online digital signature recognition using kernel based ...joics.org/gallery/ics-1282.pdf ·...
TRANSCRIPT
Online Digital Signature Recognition Using Kernel based
SOM Method to enhance its performance and Accuracy
R. Ravi Chakravarthi
Research Scholar in Computer Technology,
Manonmaniam Sundaranar University,
Tirunelveli-627 012, Tamil Nadu, India
Dr. E. Chandra
Professor & Head of the Computer Science Department,
Bharathiar University, Coimbatore - 641046
Tamil Nadu. India
Abstract:
Signature is the world’s most popular authentication function which consists of one’s
own initials which differentiate their identity from others. It’s a unique way of identifying
individuals because each one in this world has a different kind of writing. Every two
signatures in the world differ from each other in terms of the individual's pressure given
while writing the language, the shape of the loops in the signature, the speed at which they
write, and limitless alternate selections. The unique features used in these signatures are
identified by using different algorithms to extract the unique options of each set and classified
using different classification techniques. OCR is the electronic conversion of handwritten
signatures into machine-encoded texts gathered from various sources. The character is
identified through various techniques such as division, highlight extraction, and
arrangement. The various techniques used in the OCR frameworks is optical checking, area
division, preprocessing, division, portrayal, highlight extraction, preparing, and
acknowledgment, and goes on. The clustering of input vectors id done by using a Kernel Self-
Organizing Feature Map (SOM) which uses a neural network for classification and trained
using unsupervised learning. This research focuses on using the Kernel SOM neural network
model to depict the signature verification. A Triple Crown implementation is implied for the
automatic biometric system which depends upon the training set consistency. Similar
signature’s which has a comparable author square measurement, but not identical to each
other disagrees in terms of location, scale and orientation both locally and globally.
Although there is a distinctive fingerprint, it is proved that there is no single square measure
for a signature. This paper insists on applying the competitive neural specification for
checking the information set consistency in a massive private biometric information set
obtained. A neural network is used to validate the consistency of the intra-variability of the
people signatures. A novel replacement democratic neural specification is used to minimize
the rejection error and maximize the share of correct classification for the renowned options
and replacement feature set.
Keywords: K-nearest neighbor’s algorithm, Optical Character Recognition, Kernel Self-
Organizing Map, Image segmentation, feature extraction, and classification.
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org869
I. Introduction
OCR is an automatic identification scheme which relies on machine identification
methods.
The programmedrecognitionmethod is a data capturing technique used to automatically
identify objects, gather knowledge from them, and store the gathered knowledge into a
computer system without any manual intervention. The external knowledge gathered is from
analyzing a series of images, audios or videos. An electrical device is used to convert the
input of images, audios or videos into a digitalreport. It is kept and analyzed for future
purposes. In order to detect and identify the text in the pictures and videos, an Automatic
Data Processing (ADP) system is developed. The ADP allows the system to mechanically
scan the text embedded deep within advanced backgrounds.
The final example can be illustrated using an associate example of identifying what
the ADP system is capable of doing. The ADP system should be always able to answer two
questions “Where? and What?”. The “Where” part determines where the text string is present
in the picture or video.The “What” part determines what the text string says about the picture
or video. In other words, the system mechanically identifies the individual text embedded in
advanced backgrounds and the necessary details such as each and every word in the text is
recognized. Various existingprogrammed recognition methods are reviewed and the OCR’s
location is calculated in between them and how superior is OCR than the techniques is also
calculated. An ordinary PC gains knowledge from the data entered from the keyboard. But
this method is not applicable in various cases. The automated identification network can
function in many cases which are categorized as important. This makes the ADP system
function in various application areas. OCR can be used in various applications. The
applications listed above is mainly used in the past eras[10] for that reason the applications
are not elaborated in details,its just used for understanding. (a) Speech recognition systems is a machine or program that able to recognize the phrases
and words in a spoken language. This language is converted into machine readable format. It
would recognize the voice using the algorithm.The voice is measured depends on the
accuracy and speed.
(b) This oftenness identification is act as a tagging system for the automatic frequency
identification of municipal solid waste stream in reusable elements. This radio frequency
identification technique was widely used by some various organization in different industrial
countries.
