Vous êtes sur la page 1sur 16

KERNEL BASED ARTIFICIAL NEURAL NETWORK

TECHNIQUE TO ENHANCE THE PERFORMANCE AND


ACCURACY OF ON-LINE SIGNATURE RECOGNITION
1
R. RAVI CHAKRAVARTHI Research Scholar in CSE,
Manonmaniam Sundaranar University,
Abishekapatti, Tirunelveli, Tamil Nadu, India
- 627 012.
email: vijayaraviphd30@gmail.com
2
Dr. E. CHANDRA ,
Professor & Head - Computer Science Department,
Bharathiar University,
Coimbatore - 641046 Tamil Nadu.
Email: crcspeech@gmail.com
Abstract check acknowledgment have a place with this
gathering. Social biometrics acknowledgment
Signature recognition is a standout amongst depends on estimations and information got
the most essential biometrics verification from an activity performed by the client, and
strategies, is an indispensable piece of current consequently by implication measure a few
business exercises, and is considered a attributes of the human body. Be that as it may,
noninvasive and non-undermining process. For this classification is very fake. For example,
online signature recognition, many techniques the speech signal relies upon conduct
had been presented before. Nevertheless, attributes, for example, semantics, word usage,
accuracy of the recognition system is further to elocution, quirk, and so forth (identified with
be improved and also equal error rate is further financial status, training, place of birth, and so
to be reduced. To solve these problems, a on.) [4] [5]. In any case, it likewise relies upon
novel classification technique has to be the speaker's physiology, for example, the state
proposed. In this paper, Kernel Based Artificial of the vocal tract. Then again physiological
Neural Network (KANN) is presented for attributes are additionally affected by client
online signature recognition. For experimental conduct, for example, the way in which a
analysis, two datasets are utilized that are client shows a finger, takes a gander at a
ICDAR Deutsche and ACT college dataset. camera, and so on.
Simulation results show that, the performance
of the proposed recognition technique than that Signature recognition has a place with this last
of the existing techniques in terms of accuracy class, and as indicated by piece of the pie
and equal error rate. reports [6] it is the second most critical inside
this gathering, simply behind discourse
Keywords: Online signature recognition, acknowledgment and over keystroke, walk,
Kernel Based Artificial Neural Network signal, and so forth. Signature
(KANN), accuracy, equal error rate. acknowledgment can be part into two
categories that are static and dynamic. In static
1. Introduction
mode, clients compose their mark on paper,
Individuals acknowledgment by methods for digitize it through an optical scanner or a
biometrics [1– 3] can be part into two camera, and the biometric framework
fundamental classes that are Physiological perceives the mark breaking down its shape.
biometrics and Behavioral biometrics. This gathering is otherwise called "off-line"
Physiological biometrics acknowledgment [7]. In dynamic mode, clients compose their
depends on direct estimations of a piece of the mark in a digitizing tablet, for example, the
human body. Unique mark, face, iris and hand- gadget known as CADIX digitizing tablet
which procures the mark continuously. proposed an online signature verification
Another probability is the procurement by system based on improved Dynamic Time
methods for stylus-worked PDAs. Dynamic Warping which abbreviated as DTW. Using
acknowledgment is otherwise called "on-line" this algorithm, distance score between e
[8]. Dynamic highlights [9] are elements of genuine enrolled signatures and test signature.
time and static highlights are time The optimal warping path was constructed by
autonomous. The utilization of pen elements placing constraints between the pairs of the
over state of mark would be increasingly sample points of the signatures. The authors
helpful in phony recognition since dynamic had aimed to discriminate the genuine and
highlights of a mark are not promptly forgery signatures of a user, especially, when
accessible to counterfeiter as in the state of their values are very close. To achieve their
offline signature. For online signature goal, they had presented a novel scheme of
recognition, many techniques such as ANN scoring/voting the aligned pairs in the warping
and SVM had been presented before. Those path by a set of code-vectors constructed from
techniques seemed to improve the a VQ step. Also they had reduced the equal
performance of recognition. Nevertheless, error rate of the verification system by
accuracy of the recognition is further to be presenting the incorporation of contextual
improved. information in the formulation.

