hyperspectral face recognition by texture feature extraction using hybrid wavelets type i %2c type...

24
Hyperspectral Face Recognition by Texture Feature Extraction using Hybrid Wavelets Type I , Type II Transforms and Kekre Wavelet Transform Authors Pallavi P. Vartak & Dr. Vinayak A. Bharadi

Upload: dr-vinayak-bharadi

Post on 16-Jul-2015

87 views

Category:

Engineering


4 download

TRANSCRIPT

Hyperspectral Face Recognition by

Texture Feature Extraction using

Hybrid Wavelets Type I , Type II

Transforms and Kekre Wavelet

Transform

Authors

Pallavi P. Vartak

&

Dr. Vinayak A. Bharadi

Contents

• Introduction

• Literature Survey

• Proposed System

• Proposed Algorithm

• Results & Discussions

Introduction

• Biometrics

Technologies that are automated to make attempts

for conformation of an individuals claimed identity.

• How ?

By comparing a submitted sample to one or more

previously enrolled templates.

• Types

Hand based Biometrics & Face based Biometrics.

Introduction contd.

• Classification based on requirement

Unimodal Systems

&

Multimodal Systems

Introduction contd.

• Identifying Humans by their faces is the oldest technique

used.

• What is Face Recognition of Hyperspectral Images ?

Introduction contd.

• HYPERSPECTRAL IMAGES

• They contain a great number of spectral bands or spectra.

• They can acquire the intrinsic spectral information of the

skin at many delicate wavelengths.

• It has ability to capture distinct personal identification

patterns shaped by their molecular composition that

relates to tissues, blood and structure.

• Can overcome the difficulties faced in traditional face

recognition systems, like variance of face orientation,

light distortion or expressions.

Introduction contd.

• THE BIOMETRIC RESEARCH CENTRE AT

HONG KONG POLYTECHNIC UNIVERSITY

• In this research we have used Hyperspectral face database

developed by them which provides us an opportunity to

advance the research in face recognition and compare its

effectiveness. In this existing system individual image

band is used for feature extraction and recognition.

Introduction contd.

Illustration of a set of 33 Hyperspectral face bands The Hong Kong

Polytechnic University Hyperspectral Face database (Poly U-HSFD)

Literature Survey

• Face biometric belongs to both physiological and behavioral

categories.

• Face has advantage over other biometrics because it is a

natural, non-intrusive, and easy-to-use biometric. [1] ,[9] &

[10].

• Statistical techniques, such as PCA [11], LDA [12], ICA [13]

and Bayes [14] etc., are used to extract low dimensional

features from the intensity image directly for recognition

Literature Survey contd.

• Multi resolution Transform such as, Gabor Wavelet Transform, was

used to extract the spatial frequency, spatial locality and orientation

selectivity from faces irrespective of the variations in the expressions,

illumination and pose [18]

• 3 methods are proposed for hyperspectral face recognition, including

whole band (2D)2PCA, single band (2D)2PCA with decision level

fusion, and band subset fusion based (2D)2PCA

• H. B Kekre and V. A Bharadi [19] detailed the concept of hybrid wavelet

transform in interpretation of combining traits of two different

orthogonal transform wavelets to achieve the strength of both the

transform wavelets.

Literature Survey contd.

• The hybrid of DCT and DKT gives best results among the

combination of four transforms used for generating hybrid wavelet

transforms.

• Kekre, Sarode and Dhannawat [20] used Kekre’s wavelet combine

images of same object or scene so that the final output image

contains more information such image fusion gave comparatively

better results just closer to best results with an added advantage

wherein it can be used for images of any sizes, not necessarily integer

power of 2

Literature Survey contd.

• V. A. Bharadi, P. Mishra and B. Pandya [15] used hyperspectral

images with 33 band are used for generation of feature vector

based on Vector Quantization (VQ) process. Popular VQ

Algorithms like Kekre’s Fast Codebook Generation (KFCG)

Algorithm and Kekre’s Median Codebook Generation (KMCG)

Algorithm are used to generate codebooks. These results clearly

indicated that the security performance index of KMCG and KFCG Front is

better than that of Left, Right, Left + Right, Front + Left + Right. The

PI of KFCG and KMCG Front + Left + Right is better than other feature

vector type.

Proposed System

Proposed System Block Diagram Description

• Pre-Processing block is used to accept the

hyperspectral face image data. The PolyU

Hyperspectral Face Database [7] is used for

this current research. The database

contains face images each with 33

frequency bands. These instances of the

image are taken at 33 diverse frequencies

with the help of hyperspectral image

capturing sensors. In which Front, Left

and Right side face images are captured.

These images are stored in Hypercube

MAT [21] format; they are also called as

‘Face cubes’.

Proposed System Block Diagram Description

• Hybrid Wavelet Transforms are

performed on this data. The Hybrid

Wavelet Type I (HW TI)

Transform, Hybrid Wavelet Type II

(HW TII) Transform and Kekre’s

Wavelet (KW) Transform is used

in order to generate Feature Vector.

This process generates feature

vectors for each user by HWTI,

HWTII and KW.

Proposed System Block Diagram Description

• These Feature Vectors are stored inthe database. Further Analyzed byIntra class testing and Inter classtesting, which results in Genuine andforgery data sheets. Final step is toanalyze the performance of theproposed technique for biometricauthentication based on MultiInstance Fusion and MultiAlgorithmic Fusion for TAR, TRRwill be performed on above featurevector. Distance between two facescan be evaluated by evaluating theEuclidean Distance using KNNClassifier .

Proposed Algorithm.

• Step 1: First read MAT file and its face cubes, this gives a composite

Array for 33 Bands of the Facecubes data.