(c) The vision frameworks territory unit actualized through the used camera any place the
region of the item unit known by their structure or size. This methodology is generally used
in machines for the distribution of containers. Sort the sort of jug ought to be perceived as
introductory because the amount repaid for a jug relies upon its sort.
(d) The data contains in attractive stripes territory unit wide utilized on credit cards, and so on
a serious extraordinary measure of learning is kept on the attractive tape anyway
exceptionally planned scanners zone unit required and thus the data can't peruse by people.
(e) The Universal Product Code comprises of numerous dim and lightweight lines speaking
to code for partner 11 digit run, 10 of that decide genuine item. The Universal Product Code
is filtered optically once product transfersthrough glass window by aabsorbed beam of light
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org870
of frail force that is cleared crosswise over the glass window in an in all respects uniquely
structured checking design. The reflected light-weight is estimated and dissected by PC.
Because of early institutionalization scanner tags territory unit, these days wide utilized and
speak to a genuine portion of the general commercial center for programmed distinguishing
proof. The Universal Product Code speaks to a novel range that recognizes the product and a
value activity is essential to recover information concerning the worth. The twofold example
speaking to the Universal Product Code takes up a great deal of territory as the tiny quantity
of learning it truly contains. The scanner tags don't appear to be decipherable to people.
(f) The imprinting in ink is particularly utilized at interims stockpile appliance. The nature
zone unit filled up ink that coverssoftly diced attractive components. They arefilled up
counterfeit text styles that are explicitly intended for the applying. Before the characters zone
unit peruse the ink is presented to an attractive motion. This strategy highlights each
character and modifies the location. The characters zone unit peruse by translating the wave
structure got once checking the characters on a level plane. Each character is expected to
have its very own particular wave structure. Even though intended for machine perusing, the
characters region unit still neat to people. Nonetheless, perusing depends on the characters
being composed of ink.
(g) The optical imprint perusing innovation is utilized to enroll in the area of imprints. It's
acquainted with peruse shapes and place the information is given by stamping pre-
characterized options. Such structures are obvious to people. This methodology is affordable
once the info is influenced. It's predefined with a secured assortment of options.
OCR attempts to manage numerous issues with higher than referenced strategies for
programmed ID. They're required once the information is decipherable each to people and
machines. OCR frameworks have engraved an unmistakable section place in example
acknowledgment. Their peculiarity exists in the certain reality that it needn't bother with the
executives of a strategy that produces information. OCR manages the matter of perceiving
optically or printing has been finished while the net acknowledgment is accomplished where
pc identifies the characters as they're drawn. Each manually written and composed characters
could likewise be perceived anyway the exhibition is legitimately needy upon the standard of
information records. A great deal of unnatural the information is, higher is that the exhibition
of the OCR framework. Be that as it may, when it includes aggregates free penmanship
execution of OCR machines stays faulty.
Individual personality confirmation is a significant and every now and again essential
application. There are a few methodologies taken utilizing biometrics [5], [6], [8], [15].
Biometrics perceives an individual reliant on his way of life instead of what he has or what he
understands. That is completely masterminded to physical and direct biometrics. Physical
biometrics relies upon quick estimations of the physical parts, of the human body. Social
biometrics relies upon the estimation of a movement accomplished, by the individual [2]. An
imprint is perhaps the most prepared and most extensively recognized sort of individual
character affirmation, especially in genuine and fiscal issues.. The attributes of a mark can be
partitioned into 2 sorts; i.e., dynamicand static. Thedynamic highlights are those that are
covered up, for example, the weight connected at pen-tip, situation of pen-tip, and so forth.
Mark check procedures dependent on these qualities are alluded to as online mark
confirmation. Then again, static highlights the basic unmistakable highlights, for example,
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org871
the shape. Signature check strategies dependent on these qualities are alluded to as
disconnected mark confirmation. Since the static attributes are noticeable, it is feasible for a
talented falsifier to replicate these qualities. Since the dynamic attributes are not noticeable, it
isn't workable for any individual to duplicate these qualities as they join the essayist's close to
home style.
Programmed signature confirmation is fundamental because of the powerlessness to
recognize signature essentially by the visual investigation. Programmed assessment, however
not flawless, is very solid and effective. Impressive work has been done in the course of the
most recent couple of decades [9], [10], [12]. The most recent decade has realized an
expanded accentuation on utilization of neural systems for this reason [4], [14].