Contribution of this proposed approach is Muhammad Imran Razzak and Bandar


described as follows. Alhaqbani [11] had proposed a multi-section
vector quantization based online signature
 For online signature recognition, verification system. The authors had aimed to
Kernel Based Artificial Neural improve the accuracy of the recognition
Network (KANN) Technique is system. To achieve their aim, they had utilized
presented in this paper. multi-section codebooks by splitting the
 Experimental analysis of this proposed
signature into several sections with every
approach is done with the ACT
section having its own codebook. By
College and ICDAR datasets.
presenting this proposed recognition system,
 The performance of this proposed
the authors had improved the accuracy with
recognition method is evaluated in
least equal error rate.
terms of ROC accuracy and equal
error rate. Rajib GhoshPradeep Kumar and Partha Pratim
Rest of this paper is organized as follows. Roy [12] had proposed an innovative
Section 2 surveys some previous literatures biometric approach for online handwritten
which focused on the research of online signature recognition and verification system.
signature recognition. Section 3 summarizes The authors had aimed to improve recognition
the background of this research work. Section system reliability. To achieve their goal, they
4 presents the proposed online signature had presented Dempster–Shafer theory based
recognition based on kernel-Artificial Neural classifier combination. This Dempster–Shafer
Network. Results of this proposed approach was utilized for combination of different
are discussed in section 5. Finally this paper is information sources. For classification, the
concluded with section 6. authors combined two different classifiers such
as Support Vector Machine and Hidden
2. Related works
Markov Model. The authors had improved the
In this section, online signature recognition accuracy and reliability of the system by
based literatures are survived. Abhishek presenting this proposed approach.
Sharma and Suresh Sundaram [10] had
Mehr Yahya Durrani, Salabat Khan1 and acceptance rate as well as false recognition
Shehzad Khalid [13] had proposed an rate.
innovative method for online signature
recognition that was known as VerSig. The Sudhir Rohilla1 and Anuj Sharma [15] had
authors had aimed to improve the overall presented an innovative algorithm for online
accuracy of prediction. They had achieved signature recognition. The authors had aimed
their aim by presenting this new signature to reduce the false acceptance rate and false
verification method which was based on recognition rate for the proposed recognition
creation of a signature envelope by employing system. To achieve their aim, they had
dynamic time warping method. This proposed presented two class supportive vector machine
scheme utilized the general attributes of such which is abbreviated as SVM for signature
as X, Y coordinates of the signature. By recognition. By presenting this proposed
presenting this proposed approach, the authors approach, the authors had reduced the average
had offered important enhancement in false acceptance rate and false recognition rate
recognition system. to minimum level.

Li Yang et al [14] had proposed a feature 3. Background of the research


weighting algorithm relief based Online An ANN is framed from many single units,
handwritten signature verification. This counterfeit neurons or preparing components
proposed approach aimed to increase the (PE), associated with coefficients (weights),
stability of the dynamic characteristics. To which constitute the neural structure and are
achieve this goal, the authors had proposed an composed in layers. The intensity of neural
innovative relief based writer-dependent calculations originates from associating
online signature verification technique. This neurons in a system. Figure 1 shows the
proposed scheme included two steps that are different layers of a neural network.
training and testing step. In training step, the Fundamentally, there are 3 distinct layers in a
authors had chosen a signature as the base neural system Information Layer (All the
signature from real signatures and more stable sources of info are bolstered in the model
combined features were selected based on the through this layer). Shrouded Layers (There
Relief algorithm. In the testing step, can be in excess of one concealed layers which
classification was done using the K-nearest are utilized for preparing the sources of info
neighbor. By presenting these proposed got from the information layers). Yield Layer
approaches, the authors had reduced the false (The information subsequent to handling is
influenced accessible at the yield to layer).
Figure 1: Different layers of a neural network

The shrouded layer is the accumulation of V - Set {(i, j) |I, j ∈ N}, denotes
neurons which has enactment work connected connections between neuron i and neuron j.
on it and it is a moderate layer found between
the information layer and the yield layer. Its W (function) - V → R defines the weights,
activity is to process the sources of info got by where w((i, j)) are the weights of the
its past layer. So it is the layer which is
connection between neuron i and neuron j.
capable separating the required highlights
from the info information. Numerous looks The weights can be implemented in a square
into has been made in assessing the quantity of weight matrix W.
neurons in the shrouded layer yet at the same
time none of them was fruitful in finding the In this weight Matrix
precise outcome. Additionally, there can be
 Row number of the matrix indicates -
various shrouded layers in a Neural Network.
beginning of connection
So you should imagine that what number of
shrouded layers must be utilized for which sort  Column number of the matrix
of issue. The yield layer of the neural system indicates - target neuron.
gathers and transmits the data in like manner
in way it has been intended to give. The In the matrix, if the numeric value is zero
example displayed by the yield layer can be which indicates a non-existing connection.
specifically followed back to the information 3.1. Propagation function and network
layer. Input
A neural network is a sorted triple (N, V, w) Let I = {i1,i2,i3….,in}be the set of neurons, such
Where that ∀ z
¿{1,....,n}:∃ω iz , j .
N - Set of neurons
Then the network input of j, called netj, is Once a system has been organized for a chosen
calculated by the propagation function as application, that system is set up to be
follows: prepared to start this method the underlying
weights zone unit picked. At that point, the
net j = fprop(Oi1, Oi2,,……… (1) preparation, or learning, starts. There region
Oin,
ωi 1, j ωin , j unit 2 ways to deal with instructing - directed
,……… )
and unattended. administered training includes
Here the weighted sum is more famous: The an instrument of furnishing the system with
multiplication of the output of each neuron i the required yield either by physically

