visible as conglomeration of bright regions and is clearly visible with high contrast in
the dark field image. This implies that the defective region shown in the image can be
segmented out with minimum image processing steps. The corresponding defective
region under bright field illumination is shown in Figure 4.7. The bright streaks
present in bright field image are absent in the dark field image. Figure 4.17 shows
deep line defect which is visible as straight bright line in the dark field image. This
defect is due to the metal cutting process and not just due to the coating process. The
defect occurs around the ring surface as a feed mark. The dark field image (Figure
4.17) reveals the defect area with more details as compared with that shown in Figure
44
4.5, which will aid in quicker and accurate segmentation of defective region. Damage
is a similar defect which occurs due to manufacturing process prior to the coating
process. The coating in this case is without defect but the surface on which coating is
done is defective. Figure 4.18 shows damage region under dark field illumination.
The defect is seen as a very small bright region in this image. Just like deep line,
damage defect also is hard to segment out from bright field image but not so in the
case of dark field since it is comparatively more prominent. Figure 4.19 shows rough
finish which occurs as a defect which continues around the ring surface. Since the
spread is more, it can be easily detected and is visible clearly in both dark field and
bright field illumination (Figure 4.12). It is a defect which is more clearly visible
under bright field illumination. Dark field illumination is suitable for detecting small
sized defects. Figure 4.20 shows the images of pitting defect which appears as small
bright spots and as small clusters. It is distinctly visible both in bright field and dark
field images (Figure 4.11). Built up is visible as a bright spot without much spread
and is visible in both the illuminations (Figure 4.21 and Figure 4.8). Peel off is due to
the absence of coating over a large area and is visible at the boundaries as bright
regions in dark field image (Figure 4.22). This defect is distinctly visible in bright
field image shown in Figure 4.10 also. Figure 4.23 shows the image of the surface
with yellow stain which is visible as yellow colored regions in the inner diameter of
the ring (Figure 4.6). Though it is difficult to detect the defect, there is a unique dark
pattern in the dark field image which implies that it can be processed to segment out
the defective region. It can be concluded that most of the defects can be imaged with
greater contrast using dark field illumination.
Figuillum
ure 4.14 mination; (a
Set up foa) Schemati
or inspectioic; (b) Labo
45
(a)
(b) on of ringoratory set u
g outer diup
iameter usiing dark f
field
Figu
Figu
(a)
(d)
ure 4.15 I
(a
ure 4.16 I
Images sho
a)
(d) Images sho
wing non d
wing blister
46
(b)
(e)
defective sur
(b)
(e)r under dark
rface under
k field illum
(c
(
dark field i
(c
(f)mination
)
(f)
illumination
c)
f)
n
47
(a) (b) (c)
(d) (e) (f) Figure 4.17 Images showing deep line under dark field illumination
(a) (b) (c)
Figure 4.18 Images showing damage under dark field illumination
(a) (b) (c)
(d) (e) (f)
Figure 4.19 Images showing rough finish under dark field illumination
48
(a) (b) (d)
(e)
Figure 4.20 Images showing pitting under dark field illumination
(a) (b) (c)
(d) (e) (f)
Figure 4.21 Images showing built up under dark field illumination
49
(a) (b) (c)
(d) (e) (f)
Figure 4.22 Images showing peel off under dark field illumination
(a) (b) (c)
(d)
Figure 4.23 Images showing yellow stain under dark field illumination
50
4.3.1 Implementation of Dark Field Illumination
The camera, oriented with optical axis normal to the surface of inspection along with
the ring indexing mechanism has been used to create images of all the defect prone
surfaces of the ring which are to be inspected. Four different camera positions are
required to create images of all the required defect prone surfaces and they are shown
in Figure 4.3c (The same positions are used for both the bright field and dark field
illumination set ups). The geometry of the ring surfaces to be inspected has been used
to decide the location of these cameras and this was discussed in Section 4.2.
The proposed system in the industry will have four cameras oriented in these
positions. In the present study, a single camera has been used to take images for
development of algorithms for defect inspection. Figure 4.3a and Figure 4.3b show
the camera being oriented to inspect a specific area on the top of the ring. The
resolution of images captured by the camera is 816 x 612 pixels, where each pixel
represents 5.7 μm x 5.7 μm ring surface area. This explains the smallest defect size
that can be detected which is 10.14 μm i.e. the presence of two defect pixels. The
algorithms were implemented on Intel Core 2 Duo, 2.8 GHz PC with 2GB RAM
using National Instruments Labview software, such that it will be compatible with the
proposed set up in the industry. Equal number of non defective sample images and
defective sample images were used for testing the algorithm. The performance of the
four approaches has been evaluated based on 210 images of ring samples obtained
from the manufacturers.
The expression for accuracy as given by Juran et al. (1994) is,
51
Accuracy of inspector = Percentage of defects correctly identified= (4.1)
Where,
d = defects reported by the inspector
k = number of defects reported by the inspector but determined by the check inspector
not to be defects
(d – k) = true defects found by the inspector = correct detections by the algorithm
b = defects missed by the inspector, as determined by check inspection
(d - k + b) = true defects originally in the product = the total number of samples of
known status that is supplied to the algorithm
The expression is defined especially for correct detection of defective surface
(without considering correct detection of non defective surface). Therefore, in this
study, the accuracy of the algorithm was found out using the expression,
(4.2)
Number of defective sample images, classified as defective by algorithm
Number of non defective sample images, classified as non defective by
algorithm
N Total number of images
Therefore the equation 4.2 can be rewritten as,
Accuracy = N
T (4.3)
For example, if 100 defective image samples are input to the algorithm and the
52
algorithm correctly detects say, 80 of them; accuracy will be 80%. The expression for
accuracy holds true for 100% visual inspection. In the automation of visual inspection
the accuracy of the method is determined by first getting the samples tested by the
inspection system and then by an experienced inspector. A similar methodology has
been followed in this case also. The quality control team from the industry had
supplied chrome coated rings with the defective areas carefully marked out. The
images of these defective regions of the samples were used to test all the algorithms.
This was done by running the algorithm on the images that were predetermined to be
defective and then evaluating the results in terms of number of sample images
detected as defective. The same procedure was repeated for non defective regions of
the ring.
4.4 SUMMARY
A machine vision experimental set up for inspecting the chrome coated ring surface
has been made. All the critical regions of the ring component can be imaged and
processed for presence of defects. Classification can be done if defects are present.
The complete system is a result of a proper organization of camera, frame grabber,
computer system, software, optics, illumination and object locating mechanisms. It
has been observed that the image quality is highly dependent on the type of
illumination to be used. Different illumination methodologies are possible by
variation in the position of illumination. The images of the coated ring surface were
taken under different illuminations namely dark field and bright field. The basic
principle behind each illumination set up has been explained in detail. Each of the
defects was imaged using both the illumination systems and the resultant image has
been studied. The use of suitable magnification lenses and camera resolution will
53
enable detection of smaller sized details. Cameras at different angular positions will
enable inspection of all the defect prone areas of the ring. The indexing mechanism
allows for inspection of entire 360° of the ring surface. As a result all the defect prone
areas of the whole ring can be successfully inspected. The method of evaluation of the
accuracy of the defect detection system has also been discussed.
The types of illuminations used are bright field, axial and dark field. The aim of the
imaging set up is to capture images of the surface to reveal complete details with
suitable magnification. The details should have good contrast, such that they can be
identified and separated out very easily by a simple thresholding algorithm. Bright
field illumination reveals more details about the surface under inspection. The
defective regions are shown as darker areas in the image of the ring surface. There are
significant specular reflections, which might complicate the image processing
algorithm that is to be developed at the next stage. Though axial illumination, which
is a special case of bright field illumination, has similar characteristics, it gives lesser
specular reflection from the surfaces that are near the inspected area. But specular
reflections are still prevalent and this might lead to the loss of some information in the
final image. In dark field illumination, the illumination hits the surface of inspection
at oblique angles. As a result, there is very less specular reflection back into the vision
sensor. Those surfaces that are with abnormalities will be specially highlighted in the
image. This might make the image thresholding operation very simple and fast as
compared to that when the bright field images are used. Though the scattered light
from the surface under consideration is important and influences the image quality,
there is a possibility of losing some light reflected from the inspected surface. With
the experimental set up to capture the images of defective surface images at various
angles and under various illuminations, experiments and analysis have been carried
54
out and the details are presented in the following chapter.
CHAPTER 5
DEFECT IDENTIFICATION
5.1 INTRODUCTION
Image processing methodology to be adopted generally is based on the type of
illumination, the characteristics of object under inspection, the data to be extracted
and type of machine vision system. Though some of the image processing procedures
are adaptable, most of the new applications require newer and different techniques of
pre-processing of the images. The complete vision system has to be designed based on
the inspection requirement which vary for different applications. In many of the
industrial applications the size, dimension, pattern or shape of the objects are
assessed. The application discussed here in this work is principally to detect the
presence of defects and classify them. Different methodologies tried out with regards
to the kind of illumination will be discussed in this chapter. Since speed and accuracy
are the major requirements, each of the methods will be analyzed for suitability,
according to the speed and the accuracy at which the inspection process can be
completed.
5.2 BRIGHT FIELD ILLUMINATION
The presence of more details in the image than what is actually required makes it
difficult to separate out the defective region effectively from the rest of the details. As
a result, image processing is made more complex and error prone. The initial images
of coating defects received from the industry were taken using bright field
illumination technique and as a result, the initial approach was to segment out the
55
defects using those bright field images. The Fast Fourier Transform filter was tried out
for this purpose and the procedure is explained below.
5.2.1 FFT Filtering Based on Critical Radius
This approach is based on the Fourier image reconstruction proposed by Tsai and Kuo
(2007) which they had used for analyzing sputtered surfaces. The defects are known
to be normally the high frequency details in the image. Hence a high pass filter in
frequency domain was used for defect detection.