Here, we deliberate acknowledgment of separate dependent on social biometric, explicitly
over his mark. We think about the dynamic attributes of a mark and thus, utilize online
signature acknowledgment.
1.1 Goals
In the examination of content identification and acknowledgment in a convoluted
foundation is impelled by driving edge utilizations of the computerized interactive media
frameworks. These days extra and extra sound and visual data is caught, put away, conveyed
and oversaw in advanced structures. The wide utilization of advanced media records incites a
few new difficulties in versatile data obtaining an immense sight and sound framework for
the executives. Among the first exceptional are:
1. Automatic broadcast annotation: creates a structured, searchable read of archives of
the
broadcast content.
2. Digital media quality management: archives digital media files for economical media
management.
3. Video piece of writing and cataloging: catalogs video databases on basis of content
relevance;
4. Librarydigitizing:digitizescowlofjournals,magazinesandvariedvideosvictimizationadv
ancedimageandvideo OCR.
II. Proposed Method
A run of the mill OCR framework comprises of numerous components as appeared in
Fig. 1 [3, 7]. The digitize Analog archive exploitation an optical scanner in the essential
advance method. When areas containing content are settled each picture is extricated through
the division technique. The separated images are pre-prepared, taking out clamor to
encourage highlight extraction. The personality of every picture is initiate by correlation
separated alternatives with depictions of picture classes got over a past learning area. At last,
talk information is used to re-creation words and amounts of underlying content. These
means are in a matter of seconds gave here. Intrigued persuser will allude [11] for an extra
intricate discourse of OCR framework components.
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org872
2.1 Optical Scanning
In the main sectionof Optical Character Recognition is optical checking. Over the
filtering technique, an advanced picture of the first report is caught. In OCR optical scanners
are utilized that consolidate transport system and detecting gadget that changes over quality
into dark levels.
Fig 1: Flow chart of the proposed Technique
Composed records join dark print on a white foundation. When playing OCR
structure picture is recover into a bi-level high contrast picture. This technique alluded to as
thresholding is achieved on detector to abstain from squandering remember zone and
procedure exertion. The thresholding strategy is essential on grounds that the aftereffects of
acknowledgment are totally excited about the nature of the bi-level picture. A firm limit is
utilized any place dim levels beneath this edge are dark and levels over white. For top
refinement records with the uniform foundation, a pre-picked mounted limit is frequently
adequate. Be that as it may, reports experienced in pursue have rather enormous shifts. In
these cases, a great deal of refined systems for thresholding is expected to get savvy results.
The most straightforward thresholding techniques shift limit adjusting to local properties of
record like refinement and splendor. Be that as it may, such methodologies some of the time
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org873
depend upon structure filtering of the record which needs a ton of memory and procedure
capacity. The stream outline of the proposed strategy is demonstrated as above.
2.2 LOCATION SEGMENTATION
The following OCR portion is area division. Division decides the comprises of an
image. It's important to discover districts of the report that have composed learning and are
recognized by illustrations and figures. Such as, while action programmed email dealing with
wrapping locationhave to be set isolated from elective prints like stamps and friend’s logo’s,
before acknowledgment. When connected to content, division is the separation of characters
or words. The greater part of OCR calculations stage words into detached characters that are
perceived independently. Ordinarily division is performed by investigative each associated
part. This framework is clear to execute anyway issues emerge if characters bit or they're
divided and include numerous components. The most issues in division are: (a) unmistakable
commotion from the content (b) withdrawal of contacting and divided characters (c)
misconstruing designs and unadulterated science with content and a different way. For
intrigued perusers, extra subtleties are reachable in [11].
2.3 PREPROCESSING
Pre-processing is the 3rd section in Optical Character Recognition. The information
figuring on the data procurement kind is exposed to an assortment of fundamental procedure
ventures to make it usable inside the expressive phases of character investigation. The picture
following from the checking strategy may contain a specific amount of clamor. Retribution
on the scanner goals and in this manner the characteristic thresholding, characters could
likewise be filthy or cracked. Smoothing infers each fulfilling and weakening. Filling wipes
out little breaks, holes, and gaps in digitized characters whereas weakening diminishes the
component of the line. The principal regular strategy for smoothing moves a window over a
double picture of the character and applies sure standards to substance of the window. Pre-
preparing conjointly incorporates standardization related to smoothing. The standardization is
connected to get natures of unvaryingscale, inclination, revolution. The accurate pivot is
initiateover its edge. For turned pages and highlights of the content, variations of Hough
rebuild square measure usually utilized for police work slant.