by
ωi , j , and summation of the results:
"reviewing" the system's execution or by
giving the required yields the sources of
net j (2 information unattended training is wherever
) the system should be of the information
=∑ ( Oi .
i ∈I
ω i,j ) sources while not outside encourage.
The Iterative Learning Process: A key element
The activation function finds the activation of of neural systems in K-ANN is Associate in
a neuron dependent on network input and Nursing unvaried learning technique inside
threshold value. which data cases (columns) are given to the
system each one in turn, and along these lines
At a certain time – the activation aj of a neuron the weights identified with the info esteems
j depends on the previous activation state of are balanced unfailingly. Despite everything
the neuron and the external input. cases are given, the strategy typically begins
3.2. Activation function once more. All through this learning area, the
system learns by modifying the weights
Let j be a neuron, then the activation function accordingly on be prepared to foresee the
is shown as: correct class mark of info tests. Neural system
learning is also named as "connectionist
a j(t) = fact (3) learning," because of associations between the
(netj(t), aj(t - 1), θ j ¿ units. Advantages of neural systems in K-ANN
encapsulate their high resilience to clangorous
This converts the network input netj, along
data, moreover as their capacity to order
with previous activation state aj(t - 1) into a
designs on that they require not been prepared.
new activation state aj(t), with the threshold
The chief standard neural system recipe is
value θ playing a vital role. Unlike the
back-engendering equation anticipated inside
other parameters within the neural network the
the 1980's.
activation function is defined universally for
Once a system has been organized for a chosen
all neurons or at least for a set of neurons and
application, that system in K-ANN is set up to
only the threshold values will be varied for
be prepared. To start this method, the
every neuron. Learning procedures helps to
underlying weights (depicted inside the
change the threshold values. Hence the
following segment) are picked all over. At that
threshold values can be related to time as
point the training, or learning, starts. The
θ j(t ). Transfer function is another name system forms the records inside the training
for activation function. data each one in turn, abuse the weights and
capacities inside the shrouded layers, at that
4. Kernel Based Artificial Neural Network
point thinks about the resulting yields against
Technique for On-Line Signature
the predetermined yields of the K-ANN
Recognition
method. Blunders are then proliferated back
4.1. Proposed K-ANN through the framework, dispensing the
framework to direct the weights for application (b) Normalization: The standardization K-
to succeeding record to be handled. ANN strategies intend to evacuate the varieties
This technique occurs again and again of the composition and acquire
in light of the fact that the weights K-ANN institutionalized information. A portion of the
much of the time changed. All through the usually utilized strategies for standardization
instructing of a system indistinguishable are:
arrangement of information is prepared over (i) Skew standardization and bench
and over in light of the fact that the alliance signature extraction
weights are much of the time refined. Note (ii) Slant standardization
that a few systems ne'er learn K-ANN. This (iii) Size standardization and
may be because of the information record don't (iv) contour smoothing.
contain the specific information from that the (i) Skew institutionalization and bench
predetermined yield springs. Systems signature extraction: Due to botches in the
moreover don't unite if there's insufficient data checking K-ANN methodology and forming
to change finish learning. style the synthesis may be hardly tilted or
The fundamental thought is to twisted inside the photo. This can hurt the
convolute a predefined cover with the picture suitability of the counts and thusly should be
to allocate an incentive to a pixel as an perceived and amended. In addition, a couple
element of the dark estimations of its of characters are perceived by the relative
neighboring pixels. A few channels have been position concerning the standard, for instance,
intended for smoothing, honing, thresholding, 9 and g. The methods for bench signature
expelling somewhat finished or shaded extraction consolidate using the projection
foundation and differentiation change profile of the photo, nearest neighbor
purposes. (ii) The fundamental thought behind bundling, cross association procedure among
the morphological tasks is to channel the lines and Hough change. After skew
character picture supplanting the convolution recognizable proof the character or word is
activity by the consistent activities. Different implied the root, turned or reached out until
morphological activities have been intended to the point that the moment that the bench
associate the broken strokes, disintegrate the signature is level and retranslated yet again
associated strokes, smooth the shapes, prune into the show screen space.
wild focuses, thin the characters and (ii) Slant standardization: One of the
concentrate the limits. The morphological quantifiable elements of various penmanship
tasks can be effectively used to expel clamor styles is the inclination point between
on the character pictures because of low nature longest stroke in a word and the vertical
of paper and ink and additionally inconsistent heading. Inclination standardization is
hand development. utilized to standardize all characters to a
(iii) Noise can by and large be expelled by standard frame. The vertical line
adjustment methods in K-ANN on the off components from forms are separated by
chance that it would have been conceivable to following chain code segments utilizing a
display it. Be that as it may, clamor couple of one dimensional channel. The
demonstrating isn't conceivable in a large directions of the start and end purposes of
portion of the applications. There exists some each line component gives the inclination
accessible writing on commotion edge. The projection profiles are processed
demonstrating presented by optical bending, for various points from the vertical course.
for example, spot, skew and obscure. It is The edge relating to the projection with the
additionally conceivable to evaluate the nature best positive subordinate is utilized to
of the character pictures and expel the identify minimal measure of cover between
commotion to a specific degree. vertical strokes and the predominant
inclination point. The inclination location is (c) Compression: It is extraordinary that
performed by isolating the picture into customary picture weight frameworks
vertical and flat windows. The inclination is change the photo from the space territory to
assessed in view of the focal point of gravity regions which are not suitable for
of the upper and lower half of each window affirmation. The weight for OCR requires
found the middle value of over every one of space territory strategies for shielding the
the windows. A variation in K-ANN of the shape information. The two pervasive
Hough change is utilized by examining left weight methodologies used are: (i)
to ideal over the picture and ascertaining Thresholding and (ii) Thinning.