Conversion from spatial domain to frequency domain was done using 2D FFT given
by,
11
0 0
( , ) ( , ) exp 2yx
NN
x y x y
ux vyF u v f x y j
N N
(5.1)
for frequency variables u = 0, 1,2…Nx-1
and v = 0, 1,2…Ny-1
The power spectrum was calculated from the real and imaginary components of the
FFT. That is
2 2 2( , ) ( , ) , ,P u v F u v R u v I u v (5.2)
Where, R(u,v) and I(u,v) are the real and imaginary parts of F(u,v), respectively
56
The logarithmic value of the amplitude of the FFT of the image is shown in Figure
5.1a. Logarithmic value is useful when it comes to visualizing sudden variations in
value. A radial line from the center of the image to the rightmost end of the image has
been considered and the intensity of the logarithmic plot along this line has been
studied. It was observed that after an initial surge (due to contribution from metallic
streaks) in the value of intensities, the plot begins to stabilize at a particular intensity
value (Figure 5.1b). This point is of significance, since it is the point at which all the
lower frequency components are eliminated and the higher frequency components
start to appear. The radial cut-off distance has to be close to this point. To find the
exact cut-off radius, a disk of increasing radius was used to mask all the intensity
values proceeding from the center outward. The mask size is the radial distance along
the radial line from the center of logarithmic FFT image. Figure 5.1c shows
logarithmic value of the amplitude of the FFT of gray scale values of the image and
the overlapping disk of optimum diameter. Corresponding plot of intensity along the
radial line is shown in Figure 5.1d. Original image before Fourier filtering is shown in
Figure 5.1e. The image is reconstructed using 2D inverse FFT after making zero all
the intensity values within the disk of optimum size (Figure 5.1f). This reconstructed
image predominantly gives the defective area segmented out from the statistical
texture. After running it on many images, the frequency corresponding to the optimum
size of the disk has been found as 60% of maximum frequency. In the algorithm for
the FFT approach discussed, this observation was used to filter out the FFT at single
step with a high pass cut-off frequency of 60% to save processing time and to reduce
the complexity of the approach. This process of training the model and then using it to
predict future results is a methodology followed in classifier design for pattern
recognition algorithms. The filtered frequency image was again converted to spatial
domain, using 2D inverse FFT. The reconstructed image showed the high frequency
57
details (defects), which were then segmented by thresholding.
(a) (b)
(c) (d)
(e) (f) Figure 5.1 FFT filtering(a) FFT image (b) Intensity (logarithmic value of FFT
amplitude) variation along the radial line indicated in Figure 5.1a (c)
Radial line and disk of optimum radius used to mask the intensity
(logarithmic value of FFT amplitude) values within it (d) Intensity
(logarithmic value of FFT amplitude) variation in Figure 5.1c along the
radial line after the filtering operation using the mask (e) Original
image used (f) Reconstructed (Inverse FFT) image of Figure 5.1c.
Intensity
Distance along profile (pixels)
Intensity
Distance along profile (pixels)
58
5.2.2 Auto-Median Based Approach in Image Segmentation
The accuracy of the detection was improved when the reconstructed image was
subjected to a contrast enhancement and then thresholded. Later, the auto-median
morphology operation was used to remove the small objects which used to produce
false detections. The expression for the auto-median operation is given below
· · · (5.3)
Where,
operation
· operation
M – 3 x 3 Structuring element
M = 1 1 11 1 11 1 1
I – Image
Auto-median operation( ) on image(I), using a 3 x 3 square structuring element(M),
is a morphological operation which consists of producing two images by
implementing opening( ) and closing(·) operations in the sequence shown in equation
5.3, and then applying logical AND( ) operation on them (Thomas, 2003). Opening
is erosion followed by dilation whereas closing is dilation followed by erosion. Auto-
median is used to remove noise (which might appear as defect in the FFT approach)
in the analyzed image.
59
5.3 DARK FIELD ILLUMINATION
From the images of defective area under dark field illumination it can be understood
that a simpler thresholding operation may suffice for image segmentation of these
images. Just as bright field illuminated images are processed by auto-median based
algorithm, a new algorithm has been developed to successfully detect defects and then
segment them out while inspecting rings using dark field illumination. Figure 5.2a
shows the histogram of the image of a non defective component under dark field
illumination and Figure 5.2b shows the first derivative plot of the histogram. The first
derivative is given by,
1 (5.4)
Where,
.
It can be observed that, lim 0. Histogram in Figure 5.3a (of defective
sample image) does not show any major spike in the number of pixels but there is
noticeable spike in the plot of at g→255 (Figure 5.3b) unlike in the case of non
defective sample mentioned before. This observation has been used to fix a threshold
value of 230 for all the dark field images, such that details which are corresponding to
the defects are segmented out. The major spike in the plot of h’(g) corresponds to the
boundary between the ring surface and the background (visible at g≈50 in Figure 5.2b
and Figure 5.3b. The ring surface information is visible from g=50 or g=70. The
defective region if present will be visible above the value of g=70. Thus the lower
60
limit of G is defined as g=70. The dark field illuminated images after applying this
threshold is shown in Figure 5.4. After thresholding, the number of particles or the
number of white (g=255) pixels can be counted. Here the number of particles (any
inter connected group of white pixels) can be counted as a single particle. After
testing threshold g=230 for a large sample of images it has been found that for a
threshold value of 230, detection of a single particle deems the component to be
defective. Lower the threshold value, more the particles will be detected but at the
same time non defective areas may also get detected as defective areas. This is
illustrated in Figure 5.4.
(a) (b)
Figure 5.2 Image characteristics for non defective surface; (a) Histogram of image of non defective surface of ring under dark field illumination, (b) Plot of first derivative of h(g) in 5.2a where h’(g)≈0 as g→255
(a) (b) Figure 5.3 Image characteristics for defective surface; (a) Histogram of image of
defective surface (blister) of ring under dark field illumination, (b) Plot of first derivative of h(g) in 5.3a showing the peak in the value when g→255
Gray scale values
First Derivative
Gray scale values
First Derivative
Peak
61
Original dark field image
Thresholded image using FCD approach
Thresholded at g=127
Thresholded at g=230 (Few white pixels present for
defective)
Blister (a) (b) (c) (d)
Yellow stain (e) (f) (g) (h)
Blister (i) (j) (k) (l)
Non defective False detection Few white pixels
detected No white pixels
detected (m) (n) (o) (p)
Figure 5.4 Effect of using different threshold values Note: White pixels represent the defective regions. Images in each row are obtained by thresholding the images shown in the first column. Images were taken from different rings.
62
5.3.1 FRACTIONAL CHANGE IN DERIVATIVE (FCD) METHOD FOR
FINDING OPTIMUM THRESHOLD VALUE
FCD (Zhang et al., 2006) can be used to segment out the defect completely from the
surrounding non defective surface in the image of ring surface. The basics of this
approach are explained in the following paragraph. Details in segmented image can be
used to quantitatively assess the type of defect. The image has to be segmented at the
point of steep increase in the histogram plot of dark field image. This can be found
out using the value of the variation of slope of histogram plot. Figure 5.4b, Figure 5.4f
and Figure 5.4j show the defective region segmented out distinctly by using this
observation. For finding optimum threshold g’, the increase in the value of with
respect to the immediate neighboring gray scale value has been found out. This is
given by,
FCD(g)= 1 / (5.5)
g’: =maximum value of FCD(g) in G, where (5.6)
G={ , 255 70
FCD(g)=Fractional Change in Derivative
g’=optimum threshold value
The value g=70 has been found to be the least observed value for getting the defects
segmented out for all the thresholded images and hence the lower limit for G was
fixed as g=70. In Figure 5.4 c, Figure 5.4g and Figure 5.4k the threshold value was set
close to half of the maximum value of 255 for illustration purpose. This gives a fair
idea about the approximate spread of the defective area, if present in the image. Both
the threshold values obtained by maximum FCD(g) and g=127 as it is in this case,
63
fails for non defective images. Hence a very high threshold value of 230 was fixed
such that none of the non defective areas are segmented out as defects and such
images are shown in Figure 5.4d, Figure 5.4h and Figure 5.4l. The image of non
defective sample in Figure 5.4m has been subjected to thresholding based on all the
three values and the respective resultant images are shown in Figure 5.4(n), Figure
5.4(o) and Figure 5.4p. It can be found that only from Figure 5.4p with threshold
value, 230 as discussed before; resultant image is obtained with no objects being
detected. The conclusions based on these observations are given below.
1. Threshold value of 230 can be used to identify the presence of defects on the
imaged surface.
2. After ascertaining the presence of defects, the maximum FCD(g) method can
be used to correctly segment out the defects for further processing. This may
be used only on images of defective regions after the identification of defect
by using single threshold value of 230, since the non defective images will
give erroneous results when directly operated upon by the algorithm.
3. The images of non defective surfaces will not show any distinct area
representing defects while threshold value of 230 is used.
5.4 DEFECT CLASSIFIER USING BRIGHT FIELD ILLUMINATION
IMAGES
Defect classifier need to be used, only if the image is found to be having a defective
area. The defect classes for classification problem were redefined as,
64
1. Blister – Defect class containing blister defect samples. This was kept as a
separate class because the defect contained larger spread of small objects as
compared to the other defects. This is non-repairable defect.
2. Built-up – Defect class containing built up defect samples. This was also a
separate class because this kind of defective ring can be reworked to salvage
the ring.
3. Damage – Defect class containing all the other kinds of defects for example;
damage, peel off, pitting, deep lines. This is non-repairable defect.
The defective region parameters that have been considered while designing the
classifier are:
1. Area- Scalar which is equal to the actual number of pixels in the region.
2. Perimeter- The length of the boundary of a region.
3. Extent - Scalar that specifies the proportion of the pixels in the bounding box
those are also in the region. Computed as the pixel area of defect divided by
the pixel area of the bounding box.
5.4.1 Classification Parameter Selection
Though, the segmented defect area can be characterized by different parameters, only
in few of them may be able to give an efficient classification of defects (Duda et al.,
2001). The parameters that give clear classification between the classes of defects can
be used for classification. For this purpose conditional probability of different kinds
of d
peri
clas
the
the
can
The
Figu
5.5
5.5.