The pre-preparing part in this way intends to create information that is simple for the
OCR frameworks to work precisely. It is a significant action to be achieved before real
information examination. The primary targets of pre-preparing can be directed as [1, 3]:
(a) Noise decrease (b) compressionand (c)normalization of the data.
(a) Noise reduction:
The commotion presented by optical examining gadget or composition
device causes detached line sections, knocks, and holes in lines, stuffed circles, and so
on. The mutilation together with local varieties, misreckoning of corners,
enlargement, and disintegration might be a potential downside. It's important to
dispense with these flaws before the genuine procedure of the data. The commotion
decrease methods are regularly grouped in 3 noteworthy sets [1, 3]: (i) noise modeling
(ii) morphological processes and (iii) filtering.
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org874
(i) Noise will more often than not be evacuated by movement procedures in the
event that it may have been feasible to demonstrate it. Nonetheless,
commotion demonstrating is beyond the realm of imagination in the greater
part of applications. There occurs some offered writing on commotion
displaying presented by optical twisting like dot, slant, and obscure. It's
conjointly feasible to survey the standard of the character pictures and expel
the clamor to a specific level [1, 3].
(ii) In the fundamental arrangement following the morphological tasks to channel
the character picture trade the difficulty in the activity by intelligent activities.
Various morphological activities are intended to append the messed upstrokes,
decay the associated strokes, smooth the shapes prune wild focuses, thin the
characters and concentrate the limits [1, 3]. The morphological activities are
with progress use to expel commotion on the character pictures because of the
nature of paper and ink besides as whimsical hand development.
(iii) Separating means to dispose of clamor and decrease deceptive focuses
ordinarily presented by poor rate of the data obtaining gadget and uneven
composition surface. Differed reflection and recurrence area channels have
been intended for this reason. The elementary arrangement is to tangle a pre-
characterized veil with picture to allot a cost to an image component as
perform of dim estimations of its adjacent pixels. Numerous channels are
intended for smoothing, honing, thresholding, evacuating somewhat finished
or hued foundation and refinement change capacities [1, 3].
(b) Compression:
It is outstanding that old style picture pressure methods change the picture
from space area to areas that are not reasonable for acknowledgment. The pressure for
OCR needs space area methods for saving shape data. The two prevalent pressure
methods utilized are: (i) thresholding and (ii) thinning.
(c) Normalization:
The institutionalization ways intend of dispose the varieties of the
composition and gain institutionalized information. Some of the unremarkably
utilized ways for institutionalization square measure [1, 3]: (i)
skew standardization and baseline extraction (ii) slant standardization (iii)
size standardization and (iv) Contour smoothing.
(i) Skew normalization and baseline extraction: owing to mistakes inside the
checking technique and expressive style the composing could likewise be
marginally inclining or arced at interim the picture. This may offended the
viability of calculations therefore should be recognized and remedied. For
sure, a few characters recognized in accordance with the relative position with
the importance the standard, the methods for benchmark extraction grasp
exploitation the projection profile of picture, closest neighbor bunch, cross-
connection procedure among lines and Hough revamp [1, 3]. An entirely
appalling closest neighbor is utilized for extricating standard of modern
penmanship in huge commotion [1, 3]. At the point when slant recognition the
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org875
character or word is meant the root, turned or extended till the benchmark is
level and retranslated into the visual presentation unit zone.
(ii) Slant normalization: One of the quantifiable elements in contrast hand-
composing structures is that the inclination edges between the longest strokes
in a word and in this manner the vertical heading. Inclination
institutionalization is utilized to standardize all characters to an ordinary sort.