projections toward 21 unique inclinations. (i) Thresholding: so as to diminish capacity
The main three projections for any necessities and to build preparing speed
inclination are included and the inclination usually alluring to speak to dim scale or
with the biggest consider is taken the shading pictures as twofold pictures by
inclination esteem. Sometimes, the picking a limit esteem. The two imperative
acknowledgment K-ANN frameworks don't classifications of thresholding are via
utilize incline remedy and remunerate it worldwide and neighborhood. The
amid preparing stage. worldwide thresholding picks one limit an
(iii) Size Normalization: It is utilized to modify incentive for the whole character picture
the character size to a specific standard. The which is regularly in light of an estimation
K-ANN techniques may apply for both flat of the foundation level from the force
and vertical size normalizations. The histogram of the picture. The nearby or
character is partitioned into number of zones versatile thresholding use distinctive
and every one of these zones is qualities for every pixel as indicated by the
independently scaled. The size neighborhood. A correlation of regular
standardization can likewise be executed as worldwide and neighborhood thresholding
a piece of the preparation arranges and the strategies is given by utilizing an
size parameters are evaluated independently. assessment standard that is objective
Here the two example characters are step by coordinated keeping in perspective of the
step contracted to the ideal size which coveted precision of the OCR framework. It
augments the acknowledgment rate in the has been demonstrated that Niblack's
preparation information in K-ANN. The locally versatile strategy produces the best
word acknowledgment safeguards outcome. A versatile coherent technique is
substantial intra class contrasts in the length produced by investigating the bunching and
of words so they may likewise aid association attributes of the characters in
acknowledgment; it keeps an eye on just debased pictures.
include vertical stature standardization or (ii) Thinning: While it gives an enormous decrease
bases the flat size standardization on the in information measure, diminishing
scale factor figure for vertical concentrates the shape data of the characters.
standardization. Diminishing can be considered as
(iv) Contour smoothing: It takes out the blunders transformation of disconnected penmanship to
because of the flighty K-ANNfor hand relatively online like information with
movement amid the composition. It by and deceptive branches and antiques. The two
large diminishes the quantity of test guides fundamental K-ANN methodologies for
required toward speak to the content and diminishing depend on pixel savvy and non-
subsequently enhances effectiveness in pixel insightful diminishing. The pixel
outstanding pre-preparing steps. insightful diminishing techniques locally and
iteratively process the picture until one pixel
wide skeleton remains. They are extremely
delicate to clamor and distort the state of the division methodologies are partitioned into
character. The non-pixel shrewd techniques three classifications: (a) express division (b)
utilize some worldwide data about the understood division and (c) blended
character amid the diminishing. They deliver a procedures.
specific middle for K-ANN or focus line of the (a) In unequivocal division the sections are
example specifically without analyzing all the recognized in light of character like properties.
individual pixels. The grouping based The way toward cutting up the character
diminishing strategy characterizes the skeleton picture into important parts is accomplished
of character as the bunch focuses. Some through dismemberment. Analyzation
diminishing calculations distinguish the examinations the character picture without
solitary purposes of the characters, for utilizing a particular class of shape data. The
example, end focuses, cross focuses and paradigm for good division is the assertion of
circles. These focuses are the wellspring of general properties of the fragments with those
issues. In a non-pixel shrewd diminishing they normal for substantial characters. The
are taken care of with worldwide accessible-ANN techniques in light of the
methodologies. The cycles for diminishing can dismemberment of the character picture utilize
be performed either in successive or parallel blank area and pitch, vertical projection
calculations. The successive calculations look examination, associated segment investigation
at the form focuses by raster output or shape and points of interest. The unequivocal
following. The parallel calculations are better division can be subjected to assessment
than successive K-ANN ones since they utilizing the phonetic setting.
analyze every one of the pixels at the same (b) The verifiable division technique depends on
time utilizing a similar arrangement of acknowledgment. It looks the picture for
conditions for erasure. They can be segments that matches the predefined classes.
productively actualized in parallel equipment. The division is performed by utilizing the
It is to be noticed that the above systems acknowledgment certainty including syntactic
influence the information K-ANN and may or semantic accuracy of the general outcome.
acquaint surprising bends with the character In this approach two classes of techniques are
picture. Therefore, these systems may cause utilized viz (I) strategies that make some
the loss of vital data about composing and in inquiry procedure and (ii) techniques that
this way ought to be connected with mind section an element portrayal of the picture.
division. The pre-handling stage yields a The top notch endeavors to portion characters
spotless K-ANN character picture as in an into units without utilization of highlight
adequate measure of shape data, high pressure, based dismemberment calculations. The
and low commotion on a standardized picture picture is isolated deliberately into numerous
is gotten. The following OCR part is division. covering pieces without respect to content.
Here the character picture is sectioned into its These strategies begin from plans created for
subcomponents. Division is imperative in light the acknowledgment of machine printed
of the fact that the degree one can reach in words. The fundamental standard is to utilize
detachment of the different lines in the a versatile window of variable width to give
characters specifically influences the groupings of speculative divisions which are
acknowledgment rate. Inward division is affirmed by OCR. The below average of
utilized here which separates lines and bends strategies sections the picture verifiably by
in the cursively composed characters. In spite characterization of subsets of spatial
of the fact that few amazing strategies have highlights gathered from the picture in
created before and an assortment of systems in general. This should be possible either
K-ANN has risen, the division of cursive through shrouded signature of chains or non-
characters is an unsolved issue. The character signature based K-ANN methodologies. The
non-signature approach originates from the Step 6: Use the following activation
ideas utilized as a part of machine vision for function to get the final result for
acknowledgment of blocked question. This every output unit j = 1 to m- (6)
acknowledgment based approach utilizes