In a
pixe
defect param
imeter/area
ssification a
curve perta
parameter
be underst
erefore these
ure 5.5
RESUL
.1 Image P
all the algor
els) with an
meters have
and exte
and this is i
aining to bl
clearly sepa
tood that b
e two param
Conditionadamage cextent/area
LTS AND D
Processing
rithms, to re
n area of mo
been plotte
ent/area wi
llustrated in
lister defect
arates out t
uilt up def
meters can b
al probabiliclasses (a) a(Ex/A)
DISCUSSIO
Algorithm
educe the e
ore than 5 b
65
ed. It has be
ill serve t
n Figure 5.5
t class is se
the blister d
fect class ca
be used for t
(a)
(b)
ity density Plot of
ONS
ms for Brigh
effect of noi
by 5 pixels,
een found th
the purpos
5 (Eiden et
eparate from
defects. Sim
an be easily
the classifie
functions perimeter
ht Field Im
ise, the obj
, were set a
hat two para
se of givin
al., 2000).
m the other
milarly from
y separated
er.
for blisterr/area(P/A)
ages
ects (repres
as defects. A
ameters nam
ng very g
In Figure 5
ones, impl
m Figure 5.5
d from the
r, built up (b) Plot
sented by w
Any presenc
mely
good
5.5a,
ying
5b it
rest.
and t of
white
ce of
66
such objects will imply that the work piece sample that has been imaged is defective,
and the pixel count or pixel area of such particles will indicate the intensity of the
defect. The FFT approach for defect detection on bright field images is shown in
Figure 5.6. This approach has been discussed by Tsai and Kuo, (2007) to detect
defects in sputtered surfaces. In this particular case, non defective ring surfaces were
not correctly detected from the corresponding images (Figure 5.7). Figure 5.7b shows
the plot of intensity function after filtering frequencies below 60% in the FFT of
image in Figure 5.7a. Figure 5.7d shows the final reconstructed image after the
filtering operation. There are some frequencies in a non defective component image
that were above the cut-off frequency and appeared as defective areas in the
segmented images. There is no ideal cut-off radius or frequency for defects which
would segregate only the defective region. It could be due to the fact that the defects
in the components are diverse in nature, localized and prominent, unlike the defects
that occur on the sputtered surfaces.
The auto-median operation was appended to the Fourier filtering approach in an
attempt to improve the accuracy. This approach removes the smaller particles left
behind, after the Fourier filtering method and the final resulting image is shown in
Figure 5.8f. Figure 5.8d is obtained after contrast enhancement on the image obtained
from the FFT approach. The resulting image is thresholded and then closing operation
is done (Figure 5.8e). The final image, with defects segmented out, is shown in Figure
5.8f and this is the result of auto-median operation and removal of border objects. The
89% accuracy of the system in detecting and classifying the components as defective
and non defective parts, was at the cost of increased time of inspection to 0.292
seconds from 0.248 seconds for the FFT method. The reason for this is the fact that an
additional auto-median operation was used after the Fourier filtering.
67
Figure 5.6 Image segmentation using FFT approach to highlight the defective region in bright field image: (a) Original image of blister (b) Logarithmic plot of intensity as a function of radius from center of FFT power spectrum (c) High-pass filtered FFT power spectrum, showing radial line considered (d) Logarithmic plot of intensity as a function of radius from center of FFT power spectrum in 5.6c (e) Reconstructed image of 5.6c.
Figure 5.7 Failure of the FFT method on bright field images of non defective surface: (a) Original image of non defective surface (b) Logarithmic plot of intensity as a function of radius from center of FFT power spectrum (c) High-pass filtered FFT power spectrum, showing radial line (d) Reconstructed image of 5.7c
As mentioned earlier, the rings are to be inspected for all the defect-prone areas by
indexing, which necessitate the reduction in run time for an image. Finally, it was
observed that the performance of the auto-median approach is good in terms of
average processing time and also has a fairly high accuracy of detection. Auto-median
68
approach which gave an accuracy of detection of 96% when only defective samples
were tested makes sure that very less defective samples will make it to the customer.
When only non defective samples were tested, auto-median approach gave an
accuracy of 81%.
Figure 5.8 Image segmentation using auto-median approach on bright field image: (a) Original image of blister (b) Value plane extracted from 5.8a (c) Reconstructed image of High-pass filtered FFT (altered by brightness and contrast changes to make details visible to reader) (d) Contrast enhancement of 5.8c (e) Close operation on 5.8d (f) Auto-median operation followed by removal of border objects in 5.8e
Different rings were inspected and the images of the defective as well as non
defective surfaces were captured. The algorithms for bright field image were applied
on these images. Figures 5.9 a-r show the resultant images after applying each of the
two algorithms on the defective sample image (Figures 5.9 a, 5.9 d, 5.9 g, 5.9 j, 5.9 m
and 5.9 p). Figures 5.10 a-i show the resultant images after applying each of the two
algorithms on a non defective sample image (Figures. 5.10 a, 5.10 d and 5.10 g). It
can be observed that FFT approach fails for many of the non defective surface images
unlike auto-median approach. Both the approaches correctly detect and segment the
defective region for images of defective surfaces.
Br
(
Figu
Wh
resu
surf
right field im
(a) Built u
(d) Peel of
(g) Rough fin
(j) Deep lin
(m) Damag
(p) Pittingure 5.9
en used und
ult, while im
face area wi
mage
up
ff
nish
nes
ge
g Performancsurface of rshown in tImages wer
der 4X mag
maging all
ill be in foc
FFT Appro
(b)
(e)
(h)
(k)
(n)
(q) ce of FFT aring. Note: the first colre taken fro
gnifications,
the surface
cus (approx
69
oach A
and auto-meWhite pixe
lumn is usem selected
, the depth
es other tha
imately 17m
Auto-median
(c)
(f)
(i)
(l)
(o)
(r)edian approaels represened to obtaindefective su
of focus is
an the top
mm) and th
n Approach
)
)
)
)
)
) ach on imag
nt defective n other imaurfaces on d
approximat
flat surface
e rest of the
ges of defecregions. Im
ages in the different rin
tely 2 mm. A
e, only a s
e surface im
ctive mage row.
ngs.
As a
mall
mage
will
resu
Suc
algo
was
B
Figu
5.5.
The
pres
5.11
are
l be blurred
ults while pr
ch an effec
orithm and t
s applied.
Bright field im
(a)
(d)
(g) ure 5.10
.2 Image P
e blister defe
sence of str
1d). Thus, t
four majo
d gradually
rocessing im
ct was reall
therefore ve
mage
Performancdefective sregions. Imin the row. different rin
Processing
fect imaged
reaks along
the defect d
or peaks an
y as the su
mages of cu
ly high for
ery low ove
FFT Appr
(b)
(e)
(g)ce of FFT surface of
mage shownImages we
ngs.
Algorithm
using brigh
with the d
detection ma
nd many m
70
urface heigh
urved surfac
r non defe
erall accurac
roach (
and auto-mring. Notein the first
ere taken fro
m for Dark F
ht field illum
defects make
ay not be e
more minor
ht varies. T
ce resulting
ctive samp
cy in result
Auto-media(No white pi
(
(
(median appe: White pcolumn is uom selected
Field Imag
mination is s
es the histo
easy as it w
r peaks eac
This might
in less accu
ples tested
s where the
an Approachixels detected
(c)
(f)
(h) proach on pixels repreused to obtad non defect
ges
shown in Fi
ogram multi
was discusse
ch indicatin
give erron
uracy of res
using the
e FFT algori
h d)
images of esent defec
ain other imtive surface
igure 5.11a.
i-modal (Fi
ed earlier. T
ng a partic
eous
sults.
FFT
ithm
non ctive
mages es on
The
gure
There
cular
71
important feature in the image. This shows that features other than this defect are
being focused upon. And when observed under axial illumination, this defect appears
as presented in Figure 5.11b. The streaks are still present, but the histogram (Figure
5.11e) reveals that the defect detection would be simpler than in the previous case.
Three major peaks are present and show that the number of features visible in the
image is less than that in the case of bright field illumination. But still, the specular
reflection is present and may get detected as a false defect. Further, the same defect
observed under dark field illumination (Figure 5.11c) resulted in uni-modal intensity
distribution (Figure 5.11f). This shows clearly that the features of the defect alone are
focused and the rest are neglected. Hence, the defect segmentation is now a simple
thresholding operation (i.e. simplified image processing). The setup used for such a
dark field imaging on OD of a ring is shown in Figure 4.14. With a suitable
orientation of camera and illumination, it is possible to image ID or any other curved
surface of the ring under dark field imaging conditions. Dark field illumination is
found useful, for inspecting inner and outer fillets of the ring, which was nearly
impossible when bright field or axial illumination systems were used.
Another kind of coating defect (peel off) imaged using axial illumination is shown in
Figure 5.12a. The histogram (Figure 5.12d) again follows the same trend as before.
The dark field illuminated image of the same defect (Figure 5.12c) gives uni-modal
histogram (Figure 5.12f). The mode corresponding to the bright pixel region occurs as
a sudden disturbance in the histogram towards the higher gray scale value. In short,
the dark field illumination gives an image which is uni-modal and thus easy to
threshold to detect the defects. The histograms for all the defects in different areas of
the ring follow the same pattern. All the images clearly highlight the defective areas
leaving out the details of areas that are non defective. The thresholding as discussed in
72
Section 5.3 helps in quick defect detection. The performance of defect detection
algorithms on a Core2Duo 2.8 GHz computer with 2GB RAM, for different imaging
conditions is shown in Table. 5.1. The information was deducted from a sample size
of 210 images for bright field illumination approach and 210 images for dark field
approach. Dark field images for different kinds of defects from different rings and the
corresponding thresholded images are shown in Figures 5.13 - 5.20. Dark field images
of blister (Figure 5.13), deep line (Figure 5.14), heavy blister (Figure 5.15), built up
(5.16), rough finish (5.17), pitting (5.18), damage (5.19) and non defective surface
(5.20) are processed and the resulting binary images are shown.