The chief basic procedure for inclination estimation is that the count of the run
of the mill point of near vertical segments. The vertical line segments from the
shapes region unit separated by following chain code components utilizing a
join of 1dimensional channels [1, 3]. The directions of start and completion
purposes of every line component give inclination edge. The projection
profiles zone unit figured for an assortment of edges a long way from the
vertical heading [1, 3]. The edge equal to projection with best positive spinoff
is utilized to see the quantity measure of cover among vertical strokes and
accordingly overwhelming inclination edge. Inclination location is achieved
by partitioning the picture into vertical and flat windows [1, 3]. The inclination
is evaluated bolstered center of gravity of higher and lower 1/2 each window
found the middle value of over every one of windows. A variation of Hough
change is utilized through filtering left to directly over the picture and
calculative projections inside the heading of twenty-one very surprising
inclinations [1, 3]. The most elevated 3 projections for any inclination territory
unit extra and in this way the inclination with the most significant check is
taken due to the inclination cost. At times, the ubiquity frameworks don't
utilize incline amendment and remunerate it all through training stage [1, 3].
(iii) Size standardization is employed to regulate character size to a precise
customary. The OCR ways may utilize for every even and vertical size
standardization. The character is part of an assortment of regions and all of
those regions are one by one scaled [1, 3]. The measurement
institutionalization likewise can be executed as an area of the instructing stage
and in this way, the size parameters are measurable one by one for each
express training data [1, 3]. The example characters are nibbled by a bit
contracted to the ideal size that expands the prominence rate inside the training
data. The word acknowledgment jelly mammoth intra classification varieties
inside the length of words with the goal that they may moreover aid
acknowledgment; it will in general exclusively include vertical tallness
institutionalization or bases the level size institutionalization on size issue
determined for vertical institutionalization [1, 3].
(iv) Contour smoothing disposes of mistakes because of unpredictable hand
movement all through composition. It regularly decreases the quantity of test
focuses required to speak to the content thus expands intensity in enduring
preprocessing procedure [1, 3].
It is to be noticed that the above methods influence the information and may acquaint
sudden contortions with the character picture. Subsequently, these systems may cause the loss
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org876
of significant data about the composition and along these lines ought to be connected with
consideration.
2.4 SEGMENTATION
The preprocessing stage produces a perfect character picture inside the feeling that an
adequate amount of structured data, high pressure, and low commotion on a standardized
picture is acquired. Here character picture is metameric into its subcomponents. The division
is significant because of degree one will attain in the partition of varying lines inside the
characters straightforwardly influences the prominence rate. Interior division is utilized that
separates lines and bends inside the cursively composed characters. Albeit numerous striking
systems have created inside the past and scope of strategies have developed, the division of
cursive characters is partner degree unexplained disadvantage.
2.4.1 Representation
The fifth part of OCR is a portrayal. The picture portrayal plays one in all premier
essential jobs in any acknowledgment framework. Inside the most straightforward case, a
dark level or twofold pictures square measure nourished to a recognizer. Notwithstanding, in
the greater part of the acknowledgment frameworks in order to maintain a strategic distance
from further intricacy and to expand the precision of the calculations, an extra minimized and
trademark delineation is required. For this reason, a gathering of alternatives is removed for
each class that recognizes it from various classifications though staying invariant to
trademark varieties at interims the class [1, 3]. The character picture representation
techniques square measure normally classified into 3 noteworthy gatherings:
(a) international transformation and seriesenlargement(b) applied
mathematics illustration and (c) geometrical and topological statement.
(a) Series expansionand Global transformation: An everlasting sign, for the most part,
contains a ton of information at that point must be imagined for the point of order. This
may be valid for particular approximations of ceaseless flag too. A system to speak to an
image is by a straight mix of a progression of simpler well-plot capacities. The
coefficients of the straight blend offer smaller mystery composing called change or
arrangement development. Disfigurements like interpretation and revolutions square
measure invariant underneath worldwide change and arrangement development[1, 3].
(b) Statistical illustration: The portrayal of a natural picture through the association of
focuses deals with support varieties somewhat. Despite the fact that this sort of
representation doesn't empower the remaking of the main picture, it's utilized for
diminishing the element of the list of capabilities giving rapid and low multifaceted
nature. Some of the primary connected science alternatives utilized for character
delineation are: (i) division (ii) projections and (iii) crossings and distances.