{}
probabilistic unwinding, the idea of 1 if y inj >θ
regularities and singularities and in reverse 0 if −θ ≤ y inj ≤ θ
coordinating. −1 if y inj <−θ
(c) The blended K-ANN techniques consolidate f(yin) =
unequivocal and verifiable division in a Step 7: Modify the weight and bias for x = 1 to
hybrid way. An analyzation calculation is n and j = 1 to m as given below
connected to the character picture; however,
the purpose is to over section i.e. to cut the Case 1: - if yj ≠ tj then, (7)
picture in adequately numerous spots with the
end goal that the right division limits are wij(new) = wij(old) + α txj i (8)
incorporated among the cuts made. When this
b(ne
is guaranteed, the ideal division is looked for
w) = b(old) + t
by assessment of subsets of the cuts made.
Case 2: - if yj = tj then, (9)
wij(new) = wij(old)
4.2. Proposed Classification Algorithm: K- (10)
ANN b(n
ew) = b(old)
Step 1: Initialize the below parameters for
simplicity, before starting the training: here ‘y’ is the actual output and
‘t’ is the target output
o Weights (w) = 0
Step 8: Test for the halting condition that will
o Bias = 0 occur when the weight w remains fixed.
o Learning rate ∝=1
o K(x, y) = <f(x), f(y)>
5. Results and Discussions
Step 2: Continue steps 3 to 8 until the stopping
In this experiment, the training dataset
condition is False.
comprises 80 writer’s samples in ICDAR
Step 3: Continue steps 4 to 6 for every training Deutsche and ACT college dataset from a total
vector x. of 40 dissimilar writer’s Signatures. The test
set contains 160 online digital signatures, 80 in
Step 4: Activate each input as follows: English. The training and test data are used as
a function of the tasks and evaluate the system.
xi= si (i = 1 ton) (4)
5.1. Experimental Analysis with
Step 5: Get the net input using the following ACT College Dataset:
equation:
n (5)
∑ x i ωij
yin = b + i + k(x, y)