Table 5.1 Performance comparison of algorithms
S. No
Method/ Algorithm AccuracyRun time per
image 1 FFT on bright field images 57% 0.248 s 2 Auto-median on bright field images 89% 0.292 s
3 Single step thresholding on dark field
images 96.5% 0.047 s
Figure 5.11 Image of blister in different illumination systems; (a) Bright field; (b)
Axial illumination; (c) Dark field illumination; (d) Histogram of 5.11a, (e) Histogram of 5.11b, (f) Histogram of 5.11c.
Figu
Figu
Figu
ure 5.12
(a)ure 5.13
(a)ure 5.14
Image of pilluminationHistogram
) Single stering(showinthresholdinthresholdindetected)
) Single stepring(showinthresholdinthresholdin
peel off in n; (b) Axiaof 5.12a; (e
ep thresholng blister)
ng 5.13a ang 5.13a at
p thresholdng deep lin
ng 5.14a ang 5.14a at g
73
different ilal illuminate) Histogram
(b) lding for under dark
at gray scgray scale
(b) ding for imne) under daat gray scgray scale v
lumination tion; (c) Dam of 5.12b;
images ofk field illucale value e value g=2
mages of ark field illcale value value g=230
systems; (ark field il(f) Histogra
f blister; umination (
g=127 (c230 (Few w
deep line; lumination
g=127 (c0
a) Bright Flumination;am of 5.12c
(c) (a) Image
(b) Image ac) Image awhite pixels
(c) (a) Image
(b) Image ac) Image a
Field ; (d) c
e of after after s are
e of after after
Figu
Figu
Figu
(a)ure 5.15
(a)ure 5.16
(a)
ure 5.17
) Single stepring(showinafter threshthresholdin
) Single stepring(showinthresholdinthresholdindetected)
)
Single stepring(showinafter threshthresholdin
p thresholdng heavy bholding 5.1
ng 5.15a at g
p thresholdng build up
ng 5.16a ang 5.16a at
p thresholdng rough fholding 5.1
ng 5.17a at g
74
(b) ing for im
blister) und15a at graygray scale v
(b) ding for ip) under daat gray sc
gray scale
(b)
ding for imfinish) unde17a at graygray scale v
ages of heder dark fiey scale valuvalue g=230
images of ark field illucale value e value g=2
mages of roer dark fiely scale valuvalue g=230
eavy blistereld illuminaue g=127 (
0
built up; umination (
g=127 (c230 (Few w
ough finish;ld illuminaue g=127 (
0
(c) r; (a) Imagation (b) Im(c) Image a
(c) (a) Image
(b) Image ac) Image awhite pixels
(c)
; (a) Imagation (b) Im(c) Image a
e of mage after
e of after after s are
e of mage after
Figu
Figu
Figu
The
oper
whe
dete
(a)
ure 5.18
(a)ure 5.19
(a)
ure 5.20
e algorithm
ration using
ether the im
ection. This
)
Single stering(showinthresholdinthresholdin
) Single stering(showinthresholdinthresholdin
)
Single stepof ring(shothresholdindetected) (c(No white p
to identify
g a thresho
maged ring
is illustrate
ep thresholng pitting)
ng 5.18a ang 5.18a at g
p thresholng damage
ng 5.19a ang 5.19a at g
thresholdinowing damang 5.20a at c) Image afpixels are d
y the presen
ld value of
g is defecti
ed for differ
75
(b)
lding for under dark
at gray scgray scale v
(b) ding for i) under darat gray scgray scale v
(b)
ng for imagage) under d
gray scalefter thresholetected)
nce of defec
f 230. This
ive or not
rent images
images ofk field illu
cale value value g=230
images of rk field illucale value value g=230
es of non ddark field ile value g=1lding 5.20a
ct involves
single step
and is su
in Figures
f pitting; umination (
g=127 (c0
damage; umination (
g=127 (c0
efective surlumination
127 (Few wat gray sca
a single ste
p may be us
fficient for
5.13c, 5.14
(c)
(a) Image(b) Image ac) Image a
(c) (a) Image
(b) Image ac) Image a
(c)
rface; (a) Im(b) Image a
white pixelsale value g=
ep threshol
sed to ascer
r simple de
4c, 5.15c, 5.
e of after after
e of after after
mage after s are =230
ding
rtain
efect
.16c,
76
5.17c, 5.18c, 5.19c and 5.20c. The gray scale value of g=127 is used to ascertain the
spread of defects for illustration purpose in Figures 5.13b, 5.14b, 5.15b, 5.16b, 5.17b,
5.18b, 5.19b and 5.20b. The presence of a single white pixel implies the presence of
defects. If the defects are to be segmented out distinctly, the FCD approach can be
used. This information may be used to find out the actual spread and class of defects.
However, applying dark field illumination to specific areas of ring without mutual
interference of the light from multiple illumination systems will be difficult. Since the
machine vision sensor is away from the specular lobe of reflection from the reflective
surface in the case of dark field illumination (Figure 4.2), the intensity of light
reflected from the defects are always very less and in some cases might cause the
defects not to be detected. But still in the image, the defects are in very high contrast
from the other non defective surfaces and thus the defects are easily detected in most
cases. Finer details regarding the regular textured surface are lost, since it appears as
dark pixels in image and as a result, it cannot be used directly to quantify many of the
surface texture features. There is also possibility of pixels getting saturated due to the
powerful illumination that is being used which might be harmful to machine vision
sensor.
In the past, contrast enhancement using different illumination conditions were
attempted by earlier researchers on many other engineering applications (Piper, 2008;
Bamforth, 2007). Therefore, there is further scope here in this case as well, for using
dark field illumination with similar contrast enhancement for inspection of defects in
these ring components.
77
A single step thresholding algorithm (with a higher bound as threshold value) has
been run on 210 images, taken using dark field technique, and has given an accuracy
of 96.5% with running time of 0.047s. Out of the 105 images of non defective rings
that have been tested, 97% were detected as non defective. The accuracy of detection
for images of defective rings is 96%. The comparative performance is given in Table
5.2. This is a significant improvement over the other approaches. The images taken
using dark field illumination can be easily evaluated for the presence and absence of
defects. On the other hand, there is a significant loss of information, since major
portion of the light is reflected out of the machine vision sensor (Majority of which is
specular reflections but with some from the defective surfaces). As a result,
thresholding may not necessarily give all the details of the defects as in the case of the
other discussed algorithms, which are applied to bright field images. Single step
thresholding operation, using dark field imaging technique, cannot be used to
correctly quantify the extent and spread of defects but can be used as an easy and fast
way to identify the presence of defects. This can be understood from comparing
Figure 5.17 a-c and Figure 5.9 g-i. The spread of defect is not correctly visible in the
dark field image but represented clearly in bright field image. Added to this, is the
practical difficulty in setting up directed illumination systems for each part of the ring
to be inspected. As a result, each area to be inspected will require its own specific
illumination system.
The experimental set up in the laboratory was used only to evaluate the algorithm
developed and thus indexing was done manually. The proposed set up will have the
following specifications. The camera has a field of view of 46mm x 35mm (area).
There needs to be 360/60=6 indexes to cover the whole ring (where 600 is angular
value corresponding to each index). The processing time stated in the table is for a
78
single image captured at a single indexed position of the ring. An indexing mechanism
with 0.5s for a single index can be used. The maximum image processing time for the
auto-median approach on all the four angular positions of the camera and for all the
indexed positions of the ring is given by 0.155 x 4 x 6 = 7s/ring (The indexing time of
2.5s is included in this, since both are done in parallel). The time of inspection for
different approaches is given in Table 5.3. This proves that the entire inspection
process can be completed within 8 seconds, as required by the company, by using any
of these approaches.
Table 5.2 Comparative performance with respect to FFT approach on bright field images (FFT takes 0.248s at an accuracy of detection of 57%)
S. No
Method/ Algorithm Improvement in
accuracy Change in run time
per image
1 Auto-median on bright field
images 56% -17%
2 Single step thresholding on dark
field images 69% 97%
Table 5.3 Time of inspection per ring
S. No
Method/ Algorithm Run time per
image Time/Ring
1 FFT on bright field images 0.248 s 5.95 s 2 Auto-median on bright field images 0.292 s 7 s
3 Single step thresholding on dark field
images 0.047 s
1.1 s
5.6 SUMMARY
The advantages of using dark field illumination in particular have been highlighted in
this case study. Based on the results and analysis the following conclusions are drawn.
It was clearly observed that the dark field illumination was performing better
compared to bright field illumination in the sense that the image processing became
79
simpler because of the improved quality of images obtained. Therefore, the entire
process of defect detection became faster. When using the dark field illumination, the
effect of bands of reflective regions was absent. These bands were occurring due to
specular reflections in axial and bright field illumination. The presence of defect on a
dark field illuminated image can be identified using a single thresholding operation
rather than using other complex image processing algorithms like FFT approach and
auto-median approach which are discussed in Section 5.2. The entire process of
identification of defects by using simple thresholding operation was faster by
approximately 6 times (Table 5.3), when compared with algorithm based on auto-
median approach on bright field images. There is scope for further improvement if the
image processing set up is coupled with higher magnification lenses. In that case, very
small defects can be easily detected.
CHAPTER 6
DEFECT CLASSIFICATION USING MULTIPLE IMAGES
6.1 INTRODUCTION
An accurate and quick way to detect the presence of defects has been presented and
discussed in Chapter 5. Once the presence of defects is detected it has to be classified
into the corresponding defect class which is a normal requirement in many pattern
recognition applications. The basic steps to build a classifier are discussed in Section
1.3.4. The images of ring surfaces containing particular classes of defect have been
captured. For doing this a multiple illumination based method has been proposed. The
images were segmented using a thresholding algorithm and the defective regions were
separated out. The thresholded region has been quantified using different parameters
and the most suitable one was selected. The selected parameters are to be input to the
classifier. The classifier which satisfies the requirements of classification is selected.