(c) (c)topological andGeometrical depiction: The different worldwide and neighborhood
effects of characters can be tended to by geometrical and topological geographies with
outrageous adaptability to turns and style arrangements. This kind of portrayal may in like
way encode some getting some answers concerning the structure of an article or may give
a little information with respect to what kind of parts make up that object.
All in all, the real objective of portrayal is to extricate and choose a lot of highlights
that augments the acknowledgment rate with minimal measure of components. The
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org877
element extraction and choice are characterized [1, 3] as extricating the most delegate
data from the crude information which limits the inside class design fluctuation while
upgrading the between class design inconstancy.
2.5 Feature Extraction
In the Optical Character Recognition’s sixth component is including with the
extraction. The equitability of highlightingthe extraction also to catch the fundamental
qualities of images. Highlight extraction is acknowledged together of the principal
problematic issues of example acknowledgment. The principal undemanding methods for
depicting a character is by genuine arrangement picture. Another methodology is to remove
sure choices that describe images anyway leaves the immaterial qualities. The strategies for
the extraction of such choices region unit separated into 3 gatherings' viz. (a) conveyance of
focuses (b) changes and arrangement developments and (c) basic examination. The different
groups of alternatives territory units assessed predictable with their clamor affectability,
misshapen, simple usage, and use. The gauges utilized in this investigation are (a) quality
regarding commotion, contortions, vogue variety, interpretation and revolution and (b)
reasonable use as far as acknowledgment speed, execution quality, and autonomy. Some of
the normally utilized component extraction methods zone unit model coordinating and
connection, changes, appropriation of focuses and basic examination. For intrigued perusers
extra subtleties region unit out there in [11].
Another vital assignment identified with highlight extraction is the arrangement.
Characterization is the strategy for recognizing each character and conveyance to that right
character classification. The two essential classes of the grouping approach for the OCR
territory unit call from the earlier and auxiliary methodologies. In call, from the earlier
acknowledgment character portrayal is numerically depicted in the element vector. There
may also be design attributes gotten from the assortment of character that isn't as basically
measured. Here the connection between the attributes may important once choosing class
enrollment. for instance, on the off chance that we as a whole realize that a nature comprises
1 vertical & 1 even stroke, it will be either 'L' or 'T'. The connection into 2 strokes is expected
to differentiate the nature. The chief ways to deal with a call from the earlier
acknowledgment territory unit least separation classifiers, connected math classifiers, and
neural systems. In auxiliary acknowledgment syntactic techniques region unit, the chief
winning methodologies. a top to the bottom discourse of those methodologies is existing in
[3, 11].
2.6 Training and Recognition
In the 7th element of Optical Character Recogination is preparing the
acknowledgment. Optical Character Recognition frameworks broadly utilize the approach of
example acknowledgment which allots an obscure example into a predefined class. The four
common methodologies are examined in the OCR model as proposed in [1, 3]: (a) template
matching (b) Kernel SOM (c) structural methods and (d) statistical techniques. These
methodologies are neither essentially free nor incoherent from one another. Every so often,
an OCR method in one methodology can likewise be viewed as an individual from different
methodologies. In the majority of the above methodologies, OCR procedures utilize either
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org878
all-encompassing or systematic techniques for the preparation and acknowledgment stages.
The all-encompassing technique utilizes top-down methodologies for perceiving the deep
character killing the division issue. The cost for this operational sparing is to compel the issue
of OCR to constrained language. Likewise, because of the multifaceted nature presented by
the portrayal of a solitary character or stroke the acknowledgment precision is diminished.
Then again, the scientific systems utilize a base up methodology beginning from stroke or
character level and going toward creating an important content. The express or certain
division calculations are required for this procedure, adding additional multifaceted nature to
the issue as well as acquainting division blunders with the framework. Nonetheless, with the
collaboration of the division organizes, the issue is decreased to the acknowledgment of basic
disconnected characters or strokes, which can be taken care of for boundless terminology
with grand acknowledgment price.