Here ‘n’ is the total number of input


neurons and ‘b’ is bias.
OCR~TP OCR~FP KANN~TP KANN~FP
R R R R
0.00 0.00 0.00 0.00
0.23 0.10 0.28 0.20
0.34 0.15 0.32 0.30
0.39 0.22 0.39 0.42
0.42 0.29 0.43 0.53
0.47 0.31 0.50 0.58
0.53 0.38 0.53 0.62
0.56 0.45 0.58 0.68
0.58 0.51 0.60 0.71
0.60 0.55 0.66 0.73

Figure 2: ROC - Accuracy of OCR with


KANN – ACT Dataset
The ROC curve is created for checking
whether the classifier is good or bad and
plotting the TPR against the FPR at different
threshold values. The input signatures are
given to the system and it recognizes a genuine
signature and result obtained as positive then
it’s called True Positive. If the input is
recognized as a forgery signature and result Figure 3: Performance analysis for OCR and
obtained as negative, then it’s called False KANN – ACT Dataset
Positive. In this experimental analysis, True The genuine, skilled, unskilled and random
Positive Rate is 0.66 and False Positive Rate is forgery signatures are used for training and
0.73 for K-ANN classifier and TPR, FPR is testing. In this experiment, the skilled,
0.6, 0.55 respectively for OCR method. unskilled and random forgery signatures are
Epoch Training- Testing Training Testing identified based on the dynamic features using
a threshold value of 0.5 in the 0-60 epochs.
s KANN -KANN -OCR -OCR
0 0.00 0.00 0.00 0.00
Based on the dynamic features, the genuine
20 0.26 0.22 0.18 0.15 signatures are recognized from 80 th to 180th
40 0.34 0.31 0.24 0.25 epochs.
60 0.38 0.36 0.29 0.33 ACT college Dataset – Sample:- KANN
80 0.44 0.43 0.33 0.37
100 0.54 0.52 0.37 0.41
120 0.56 0.55 0.42 0.45
140 0.58 0.57 0.48 0.47
160 0.60 0.59 0.51 0.50
180 0.63 0.61 0.55 0.52
Figure 4.A: Genuine Signature

Figure 4.E: Unskilled Forged Signature


recognized as GENUINE by KANN

The input layer accepts all the signatures with


its features and converts it into the triplet
Figure 4.B: Random Forged Signature matrix, which propagates the values to the
recognized as GENUINE by KANN input layer to hidden layer and hidden layer to
output layer. The target values compared with
actual values gathered from this process i.e.,
input signatures were compared with genuine
signatures, if the equal error rate is high, adjust
the weights from the output layer to hidden
layer and hidden layer to input layer. Based on
the weight adjustment and a threshold value >
0.5, the genuine signatures, Random forged,
Skilled Forged and Unskilled Forged are
Figure 4.C: Random Forged Signature recognized by the system.
recognized as FORGED by KNN
In this experimental analysis with ACT college
dataset, Figures shown in 4.4. A was
recognized as genuine signatures. The system
recognized the Randomly Forged Signatures
[RF] as genuine signatures based on the
threshold values, which is Shown in Figure.
4.4. B and some signatures were recognized as
Forged a signature, which is shown in Figures
4.4. C. Some signatures were recognized by
the system as Skilled Forged signatures [SF] as
Figure 4.D: Skilled Forged Signature
recognized as GENUINE by KANN genuine signature which is shown in Figure
4.4. D and some signatures were recognized
Unskilled Forged signatures [USF] as genuine
signatures, which is shown in Figure 4.4. E.
model for better accuracy, ROC curve must
have the maximum value nearer to 1. The
signatures are recognized based on the existing
features during the time of signing by applying
back propagation technique. The weights are
adjusted based on the errors at the time of
propagation.