Later the classifier is trained on the basis of gathered information. The classifier is
then tested for performance and depending on the performance all the intermediate
steps are fine tuned. This is the procedure used in this work for design of classifier.
The basic methodology to develop pattern recognition algorithm has been discussed
earlier in Chapter 1. In Chapter 5, the use and effectiveness of machine vision sensor
to detect the presence of defects has been discussed. The design and implementation
of machine vision sensor system is also explained. Now there is a requirement of
finding the most suitable parameter to be used in the defect classifier. There can be
different parameters that can be used in the classifier. Some of them are geometrical
and some are statistical. Also, the kind of algorithm to be used to process the
81
parameter needs to be identified. This will indeed help in the easy separation of rings
into their respective lots by the help of mechanical sensors.
Dark field images have the inherent disadvantage that the information from some of
the defective surfaces is lost due to reflection outside the angular field of view of the
sensor. As a result, the images in dark field reveal less information than bright field
images, though they amplify the light reflection from some of the defective regions.
The first approach used for bright field images using geometrical parameters makes
the defect detection more complex and less accurate. Parameters which aid in easier
and more accurate defect detection will help in more efficient defect detection and
classification. As discussed earlier, the modification in the initial manufacturing
process will result in some newer types of defects or some defect classes becoming
desolate. This has been experienced during the execution of the current project for the
ring manufacturer. As a result, the study has been conducted on a selected number of
classes of defects that are more frequent than the rest. The classifier can be modified
depending upon the changes in the inspection requirement in the industry.
The classes of defects that were considered in this particular approach are,
1. Blister
2. Built up
3. Damages
4. Pitting
5. Yellow stain
These were selected based on the requirement of the industry to classify the rings
specifically to avoid defective components being supplied to the customers.
82
6.2 MULTIPLE IMAGING SET UP
For a given component which is to be inspected, its underlying physical
characteristics remain unchanged for any angle of illumination or imaging. But the
specularities and the other illumination dependent aspects, as seen in the image, keep
changing their relative position; with the change in illumination set up. This aspect
can be used to find the defects separated out from the images of the inspected
component with minimum errors. This has been attempted by Khalili and Webb
(2007) using different wavelengths of illumination. In the current approach, three
illumination positions have been used, all of them giving partial dark field images.
Each illumination inclination and the position is similar to the one discussed in the
previous chapters. For discussion purpose, the images can be described as L1 for the
one with illumination at the left, L2 for the one with illumination at the centre and L3
for the one with illumination at the right sides.
The number of lighting combinations possible = 3C1+3C2+
3C3 =7,
Consider that Li is the case of ith illumination being used while image is captured.
3C1 is from using each of the illumination separately which is {{L1},{L2},{L3}}.
3C2 is from using two illumination simultaneously which is {{ L1 L2},{ L1 L3},{ L2
L3}}
3C3 is from using three illumination simultaneously which is { L1 L2 L3}
All the seven images were captured one by one and then given as input to the feature
separation algorithm. The number of images can be any depending upon the
feasibility of implementation. The image of the set up is given in Figure 6.1. In the set
up,
syst
thre
blis
Figu
Figu
an IEEE F
tem. The il
ee different
ter defect, c
ure 6.1
ure 6.2
irewire CC
lumination
illuminatio
captured usi
Experiment
(a)
(d)
Effect of vblister usin
D camera h
sources we
on positions
ing differen
tal set up fo
variation inng three diffe
83
has been att
ere three L
s. The imag
nt illuminati
or illuminati
(b)
(e)
(g) n illuminatiferent illumi
tached to a
LED (OSRA
ges of the s
on sources
ion with 3 i
ion positionination posi
suitable len
AM make) m
same area o
are given in
llumination
n (a-g) shoitions
ns and aper
mounted in
of ring show
n Figure 6.2
n sources
(c)
(f)
ows the de
rture
n the
wing
2.
efect
84
6.3 GRAY LEVEL CO-OCCURRENCE MATRIX (GLCM) APPROACH
FOR MULTIPLE IMAGES
GLCM approach is frequently used in texture measurements. Initially it was
formulated to segment satellite images of the earth for remote sensing purposes. The
basic concept is the correlation between a pixel and its nearby pixel. The neighbor to
the pixel is determined by the relations defined by ∆x, ∆y and ѳdir (Figure 6.3a). Gray
Level Co-occurence Matrix will have address indexes ranging from 1 to 256 in both
axes. Each entry in the matrix will be the number of times of occurrence of pixel
having x axis index value and y axis index value as neighbors, as defined by the
condition for neighbors. This is indicated in Figure 6.3b. This approach is particularly
useful while dealing with textures which have repetitive patterns. A host of statistical
parameters have been defined to quantify the GLCM matrix. Earlier, they were used
to classify based on single image only.
(a) (b)
Figure 6.3 GLCM computation (a) The conditions for being neighbors (b)Formulation of GLCM matrix for condition 00 [0,1].
Lindner et al. (2007) have explained how multiple illumination systems can be used
for the inspection of carbide cutting tool surfaces and the corresponding processing of
the image by using GLCM matrix approach adapted for multiple images. The
mathematical expression used in this approach is given below,
85
C∆ , ∑ ∑ ∑ , , . ∆ , ∆ , ∆ , (6.1)
1 0 (6.2)
The variables g1 and g2 are the gray values of the first and second pixel, respectively,
and the vector ∆ = (∆x, ∆y, ∆ѳ)T describes the considered displacement between two
pixels.
The increment used in this case was ∆ = (0, 0, ∆ѳ) where ∆ѳ had only two possible
increments since only three positions were used. A window of a particular length and
breadth and angular displacement has been considered and the matrix has been found
for each of the index of the pixels considering this window. The bigger window size
will result in the loss of the information from the image. A very small window size
will defeat the purpose behind using the GLCM approach. The optimum window size
has to be determined that would result in the image details getting amplified. For the
discussions which follow, window size of 3 pixels x 3 pixels is considered.
The dark field imaging has the inherent property of saturating the machine vision
sensor pixels thus leading to a loss of information. So, it is necessary to consider the
neighbouring pixels while arriving at any particular decision. This is the major
advantage of Gray Level Co-occurrence Matrix approach, since one cannot make a
decision based on the information given by certain stray pixels; especially when there
is a possibility of pixel saturation due to the enormous amount of light intensity at the
86
bright pixels. All these reasons make it well suited for applications warranting
repetition of pattern or lay.
But in the current approach, the lay can be defined across the images in the third
dimension considering the image series produced using different illumination
positions. There is an inherent repetition of patterns across the series of images since
they are of the same region of interest, but with distinctive variations. This makes the
approach, which is otherwise not well suited to non homogenous textures, applicable
to this particular problem. In this approach, the unique but less pronounced
disadvantage of dark field imaging is circumnavigated in two steps
1. By considering a window across the same image and pixel under
consideration.
2. By considering the same windows in the respective series of images.
The approach under consideration will make up for the lack of certainty, while
deciding the defect class based on geometrical features.
Here a particular window size was considered and the GLCM matrix for that
particular window across the series of images has been obtained for each of the pixels
in the image (Figure 6.4). The GLCM matrix can be used to define 14 statistical
parameters for any given pixel in the image series under consideration. Though the
parameters are not quantized into a particular dynamic range, the display shown in
Figure 6.5 has the parameter matrix values scaled to fit the dynamic range for a gray
scale image. But each of the images shows a representation of the visual contrast
between defective area and the normal non defective adjacent areas. This information
87
can be used to select the parameters giving the most optimum separation for the
defective region from the surrounding non defective region. The mathematical
definitions (Haralick et al. 1973) for these parameters are as follows.
If,
p(i, j) := (i, j)th entry in a normalized gray tone spatial dependence matrix (GLCM)
px (i) := ith entry in the marginal probability matrix obtained by summing the rows of
p(i, j), =∑ , (6.3)
Ng:=Number of distinct gray levels in the quantized image
Σi and Σj are ∑ and ∑ respectively
∑ , (6.4)
∑ ∑ , , 2,3, … ,2 (6.5)
∑ ∑ , , 0,1 … . , 1| |
(6.6)
The parameters are,
1. Angular Second moment or energy
∑ ∑ , (6.7)
2. Contrast
∑ ∑ ∑ ,| |
(6.8)
3. Correlation
∑ ∑ , (6.9)
88
Where,
µx µy σx and σy are the means and standard deviations of px and py.
4. Sum of squares
∑ ∑ , (6.10)
5. Inverse Difference moment
∑ ∑ , (6.11)
6. Sum average
∑ (6.12)
7. Sum variance
∑ (6.13)
8. Sum entropy
∑ log (6.14)
9. Entropy
∑ ∑ , log , (6.15)
10. Difference Variance
(6.16)
89
11. Difference Entropy
∑ log log (6.17)
12 Measure of correlation-1
, (6.18)
13 Measure of correlation -2
1 (6.19)
∑ ∑ , log , (6.20)
Where,
HX and HY are entropies of px and py and
1 ∑ ∑ , log (6.21)
2 ∑ ∑ log (6.22)
14 Maximal Correlation Coefficient
, ∑ , , (6.23)
Figu
From
vari
con
calc
num
requ
find
segm
ure 6.4 I
m Figure 6
iance and
sidering the
culations w
mber of com
uirement to
d the idea
mentation.