Among these strategies, the Kernel SOM strategy gives better precision when
contrasted with others. The reference unit stores information for preparing while example
unit comprises information for approval. Here, smart choice unit utilizes a neural system
model; i.e, a Kernel SOM. It is an information perception method that decreases the
components of information using self-sorting out neural systems. Kernel SOM decreases high
dimensional information into a guide of generally 1 or 2 measurements by bunching
comparable information gatherings. This makes it simple to envision and arrange the
information. Kernel SOM neural system is made for the 7 individuals utilizing MATLAB. 4
highlights were removed from every one of 3 unique factors (p, x, and y). Thus, every mark
test is spoken to by an example Fkl of twelve highlights given by
Where k signifies the signer, l sample number, x, y andp the dynamic characteristics and n, a,
d and t the tokensfeatures;.Each signer delivers5 samples for training; these arejoint to
develop a 12 × 5 matrix Fkspecified by
Lastly, data from 7 signers is joint into asingle 12×35 matrix F, which is the reference
data, specified by
The Kernel SOM made before should be prepared to utilize the reference information F.
The length of the preparation relies upon the number of ages. One age of preparing is
characterized as a solitary introduction of all information vectors to the system. The
system is then refreshed by the consequences of every one of those introductions.
Preparing happens until the most extreme number of determined ages have been come to
or the exhibition objective is met. Here, the no of ages was controlled through
experimentation. As the no of pages expanded, rate mistake diminished. For more than
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org879
200 pages, the development was not noteworthy. In this manner, the preparation was
halted at 200 pages.
III. Experimental Results
The algorithm is executed on PC for evaluation purposes. The language used for
composing these projects is MATLAB.The step by step processing of the output image for
the proposed method is shown in figure 2.Figure 2 demonstrates each step in processing the
output image by distinguishing the various characteristics used and how they differ from each
other.
(a) Input Image (b) Resized image
(c ) Thinning Image (d) Output Image
Figure 2: Step by step output for the proposed method
This area analyzes the experimental results and testing done by the Kernel
SOM system for the sample data. The functioningof the proposed
signconfirmationscheme is validated using 2 types of error rates: FRR and FAR. FRR
is also known as Type I error which calculates the percentage rate of rejected genuine
signatures.FAR is also known as Type II Error,which calculates the percentage rate of
acceptance of fakes. The error rates for the proposed Kernel SOM system is evaluated
for calculating the system’s performance for 50 epochs and 200 epochs and the results
are tabulated in Table 1.
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org880
TABLE I
SYSTEM PERFORMANCE
User Epochs = 100 Epochs = 250
TPR (%) FPR (%) TPR (%) FPR (%)
1 0.00 9.66 0.00 7.88
2 12.00 12.00 8.00 1.20
3 22.00 25.00 19.45 22.16
4 1.00 58.00 2.00 28.98
5 21.00 7.87 22.00 5.78
6 28.00 14.30 17.00 0.00
7 0.00 63.00 0.00 56.78
Total 12.00 27.12 9.78 17.54
IV. Comparative Analysis
The comparative analysis of various methods is demonstrated by the graph
given below.
0 5 10 15 20 25 30
OCR 5 25 40 28 49 56 60
KSOM 10 20 50 35 58 70 85
0
10
20
30
40
50
60
70
80
90
Tru
e P
osi
tive
Ra
te
False Positive Rate
ROC Curve Between TPR Vs FPR
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org881
Figure 3: Comparative analysis of different methods
V. Conclusions
This research focuses on building an online sign verification scheme exploiting Kernel
SOMand is evaluated for objective design. The signature possesses three dynamic
characteristics which are calculatedand 5 distinct topographies are extracted from each
characteristic. Then a Kernel SOM neural network is qualified and verified for various input
training data sets. The results obtained from these data sets tend to be fascinating. Changes
need to be made for improving the TPR and FPR by expandingthe amount of reference
sample size and the no of features. A poor choice of selecting the distinct features results in
functioningdeterioration. Here,3rd central moment possessed a high degree of sensitivity and
degradedschemeoperation. In order to expand thesystemimplementation, it was rejectedlater.
Each signature differs based on environment and state from where it is gained. In terms of
capturing all the possibilities from a reference sample set, it is ideal to obtain the reference
sample set from various environments and conditions. The proposed KSOM classification
method gives higher performance in terms of accuracy value with 99% TPR, 89% FPR
compared to 66% TPR, 73% FPR in ACT and 50% TPR,56% FPR in ICDAR datasets
respectively.