• Step 2: Next read band data for each image. The total 33 Bands of the

face image are available. These bands of the face image are taken at 33

different each image is of 180*220 Pixel sizes [24]. Perform

Normalization on the data, so that the grey levels are in-between (0-255).

Proposed Algorithm contd..

• Step 3: Then these images

are grouped into eleven sub-

bands of three images each.

We are considering 3

components (F,L and

R).Each of which will be

having 4 blocks for 5 levels

of decomposition, this gives

the size of the Featurecount

for 33 bands and 240 values

from Components*Blocks

*Blocks*Levels.

Proposed Algorithm contd.

• Step 4: This results in feature vectors form HWI, HWII and KW

Transforms for each user.

• Step 5: This feature vector database is used for Intra Class Testing

and Inter Class testing, which generates in Genuine (406 rows and 33

columns) and Forgery (5638 rows and 33 columns) results.

• Step 6: These codebooks are the feature vectors of the hyperspectral

face. In this database Front, Left and Right instances of the same face

are captured.

Proposed Algorithm contd.

• If these instances are considered to build a Multi-instance face

recognition system then the 33 columns are grouped into 11 sub

bands (L+R and/or F+L+R) and final set of codebooks is extracted

and stored in the database.

Results & Discussions.

• The results are discussed in two aspects, first the Multi Instance

Analysis and then the Multi Algorithmic Analysis.

• Evaluation metrics such as True Acceptance Rate (TAR), True

Rejection Rate (TRR), Security Performance Index (SPI) and

Performance Index (PI) are evaluated here for comparison purpose.

Euclidean Distance is calculated evaluation for classification.

• The PolyU HSFD is used for testing for the proposed method.

Some of the subjects used for feature vector extraction, intra & inter

class matching. The feature vectors are evaluated and stored in the

database.

• Security Performance Index (SPI) – This is a new parameter proposed

by Dr. H. B. Kekre [25], this parameter indicates how fast the Equal

Error Rate (EER) is achieved.

Results & Discussions contd.

Comparison of PI and SPI for algorithms by Multi Instance Analysis.

BandsPI : HWI PI : HWII PI : KW

F L R LR FLR F L R LR FLR F L R LR FLR

1 64.4 76.2 73.3 76.3 75.8 61 63 66.4 64 66.4 59.5 62.5 64 62 62.5

2 74.8 74.3 76.7 75.8 76.3 64 70.9 68.9 67.4 70.9 64 67.9 69.6 67.7 65.4

3 76.7 75.3 76.2 76.3 75.8 68.4 66.9 62.5 65.9 67.4 65.4 68.9 67.4 67.9 65.9

4 77.2 74.8 65.9 74.3 75.8 65.4 64.9 61 71.4 69.4 65.9 69.6 69.4 73.6 69.4

5 70.3 72.3 73.8 74.8 72.8 68.9 74.3 72.8 74.8 75.3 68.4 71.4 72.8 73.8 72.4

6 74.8 72.3 73.3 72.8 74.3 73.3 72.4 72.4 71.4 71.9 72.4 69.9 73.3 71.9 74.3

7 76.2 74.8 74.3 74.8 75.3 70.9 74.3 73.3 74.8 74.8 68.9 69.9 71.9 72.4 71.4

8 74.8 69.3 71.8 72.1 73.3 69.4 58.8 77.5 70.1 73.8 69.4 64 67.9 63.5 70.9

9 75.3 68.8 73.3 73.3 71.4 73.3 73.3 77.3 77.3 71.4 70.9 74.8 76.3 73.8 71.9

10 70.8 70.3 72.3 69.9 70.4 72.8 73.3 73.3 72.4 72.8 75.8 76.8 76.8 76.3 76.3

11 71.3 72.3 70.8 70.9 70.4 74.3 69.4 65.9 68.4 68.9 75.3 75.3 75.8 75.3 76.3

BandsSPI : HWI SPI : HWII SPI : KW

F L R LR FLR F L R LR FLR F L R LR FLR

1 38.5 22.2 27.3 30 36.4 54.8 57.2 31.3 40 55.6 40 27.3 14.3 25 30.8

2 15.4 7.15 7.15 7.15 14.3 15.8 14.3 15 15 15 12.5 5.89 5.89 5.89 6.25

3 7.7 18.2 25 16.7 16.7 11.1 21.1 21.1 21.1 15 6.25 11.1 15 10.5 11.1

4 25 25 37.5 43.8 35.7 22.7 48.4 62.5 54.6 54.6 10.5 26.3 33.3 34.6 25

5 20 18.2 15.4 15.4 16.6 36.8 17.6 23.5 17.6 31.6 15.4 13.3 14.3 21.4 20

6 16.7 15.4 15.4 15.4 15.4 18.8 12.5 11.8 17.6 11.8 15.4 15.4 16.7 16.6 16.6

7 15.4 13.3 14.3 33 33 12.5 12.5 11.8 11.8 12.5 16.7 15.4 7.69 15.4 15.4

8 21.4 37.5 25 20 25 46.2 46.9 25 41.6 23.5 13.3 13.3 22.2 20 33

9 25 22 20 20 22 18.2 18.2 16.7 16.6 20 16.7 14.3 12.5 33 14.3

10 20 20 18.2 18.2 18.2 16.7 15.4 11.1 20 14.3 12.5 16.7 15.8 15 15.8

11 15.4 15.4 18.8 12.5 33 17.6 22.2 42.4 34.7 28.6 15 15.8 17.4 20 15.8

Results & Discussions contd.

• As shown in fig. the PI of HWII Right is 77.5% which is

more than other PI of front, left and combined front left

right face for HWI and KW as well. Whereas SPI for

HWII Right is 62.5% which is higher than other feature

vectors for other two algorithms as well.

Thank you ! ! !