Image fusio
.5, the para
Entropy h
e higher con
ere time co
mputations w
consider a
l cut off
on for calcu
ameters Con
have been
ntrast of the
onsuming g
which gets c
a particular
value for
90
ulation of GL
ntrast, Sum
selected f
images und
given the si
compounded
region of in
each par
LCM param
average, S
for implem
der consider
ize of imag
d with the s
nterest to im
ameter for
meters
Sum of squa
menting the
ration. Sinc
ges and the
size of the im
mplement th
r better de
ares, Differe
e classifica
ce the param
e correspon
mage, there
he algorithm
efective re
ence
ation
meter
nding
e is a
m to
gion
91
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j) (k) (l)
(m) (n) (o)
Figure 6.5 Gray scale representations of GLCM parameters: (a) Contrast, (b) Sum average, (c) Sum of squares, (d) Difference variance, (e) Entropy, (f) Sum of entropy, (g) Measure of correlation-1, (h) Measure of correlation-2, (i) Energy, (j) Correlation, (k) Homogeneity, (l) Sum variance, (m) Difference entropy, (n) Maximal correlation coefficient, (o) Region of interest from Figure 6.2 under consideration for production of images Figure 6.5 a-n
92
6.4. IMAGE MASKS FOR EACH REGION OF THE RING
Fast and accurate implementation of the algorithm to find the ideal cut off value of the
parameters requires that the regions of the image should be separated out. For this
purpose, three regional masks were made with unique shapes matching the shape of
the region of the ring under inspection. These masks can be generated for any
particular region of the ring that is imaged, provided the illumination conditions are
same throughout the region under consideration. One major problem with the
processing of images was that while camera position is kept constant and more
regions are imaged using the particular position of camera under three different
illumination conditions, there is a major variation of appearance of defects in the ring
surfaces. In some cases, the background is bright and the defect regions are dark,
contrary to what happens when dark field illumination is done. Though this is not the
case in dark field illumination, the reduction in the number of cameras will force a
trade off in this aspect of the image. So, it was necessary that the regions needed to be
masked according to the nature of illumination, dividing them to different regions.
Three masks that have been used for the images are shown in Figure 6.6. They can be
generated using morphological operations on the initial image sample to be used.
First, the values of the gray scale intensities have been increased to a constant value
leaving the lower pixel values and then the regional minimum pixel values were
found out to from a binary image. Later the holes (stray pixels) have been removed,
revealing three different regions which were labelled such that they can be used to
separate out the required pixel regions from the image. Though the masks are
dependent on the position of camera, they can be easily generated for the given
positions of the camera and then used for further processing at the later stages.
93
Figure 6.6 Masks used for different regions of the image of ring surface.
6.5 IMAGE THRESHOLDING OPERATION
The image will have majority of pixels lying in the non defective region of image and
remaining very few, if at all, in the defective region of image. This information has
been used to design a thresholding algorithm. This is based on finding the ideal cut-
off value for the parameters that have been evaluated before. For this purpose, two
parameters were scaled and the scatter plot was made. The values which are forming
the bigger cluster were removed. Every combination of parameters was plotted and
the corresponding threshold values were combined to form the final threshold image
which was mapped back to the original image to find the boundaries of the defective
images. This enables the use of the parameter values from the defect threshold images
for the classification step.
For a better understanding of the thresholding process methodology, a flow chart is
provided in Figure 6.7. The input to the thresholding algorithm is any two among the
selected parameters. Each of the parameters will be used to form the two different
Top region Middle region Bottom region
94
axes of an image matrix and the data is represented as white pixels in a dark
background. Later the number of bright pixels in this image is found out by using the
number of pixel values above one half of the maximum gray scale resolution of the
image. If the number of bright pixels is more than the number of dark pixels, then
each of the axes is inverted. This is preceded by an image closing operation using a
structuring element of square shape and 5 x 5 size. This operation will cluster all the
closely lying pixels together as a single object, the extremity of which is taken as the
image index corresponding to the threshold value of the two parameters considered.
The parameter threshold value corresponding to the index that has been obtained is
used to cut off the parameter values in the input parameter matrix, finally ending up in
a thresholded binary image region. So, output of a single thresholding operation gives
two binary images each based on one each of the two parameters that was initially
given as input. These two binary images are added together using logical AND
operation to obtain a single binary image. Similarly, all such binary images from the
different parameter combinations considered are subjected to logical OR operation
culminating in a single image which has all the necessary details of the defect
segmented out. Element wise multiplication of this resultant binary image and each of
the parameter matrices will yield a matrix having parameters only in the image region
which is defective.
Figu
The
It ca
l) an
ure 6.7
e series of th
an be obser
nd by using
Find the paramthreshold corresponding tindex of image
If Axiinverson done
N
Y
Thresholdinoperation woperation
hreshold im
rved that, al
g only a sing
Input twat a tim
Find theafter remobject in
meter
to
is si
Remthan
Remgreatimag
ng operatio
with graphic
mages obtain
ll the combi
gle combina
wo parameters e
e starting index moval of the n the image
move pixels having parcut off to get two bin
move pixels having parter than cut off to get ges
95
(a
(b) on (a) Scatcal explanat
ned from th
inations wil
ation, one c
Scale the parameters to same range
Morphologicclosing on im
rameter less nary images
rameter two binary
a)
tter image tion (b) Flow
e parameter
ll not yield
can not arriv
Form imagparameter
cal mage
Invert a
Logical opfinal thresh
N
used in thwchart for t
rs are show
good result
ve at all the
ge using as each axis
If more bright pixels
axis of image
peration to get the holded image
Y
he thresholthe threshol
wn in Figure
ts (Figure 6
e correct det
ding ding
6.8.
.8 k-
tails.
96
As a result, the good images are combined to get an optimum threshold image. The
information fusion at the thresholding operation is illustrated in Figure 6.9.
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j) (k) (l)
Figure 6.8 Threshold images considering different parameters;(a) Contrast image after thresholding with Difference variance,(b) Difference variance image after thresholding with Contrast, (c) Contrast image after thresholding with Sum average,(d) Sum average image after thresholding with Contrast,(e) Contrast image after thresholding with Sum square, (f) Sum square image after thresholding with contrast, (g) Sum average image after thresholding with Sum square, (h) Sum square image after thresholding with Sum average,(k) Sum average image after thresholding with difference variance, (l) Difference variance image after thresholding with Sum average.
97
(a)
(b) Figure 6.9 Information fusion for thresholding operation (a) Schematic
representation (b) Final thresholded image from sample threshold images
98
6.6 DEFECT CLASSIFIER
The parameters after the thresholding operation have been plotted together to
understand the separability for the defect classes on the basis of the parameters
considered. The parameters can be plotted and depending on the value of the normal
distribution function, the class of defect corresponding to the parameter value can be
found out as shown in Figure 6.10.The defect parameters have been plotted against
each other in a scatter plot as shown in Figure 6.11. Though there is overlap in some
classes, considering all the parameter combinations together is a more effective
method of classification. For each defect class, six different combinations of
parameters can be plotted on the two dimensional graph.
Figure 6.10 Plot for Sum average showing the separation of defects using normal
distribution values
0 50 100 150 200 250 300 350 400 450 5000
0.02
0.04
0.06
0.08
0.1
0.12
Sum average
N(x
)
Blister
Built up
DamagePitting
Yellowstain
Pitting Blister
Yellow stain
Damage Built up
Norm
al distrib
ution, N(x)
Sum average
99
Figure 6.11 Plots showing the sum of squares against sum average with the clusters
of defects well separated out. Plots showing pixels corresponding to: (a) all the defects (b) blister, (c) built up, (d) damage, (e) pitting, (f) yellow stain.
Later, a Bayesian classifier has been applied to find the boundaries for the defects.
The parameters were approximated as normal distribution and the corresponding
graph has been plotted to visualize the boundary. Similar graphs have been plotted for
(a) (b)
(c) (d)
(e) (f)
Sum Average
Sum
of
squa
res
Sum
of
squa
res
Sum
of
squa
res
Sum
of
squa
res
Sum
of
squa
res
Sum
of
squa
res
Sum Average
Sum Average Sum Average
Sum Average Sum Average
100
all the combinations of the parameters and finally information has been fused to
decide the class membership of each pixel of the image under consideration. In fact,
the graphs give the visualization of the final decision process but the original decision
is taken based on the value of the discriminant function for the particular pixel with
respect to all the classes of defects.
6.6.1 Bayesian Classifier
Following are the basics of construction of Bayesian classifier based on the
assumption that the distribution is normal. Multivariate normal density in d
dimensions is written as
| | (6.24)
Where, x is a d-component column vector, µ is the d-component mean vector, Σ is the
d-by d covariance matrix, and | | and Σ-1 are its determinant and inverse,
respectively. If d=2,
Σµ µ µ
µ µ µ
Let x be a continuous random variable whose distribution depends upon the state of
nature and is expressed as p(x|ω) which is the class conditional probability density
function for x given the state of nature is ω. Consider the prior probabilities P(ωj) and
the conditional densities p(x|ωj) for j=1,2 i.e 2 classes. The probability density of
finding a pattern that is in category ωj and has feature value x can be written in two
ways : p(ωj,x)=P(ωj|x)p(x)=p(x|ωj)P(ωj). From rearranging this Bayes formula is
arrived at.
101
, (6.25)
Where in the case of two categories
∑ , 1 (6.26)
The basic rule is: decide that the variable belongs to class ω1 if P(ω1|x) > P(ω2|x);
otherwise decide ω2.
To represent pattern classifier, one of the most useful methods is to define a set of
discriminant functions gi(x), i=1,….,c. The classifier assigns a feature x to class ωi if
gi(x) > gj(x) for all j≠i.
Thus the classifier can be viewed as a network or machine that computes c
discriminant functions and selects the category corresponding to the largest
discriminant. A network representation can be seen in Figure 6.12
Figure 6.12 Working of a classifier (Duda et al., 2002)
X1 X2 Xd X3 ……….
g1(x) g2(x) gc(x) ……….
Action (e.g., classification)
Cost calculation from discriminant functions
input
102
There is no unique discriminant function. Each function can be multiplied by a
constant or scaled by another constant. Also, it can be written as a function of
monotonically increasing function without changing the classification. It is normally
written as
gi(x) = p(x|ωi)P(ωi) (6.27)
To simplify computation it has been written as,
gi(x)=ln p(x| ωi) + ln P(ωi) (6.28)
Though the discriminant functions are written in many forms, the purpose of it is to
divide the feature space into c decision regions, R1, R2…Rc. If gi(x)> gj(x) for all j≠i,
then x is in Ri, and the decision rule calls for us to assign x to ωi. The regions are
separated by decision boundaries, surfaces in feature space where ties occur among
the largest discriminant functions.