Table 4: Performance of KANN and OCR


with ICDAR Dataset

Epochs Training Testing Training Testing


-KANN -KANN -OCR -OCR
5.2. Experimental Analysis with ICDAR
0 0.00 0.00 0.00 0.00
Dataset: 20 0.18 0.19 0.15 0.13
40 0.21 0.22 0.16 0.17
Table 3: ROC Accuracy Comparisons for 60 0.26 0.28 0.19 0.20
OCR and KANN – ICDAR 80 0.31 0.32 0.21 0.23
100 0.38 0.43 0.25 0.28
OCR~ OCR~ KANN~ KANN~ 120 0.42 0.51 0.31 0.34
TPR FPR TPR FPR
140 0.46 0.53 0.35 0.38
0.00 0.00 0.00 0.00
0.17 0.10 0.18 0.20
160 0.53 0.56 0.41 0.42
0.23 0.13 0.23 0.26 180 0.57 0.59 0.45 0.45
0.28 0.21 0.27 0.30
0.31 0.26 0.31 0.35
0.34 0.29 0.36 0.41
0.39 0.32 0.41 0.46
0.42 0.37 0.44 0.50
0.45 0.42 0.47 0.53
0.47 0.46 0.50 0.58

Figure 5: ROC - Accuracy of OCR with


KANN – ICDAR Dataset
Figure 6: Performance analysis for OCR and

The figure 2 and 5, shows the ROC curve KANN – ICDAR Dataset

acquired at various threshold values for both Figure 3 & 6 shows the various epochs against
ACT and ICDAR dataset. Also tables 1 and 3 the accuracy values. Also tables 2 and 4 show
show shows the ROC curve acquired at the various epochs against the accuracy values.
various threshold values for both ACT and There is a threshold value used for
ICDAR dataset. To evaluate the classifier recognizing the signatures based on the
dynamic features. The threshold ( θ ) is Figure 7.C: Random Forged Signature
always denoted as a unit interval. It reveals recognized as FORGED by KANN
that the nearer θ is a decision criterion
value (threshold ( θ ) ≥ 0.5), which can be
used for recognizing genuine, skilled,
unskilled and random forgery signatures.
2) ICDAR-SigComp2012 Dataset –
Sample:1- KANN

Figure 7.D: Skilled Forged Signature


recognized as GENUINE by KANN

Figure 7.A: Genuine Signature

Figure 7.E: Unskilled Forged Signature


recognized as GENUINE by KANN

In this experimental analysis with ICDAR dataset,


Figure shown in 7.A was recognized as genuine
Figure 7.B: Random Forged Signature
recognized as GENUINE by KANN signatures. Based on the Threshold value the system
recognized the Randomly Forged Signatures [RF] as
genuine signatures, which is Shown in Figure. 7.B and
some signatures were recognized as Forged signatures,
which are shown in Figure 7.C. Some signatures were
recognized by the system as Skilled Forged signatures
[SF] as genuine signatures which are shown in Figure
7.D and some signatures were recognized Un Skilled
Forged signatures [USF] as genuine signatures, which
are shown in Figure 7.E.
Equal Error Rate (EER) of the proposed system
derived from ACT College and ICDRAI datasets when
5.3. Equal Error Rate in Training and Testing with
different number of samples was used.
ACT and ICDAR Dataset:
Table 5: EER in Training phase with ACT and ICDAR Datasets

Number of EER –SF EER - USF EER – RF


Training
ACT ICDAR ACT ICDAR ACT ICDAR
samples
2.57 3.27 2.68 2.98 1.89 2.32
2
4.28 5.32 3.49 4.35 5.62 6.12
5
7.84 8.13 5.14 6.24 6.73 7.46
9
8.32 9.46 6.21 7.89 7.94 8.64
11
9.69 10.56 7.58 8.76 8.52 9.53
13
Table 6: EER in testing phase with ACT and ICDAR Datasets

Number of EER –SF EER - USF EER – RF


Testing samples ACT ICDAR ACT ICDAR ACT ICDAR
2.79 3.23 2.13 2.18 1.53 2.46
3
3.82 4.14 3.32 4.2 4.26 4.67
6
7.24 6.85 6.65 6.94 5.47 5.79
10
8.63 8.31 7.34 7.16 6.78 7.11
12
9.12 9.11 8.77 8.49 7.42 7.83
15