References
[1]. Arica, N., Vural, F. T. Y., An Overview of Character Recognition focused on Offline
Handwriting, IEEE Transactions on Systems, Man and Cybernetics – Part C:
Applications and Reviews, 31(2), pp 216–233, 2001.
[2]. Bunke, H., Wang, P. S. P. (Editors), Handbook of Character Recognition and
Document Image Analysis, World Scientific, 1997.
10 20 30 40 50 60 70 80 90 100
OCR 0.15 0.1 0.25 0.4 0.3 0.63 0.68 0.72 0.75 0.79
KSOM 0.1 0.34 0.56 0.62 0.74 0.78 0.83 0.89 0.93 0.99
0
0.2
0.4
0.6
0.8
1
1.2
Acc
ura
cy
Epochs
Performance Between OCR Vs KSOM
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org882
[3]. Chaudhuri, A., Some Experiments on Optical Character Recognition Systems for dif-
ferent Languages using Soft Computing Techniques, Technical Report, Birla Institute
of Technology Mesra, Patna Campus, India, 2010.
[4]. Cheriet, M., Kharma, N., Liu, C. L., Suen, C. Y., Character Recognition Systems: A
Guide for Students and Practitioners, John Wiley and Sons, 2007.
[5]. Dholakia, K., A Survey on Handwritten Character Recognition Techniques for
various Indian Languages, International Journal of Computer Applications, 115(1),
pp 17–21, 2015.
[6]. Mantas, J., An Overview of Character Recognition Methodologies, Pattern
Recognition, 19(6), pp 425–430, 1986.
[7]. Rice, S. V., Nagy, G., Nartker, T. A., Optical Character Recognition: An Illustrated
Guide to the Frontier, The Springer International Series in Engineering and
Computer Science, Springer US, 1999.
[8]. Schantz, H. F., The History of OCR, Recognition Technology Users Association,
Manchester Centre, VT, 1982.
[9]. Scurmann, J., Reading Machines, Proceedings of International Joint Conference on
Pattern Recognition, Munich, pp 1031–1044, 1982.
[10]. Singh, S., Optical Character Recognition Techniques: A Survey, Journal of Emerging
Trends in Computing and Information Sciences, 6 (4), pp 545–550, 2013.
[11]. Young, T. Y., Fu, K. S., Handbook of Pattern Recognition and Image Processing,
Academic Press, 1986.
[12]. Yu, F. T. S., Jutamulia, S. (Editors), Optical Pattern Recognition, Cambridge
University Press, 1998.
[13]. Kai Ding, Zhibin Liu, LianwenJin, Xinghua Zhu, A Comparative study of GABOR
feature and gradient feature for handwritten 17hinese character recognition,
International Conference on Wavelet Analysis and Pattern Recognition, pp. 1182-
1186, Beijing, China, 2-4 Nov. 2007.
[14]. Pranob K Charles, V.Harish, M.Swathi, CH.Deepthi, "A Review on the Various
Techniques used for Optical Character Recognition", International Journal of
Engineering Research and Applications, Vol.2, Issue 1, pp. 659-662, Jan-Feb 2012.
[15]. Om Prakash Sharma, M. K. Ghose, Krishna Bikram Shah, "An Improved Zone Based
Hybrid Feature Extraction Model for Handwritten Alphabets Recognition Using
Euler Number", International Journal of Soft Computing and Engineering, Vol.2,
Issue 2.
Biographies
R. Ravi Chakravarthi, is a Computer Technology doctoral candidate in
Manonmaniam Sundaranar University, Tamilnadu, India, received his Master’s
degree in Computer Technology from Bharathiar University, Tamilnadu, India,
in 1999. His research interests include online digital signature recognition, big
data, Internet of Things, Software Testing Measurements and Project
Management.
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org883
E. Chandra, received her Ph.D. from Alagappa University, Tamilnadu,
India, in 2007. She is a Professor of Computer Science, in Bharathiar
University, Tamilnadu, India. Her research interests include online
digital signature recognition, speech recognition system, neural
networks, fuzzy logic and machine learning. Life Member in reputed
professional bodies like ACM, CSI and ACSIT.
Journal of Information and Computational Science
Volume 9 Issue 8 - 2019
ISSN: 1548-7741
www.joics.org884