After approximating the distribution as normal distribution the following has been
defined. Multivariate normal density in d dimensions is written as
| |, where (6.29)
X is a d-component column vector, µ is the d-component mean vector, Σ is the d-by d
covariance matrix, and |Σ| and Σ-1 are its determinant and inverse, respectively (All
bold faced characters represent vectors). In the general multivariate normal case, the
covariance matrices are different for each category. So the expression can be written
as,
0.5 Σ ln 2 ln|Σ | ln P (6.30)
, (6.31)
103
, (6.32)
(6.33)
and ln| | ln (6.34)
The decision boundary in this case will be hyper quadrics depending upon the spread
of the distributions (Figure. 6.13).
Finally, given an image series of the defective region, the algorithm returns the final
thresholded image with parameter representations for each of the defect pixels. The
information fusion approach used in classification is illustrated in Figure 6.14. The
final image after defect pixel classification is shown in Figure 6.15. The images
shown are without usage of masks. The masks can be used depending upon the
requirement in the real life application. The decision regarding the pixel membership
was taken after considering every combination of parameters as in the thresholding
algorithm. Whichever defect class has a better score, the allotment is made to that
particular class. Table 6.1 represents an example of decision making process based on
discriminant values after giving parameter value as input. The defect class is assigned
to the defect class having the highest value of discriminant. An example of calculation
of scores for a sample defect is shown in Table 6.2. The defect considered is pitting
and the corresponding discriminant values for every defect class and every parameter
is shown. If one of the parameter values points to the defect class as pitting, the score
for pitting is set to one. Finally the pixel is classified as the defect class having
maximum score. In this particular case the score for pitting is maximum and thus the
defect class is finally assigned as pitting.
104
Figure 6.13 Bayesian decision boundaries for defect classes in the scatter plot of
sum average against difference variance. Plots showing: (a) all the defects with under lying boundaries, (b) blister, (c) built up, (d) damage, (e) pitting (f) yellow stain
Difference variance
Difference variance
(a)
(c) (d)
(e) (f)
Sum
Ave
rage
Sum
Ave
rage
Su
m A
vera
ge
Sum
Ave
rage
Su
m A
vera
ge
Sum
Ave
rage
Difference variance
(b)
Difference variance Difference variance
Difference variance
Tab Def
Blis
Bui
Dam
Pit
Yel
Fin
Tab
P Def BlisBuiDamPittYelRes
Figu
ble 6.1 Parameter
fect
ster
lt up
mage
low stain
al Decision
ble 6.2
arameter
fect
ster ilt up mage ting llow stain sult
ure 6.14
Classificatir g(x
for
-3.7
-11
-5.0
-1.2
-7.7
n Pitt
Decision mpitting defeContrast=5
-3.74 -11.03 -5.01 -1.22 -7.73 Pitting
Illustration
on based onx) for sum ar pitting(x=
74
.03
01
22 →Larges
73
ting
making by sect
Sum Average
-6.79 -44.88 -6.19 -3.17 -17.23 Pitting
of informat105
n discriminaaverage
=5)
st
coring for d
e=60 DifVar
-24-76-84-4.3-14Pitt
tion fusion
ant (g(x)) vag(x) for sufor built u
-3.75
-3.29→La
-4.99
-220.02
-13.06
Built up
different pa
fference riance=280
4.36 6.5 4.94 39
4.78 ting
in pixel cla
alues um square up(x=300)
argest
arameters co
0 Sum squar
-4.83-17.2-5.1 -0.7 -9.4 Pittin
assification
orrespondin
re=4 Sco
0 0
0 4 0
ng Pit
ng to
ore
tting
Figu
(a
(c
(e)
(g)
(i)
ure 6.15
a)
c)
)
)
Thresholdeblue is builstain). (a) Icontrast beRepresentaFigure 6.15image usingpitting, (h)thresholdinRepresentaFigure 6.15
ed images wlt up, greenImage of blfore thresho
ation of orig5 c, (e) Img contrast b) Represent
ng for Figation of orig5i
106
with defectivn is damagelister, (b) Rolding for Fginal image mage of dambefore threshtation of o
gure 6.15gginal image
(b)
(d)
(f)
(h)
(j)
ve areas col, yellow is p
RepresentatiFigure 6.15a
using contrmage, (f) Rholding for
original imag, (i) Imag
using contr
lored (wherpitting and on of origin
a, (c) Imagerast before tRepresentatFigure 6.15
age using ge of yelrast before t
re red is bliblack is ye
nal image ue of built upthresholdingtion of orig5e, (g) Imagcontrast be
llow stain, thresholding
ister, ellow using p, (d) g for ginal ge of efore
(j) g for
107
6.7 RESULTS AND DISCUSSIONS
The defect classification requires more information about the defects, if a better
classification is to be achieved. The usage of different illumination systems gives
more information about the area under consideration. From the images in Figure 6.2,
it can be understood that the details are varied in each of the images, though the same
details are being imaged in all of them. Each of the features have varying highlights in
different images. The usage of the window has made sure that each pixel represents
information which is related to the neighbouring pixels, rather than an abnormal value
which is seldom correlated to any other pixel values in the vicinity. The comparison
of the same pixel values in different images will yield a more concrete idea about the
presence of a particular feature in image. The images in Figure 6.5 prove this aspect.
Throughout the implementation of GLCM approach, a fusion of information has been
used to make better decisions. One major disadvantage of this approach is the number
of calculations which in turn increased run time for each algorithm implementation.
As a result, it was attempted to consider a region of interest by creating a mask based
on the camera position which might reduce the computational load and run time by
many times. It also helped to simplify the thresholding operation. The thresholding
operation was done by finding the optimum cut off frequency for the parameters
under consideration. An interesting fact to be noted is that only the spatial coordinates
of the image is the same at all stages of classification operations, since the gray scale
values have been replaced by statistical parameters which have their own unique
range. As a result, they had to be normalized to a unique scale while doing operations
in unison with another pixel. Similar conversion of scale happens in all the
measurement processes in metrology. As a result, the parameters have to be
accompanied with a scaling factor while implementing the final classification process
108
also. The assumption that dark field images have predominantly all the pixel values in
the darker side have been used to make the thresholding algorithm. The images in
Figure 6.8 have the defective area separated out to a very good extent. Here again, the
information fusion has been applied through logical operations to arrive at a final
image with an optimum separation of defective region. The usage of information
fusion will help in making decisions on the basis of a large amount of valid
information, thus decreasing the possibility of error in the decision. This can be done
by avoiding all information, which is obviously erroneous, by implementing a
condition for evaluating the suitability of particular information for use in the
algorithm (For example, in a thresholding operation the whole image cannot be
showing defective area of the ring). The parameters in the index values corresponding
to the thresholded region is now passed to the defect classifier. The Bayesian
classifier based on normal distribution has been considered. All the data points could
not be fitted exactly into the normal distribution. Again, an information fusion
approach was considered where the data from different combinations could be
considered together and then the best possible decision taken. The parameters were
considered two at a time, since the classifier becomes erroneous when the covariance
matrix approaches zero (singularity). Instead of using a pseudo inverse which could
not give consistent results, inversion has been done in a manual fashion and then the
determinant multiplication done at a later stage. The results are good except for the
case of yellow stain, which is a difficult kind of defect compared with the rest.
Nevertheless many of the pixels corresponding to yellow stain were classified
correctly. The flow chart for the proposed approach is given in Figure 6.16
109
Figure 6.16 Flowchart for the defect classification approach
6.8 SUMMARY
The images from the single step thresholding cannot be used to get the exact details of
the defect which will enable classification of the rings into the respective defect
classes. A multiple imaging approach has been successfully implemented to enable
defect classification and thus, the defective ring classification in the proposed set up.
Three illumination sources have been used for imaging and these images were used to
calculate the GLCM parameters defined especially for image series. The defect region
in image was segmented out using these parameters and then the parameter
corresponding to the defective region in image was transferred into the pattern
recognition algorithm. The scatter data shows that the defects form well separated
clusters with minimal overlap. The Bayesian classifier has been used to determine the
defect class of the pixel. All these processes were done basically using an information
fusion approach in the background and this has facilitated an optimum solution to the
problem. An added advantage is that the classifier can be updated for different kinds
of defects, more kinds of images and different types of classifiers depending upon the
requirement. This aspect of defect classifier makes it adaptable for different kinds of
Image Series
Multiplication of mask
Parameter calculation
Thresholding Discriminant evaluation and scoring for each pixel
Highest Score
Mask generation
Training data
Pixel classified as defect with highest score
110
inspection methodologies and has better chances of being implemented successfully
for any given inspection task. This approach coupled with the basic dark field
illumination approach can be used in the ring inspection problem considered in this
work, successfully for classification of all the various defects.
CHAPTER 7
CONCLUSION AND SCOPE FOR FUTURE WORK
7.1 CONCLUSION
An approach to inspect highly reflective Chrome coated textile ring components using
machine vision has been presented in this work. The coated surfaces of the ring were
imaged using an experimental set up with a machine vision system using different
illumination systems. These approaches may be used to find the presence of defects
and successful classification of defects with very high accuracy and reliability.
Based on this work, the following are the specific conclusions:
1. Though bright field illumination reveals a lot of information about the ring surface
under inspection, too bright an image with large amount of light reflected due to
the curvature and high reflectivity of the ring does not suit this particular
application of inspecting ring components. Therefore, dark field illumination is
considered since high contrast images could be obtained revealing only the
information related to the abnormalities in the surface.
2. Images taken using bright field illumination can be processed to segment the
defective regions but at the cost of increased processing time. Auto-median based
image processing algorithms could segment the defective surfaces successfully
with high accuracy.
3. Dark field illuminated samples have been processed using single step thresholding
process, which can be used to find the presence or absence of defects with both
higher speed and accuracy. The segmentation for defective surfaces can be done
112
by Fractional Change in Derivative (FCD) method for dark field image proposed
in this work to get the optimum segmentation of image.