The above Tables 5 & 6, shows the EER The premier obvious technique for depicting
values, when the proposed K-ANN classifier an identity is by the specific development
algorithm was used in the various genuine and picture. The results show that the K-ANN is
forged samples of digital signatures in ACT better when compared to OCR. Another
College and ICDRAI datasets. It shows that approach is to separate bound alternatives that
Equal Error Rate (EER) value gets increased additionally portray the images, anyway
with the more number of samples that had forgets the irrelevant properties. The training
been trained and tested. set was prepared using one-against-all
6. Conclusion approach. It was due to the time complexities
involved in the back propagation process. The
The target of highlight extraction is to catch proposed K-ANN classification method gives
the fundamental K-ANN attributes of the lower performance in terms of accuracy value
images, and it's normally acknowledged this is with 66% TPR, 73% FPRin ACT and 50%
frequently one in everything about chief TPR, 56% FPR inICDAR datasets
intense issues of example acknowledgment. respectively.
References Verification", Pattern Recognition
Letters, vol. 80, pp. 84-90, 2016.
[1] Y. Chen, J. Yang, C. Wang and N. Liu, Available:
"Multimodal biometrics recognition 10.1016/j.patrec.2016.05.023
based on local fusion visual features [Accessed 23 December 2018].
and variational Bayesian extreme [8] J. Galbally, M. Diaz-Cabrera, M.
learning machine", Expert Systems Ferrer, M. Gomez-Barrero, A. Morales
with Applications, vol. 64, pp. 93-103, and J. Fierrez, "On-line signature
2016. Available: recognition through the combination
10.1016/j.eswa.2016.07.009 of real dynamic data and synthetically
[Accessed 23 December 2018]. generated static data", Pattern
[2] W. Kang, X. Chen and Q. Wu, "The Recognition, vol. 48, no. 9, pp. 2921-
biometric recognition on contactless 2934, 2015. Available:
multi-spectrum finger images", 10.1016/j.patcog.2015.03.019
Infrared Physics & Technology, vol. [Accessed 23 December 2018].
68, pp. 19-27, 2015. Available: [9] R. Doroz, P. Porwik and T. Orczyk,
10.1016/j.infrared.2014.10.007 "Dynamic signature verification
[Accessed 23 December 2018]. method based on association of
[3] K. Reshmi, P. Muhammed, V. Priya features with similarity measures",
and V. Akhila, "A Novel Approach to Neurocomputing, vol. 171, pp. 921-
Brain Biometric User Recognition", 931, 2016. Available:
Procedia Technology, vol. 25, pp. 240- 10.1016/j.neucom.2015.07.026
247, 2016. Available: [Accessed 23 December 2018].
10.1016/j.protcy.2016.08.103 [10] A. Sharma and S. Sundaram,
[Accessed 23 December 2018]. "An enhanced contextual DTW based
[4] C. YOU and B. MA, "Spectral-domain system for online signature
speech enhancement for speech verification using Vector
recognition", Speech Communication, Quantization", 2018.
vol. 94, pp. 30-41, 2017. Available: [11]M. Razzak and B. Alhaqbani,
10.1016/j.specom.2017.08.007 "Multilevel fusion for fast online
[Accessed 23 December 2018]. signature recognition using multi-
[5] T. Davis, D. Grantham and R. Gifford, section VQ and time modelling",
"Effect of motion on speech Neural Computing and Applications,
recognition", Hearing Research, vol. vol. 26, no. 5, pp. 1117-1127, 2014.
337, pp. 80-88, 2016. Available: Available: 10.1007/s00521-014-1779-
10.1016/j.heares.2016.05.011 6 [Accessed 23 December 2018].
[Accessed 23 December 2018]. [12] R. Ghosh, P. Kumar and P.
[6] C. Vivaracho-Pascual, A. Simon- Roy, "A Dempster–Shafer theory
Hurtado and E. Manso-Martinez, based classifier combination for online
"Using the score ratio with distance- Signature recognition and verification
based classifiers: A theoretical and systems", International Journal of
practical study in biometric signature Machine Learning and Cybernetics,
recognition", Neurocomputing, vol. 2018. Available: 10.1007/s13042-018-
248, pp. 57-66, 2017. Available: 0883-9 [Accessed 23 December
10.1016/j.neucom.2016.11.080 2018].
[Accessed 23 December 2018]. [13] M. Durrani, S. Khan and S.
[7] A. Soleimani, B. Araabi and K. Khalid, "VerSig: a new approach for
Fouladi, "Deep Multitask Metric online signature verification", Cluster
Learning for Offline Signature Computing, 2017. Available:
10.1007/s10586-017-1129-4
[Accessed 23 December 2018].
[14] L. Yang, Y. Cheng, X. Wang
and Q. Liu, "Online handwritten
signature verification using feature
weighting algorithm relief", Soft
Computing, vol. 22, no. 23, pp. 7811-
7823, 2018. Available:
10.1007/s00500-018-3477-2
[Accessed 23 December 2018].
[15] S. Rohilla and A. Sharma,
"SVM Based Online Signature
Verification Technique Using
Reference Feature Vector",
Proceedings of the National Academy
of Sciences, India Section A: Physical
Sciences, vol. 87, no. 1, pp. 125-136,
2016. Available: 10.1007/s40010-016-
0293-x [Accessed 23 December
2018].