4. Creating series of images by various illumination systems in the dark field
imaging and then, processing them using GLCM approach for image series is
proposed for defect classification. After segmentation of the image, the parameters
under consideration are fed into the classifier, where a Bayesian classifier is used
to make decisions namely the classification of defects in images. An information
fusion approach was used successfully to get the final results.
5. Using the present approaches it was possible to speed up the entire defect
identification and classification of ring components in particular. The accuracy
improved by 56% and the speed decreased by 17% when auto-median based
image processing algorithm on bright field images was used. The improvement in
accuracy and speed was of the order of 69% and 97% respectively when single
step thresholding on dark field illumination was used.
In short, an accurate and quick method to identify and classify the coating defects on
Chrome coated ring components using dark field illumination has been established.
7.1 SCOPE FOR FUTURE WORK
The proposed set up could be practically implemented and fine tuned to meet the
requirement of the industry. Also, instead of the normal halogen light source, other
light sources could be tried out. There can also be quantification of the surface or
geometrical features of the ring by methods such as phase shifting interferometry and
structured lighting sources etc. Such approaches can be as well used to find the
surface finish of the components under consideration.
REFERENCES
1. Abrahamovich G., Barhak J. and Spicer P. (2005) Reconfigurable array for machine vision inspection (RAMVI). Proceedings of the 3rd International CIRP Conference on Reconfigurable Manufacturing. Ann Arbor, USA, R-01, May 10–12, 2005.
2. Aluze D., Merienne F., Dumont C. and Gorria P. (2002) Vision system for defect imaging, detection, and characterization on a specular surface of a 3D object. IVC, 20, 569-580.
3. Bamforth P. E., Jackson M. R. and Williams K. (2007) Transmissive dark-field illumination method for high-accuracy automatic lace scalloping, International Journal of Advanced Manufacturing, 32(5-6), 599-607.
4. Beckmann P. and Spizzochino A. The scattering of electromagnetic waves from Rough Surfaces. New York: Pergamon, 1963.
5. Beyerer J. and León F. P. (1997) Detection of defects in groove textures of honed surfaces, International Journal of Machine Tools & Manufacture. 37(3), 371-389.
6. Beyerer J. and León F. P. (1998) Adaptive separation of random lines and background. Optical Engineering, 37(10), 2733-2741.
7. Connolly C. (2002). Lighting for industrial inspection. Sensor Review. 22(2), 106-112.
8. David M. (2008) Dark field illumination Available from: http://encyclopedia.jrank.org/articles/pages/1126/Darkfield-Illumination.html. Accessed on 8 April, 2010
9. Duda R. O, Hart P. E., and Stork D.G. Pattern Classification. New York: John Wiley & Sons, 2001.
10. Eiden G., Kayadjanian M.and Vidal C. (2000). Capturing landscape structures:Tools.Availablefrom:http://ec.europa.eu/agriculture/publi/landscape/ch1.htm#1. Accessed 25 April, 2009.
11. Gadelmawla E. S., Koura M. M., Maksoud T. M. A, Elewa I. M. and Soliman H. H. (2002) Roughness parameters. Journal of Material Processing Technology, 123, 133-145.
12. Gadelmawla E. S. (2004) A vision system for surface roughness
characterization using the Gray Level Co-occurrence Matrix. NDT&E, 37, 577-588.
13. Guise M., Poe D. K., Stafford J. and Wahba A. (2002) Automated defect
pattern recognition: An approach to defect classification and lot characterization.Availablefrom:http://www.sys.virginia.edu/students/capstone/past/cap2002/2002-01.pdf. Accessed on 25 April 2009.
14. Haralick R. M., Dinstein and Shanmugam K. (1973) Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3, 610-621.
114
15. Iivarinen J. and Visa A. (1998) An adaptive texture and shape based defect classification. 14th International Conference on Pattern Recognition (ICPR'98), 1, 117.
16. Jayavarthanavelu D, Technical Newsletter Available from: www.lakshmimach.com/tmd/training-newsletter.pdf, accessed online on 23 March 2010.
17. Juran J.M. and Godfrey A.B. Juran's Quality Control Handbook. Fifth Edition, McGraw-Hill. New York, 1998
18. Khalili K. and Webb P. (2007) The development and application of a multiple wavelength illumination technique for the vision-based process monitoring of aero-structure riveting. Machine Vision Application, 18, 73-83.
19. Lee M. F. R., de Silva C. W., Croft E. A. and Wu Q. M. J. (2000) Machine vision system for curved surface inspection. Machine Vision Applications, 12, 177-188.
20. León F. P. (1997). Enhanced imaging by fusion of illumination series, Sensors, Sensor Systems, and Sensor Data Processing. Proceedings of SPIE, 3100, 297-308.
21. León F P. (2002) Evaluation of honed cylinder bores, Annals of CIRP, 51(1),
503-506.
22. Lindner C., Schäffler F. and León F. P. (2007) Texture-based surface segmentation using second-order statistics of illumination series. Lecture notes in informatik, 109, 32-37.
23. Luo P.F. and Liou S.S. (1998) Measurement of curved surfaces by stereo vision and error analysis. Optics and. Lasers Engineering, 30, 471-486.
24. Marrs C.D. and Porteus J.O. (1985) Nondestructive defect detection in laser optical coatings. Journal of Applied Physics, 57(5), 1719-1722.
25. Martin D. (2007). A Practical guide to machine vision lighting. Advanced Illumination.Availablefrom:http://advancedillumination.com/uploads/downloads/A Practical Guide to Machine Vision Lighting.pdf. Accessed on 30 Jan 2009
26. Miyoshia T., Takahashia S., Takayaa Y. and Shimadab S. (2001) High
sensitivity optical detection of oriented micro defects on Silicon wafer surfaces using annular illumination, CIRP Annals - Manufacturing Technology, 50(1), 389-392.
27. Nayar S. K., Fang X. and Boult T. (1997) Separation of reflection components using color and polarization. International Journal of Computer Vision, 21(3), 163-186.
28. Nayar S.K., Ikeuchi K. and Kanade T. (1991) Surface reflection: physical and geometrical perspectives. IEEE Transactions on. Pattern Analysis and Machine Intelligence. 13(7), 611-634.
115
29. Noda N. and Kamimura S. (2008) A new microscope optics for laser dark-field illumination applied to high precision two dimensional measurement of specimen displacement. Review of Scientific Instruments, 79(023704), 1-7.
30. Pfeifer T. and Wiegers L. (1998) Adaptive control for the optimized adjustment of imaging parameters for surface inspection using machine vision. Annals of CIRP, 47(1), 459-462.
31. Piper J. (2008) Luminance contrast, a new illumination technique in light microscopy: Optical basics, practical evaluations, further developments, Optik - International Journal for Light and Electron Optics, 120(18), 963-975.
32. Porteus J. O., Spiker C. J. and Franck J. B. (1986) Correlation between He-Ne scatter and 2.7 μm pulsed laser damage at coating defects. Applied Optics, 25(21), 3871-3879.
33. Povolo F., Canzian A., and Favret E. (1997) Unidirectional laser oblique illumination (ULOI): a comparison with partial dark field illumination (PDFI) in the optical microscopy of metals, Applied Optics, 29(2), 67-73.
34. Rosati G., Boschetti G., Biondi A. and Rossi A. (2009) Real-time defect detection on highly reflective curved surfaces. Optics and Lasers in Engineering, 47(3-4), 379-384.
35. Salas W. A., Boles S. S. H., Frolking S., Xiao X. and Li C. (2003) The perimeter/area ratio as an index of misregistration bias in land cover change estimates. International Journal of Remote Sensing, 24(5), 1165–1170.
36. Samak D, Fischer A and Rittel D. (2007) 3D Reconstruction and visualization of microstructure surfaces from 2D images. Annals CIRP, 56(1), 149-152.
37. Sun Y., Bai P., Sun H.Y. and Zhou P. (2005) Real time automatic defect detection of weld defects in steel pipe. NDT&E International, 38, 522-528
38. Thomas K. Image Processing with LabVIEW™ and IMAQ™ Vision. Prentice, 2003.
39. Torrance K. and Sparrow E. (1967) Theory for off-specular reflection from roughened surfaces. Journal of Optical Society of. America, 57, 1105-1114.
40. Tsai D.M. and Huang T.Y. (1999) Automated surface inspection for directional textures. Image and Vision Computing, 18, 49–62.
41. Tsai D.M. and Huang T.Y. (2003). Automated surface inspection for statistical textures. Image and Vision Computing, 21, 307–323.
42. Tsai D.M. and Kuo C.C. (2007) Defect detection in in-homogenously textured sputtered surfaces using 3D Fourier image reconstruction. Machine Vision and Applications, 18, 383-400.
43. Wiltschi K., Pinz A. and Lindberg T. (2000). An automatic scheme for steel quality inspection. Machine Vision Applications, 12, 113-128.
116
44. Yi S., Haralick R. M. and Saphiro L. G. (1995) Optimal sensor and light source positioning for machine vision. Computer Vision and Image Understanding, 61(1), 122-137.
45. Zhang X., Krewet C. and Kuhlenkötter B. (2006) Automatic classification of defects on the product surface in grinding and polishing. International Journal of. Mach. Tools Manufacturing, 46, 59-69.
117
LIST OF PUBLICATIONS 1. Riby Abraham Boby, Prashant S. Sonakar, B. Ramamoorthy and M.
Singaperumal. Identification of defects on highly reflective ring components and analysis using machine vision. International Journal of Advanced Manufacturing Technology, (available online DOI:10.1007/s00170-010-2730-3).
2. Prashant S. Sonakar, Riby Abraham Boby, B. Ramamoorthy and M. Singaperumal. Identification of defects on highly reflective coated ring components using dark field illumination and image segmentation using simple thresholding technique. International Journal of Automation and Control, (accepted).
3. R. A. Boby, M. Singaperumal, B. Ramamoorthy. Automated inspection and sorting of defects of ring components using Machine Vision. IJP Tech special issue on Precison Metrology (accepted).