You are on page 1of 5

Improved Segmentation through Dynamic Time Warping

for Signature Verification using a Neural Network Classifier

Wan-Suck Lee, N. Mohankrishnan, & Mark J. Paullik


Department of Electrical & Computer Engineering
University of Detroit Mercy, Detroit, MI
mohank@udmercy. edu

Abstract stationary characterization of the signature, wherein its


A segmentation technique based on dynamic time warping spatial evolution was captured by uniformly dividing it into
is investigated as a means of obtaining improved alignment eight even-sized segments, and using an AR model to
between multiple signatures from a writer; the segments represent each segment. However, such a scheme does not
generated through this approach are represented by an allow for the natural variations (inconsistencies) in a
autoregressive model. The signature verification person’s signature from one sample to the next. In fact,
performance is evaluated with a neural-network-based close examination of the incidence of errors revealed that a
classifier, and shown to be a significant improvement over number of them corresponded to signatures that were
previous results obtained by the authors that were based on inconsistently segmented. It was felt that the incorporation
uniform segmentation. of superior segmentation techniques prior to feature
extraction should be an important next step in extending this
work. This would make for tighter clustering of similar
1. Introduction segments in the feature space and possibly result in
enhanced verification performance. The technique of
The field of biometric personal identification appears to dynamic time warping (D’IW) [5] which provides for a non-
be on a steady growth path [1,2]. One variant of it, namely linear alignment of two signatures is very suitable for this
signature verification as carried out by human beings, is purpose and has been employed in this work.
historically one of the oldest means of identity validation This paper thus extends earlier work by presenting new
both for the author of a document or the initiator of a experimental results from a study conducted to evaluate the
financial transaction. Today the electronic collection and performance of a neural network classifier. The classifier
storage of signatures at the point of a financial transaction input consisted of AR coefficients extracted from signature
is an increasingly common phenomenon. While the segments obtained through a DTW-based signature
majority of current systems use these signatures for archival segmentation scheme. The remaining sections of the paper
purposes, it is conceivable that automated authentication are organized as follows. Section 2 details the collection of
can be incorporated in the near future. the signature data base used, Section 3 discusses the
In earlier work [3] the authors have developed a non- preprocessing steps carried out on the signature, while the
stationary autoregressive (AR) model representation for the method used to segment signatures is described in Section
signature, that was used to carry out verification based on 4. The AR modeling of each segment is presented in
conventional distance measures to measure class similarity Section 5 and the neural network architecture used for
or dissimilarity. Subsequently, a neural-network-based verification is described in Section 6. Finally, the
classifier was incorporated [4] that used these same experimental results are discussed in Section 7.
distances as inputs and yielded average false acceptance
(FR) and false rejection (FA) rates of 2.52%and 1.17% 2. The Data Base
respectively. While these results were encouraging, they
only evaluated the potential of the neural network classifier
The writer population was made up of 16 writers, with
for separating clusters established on the basis of distance
each writer providing a tlotal of 150 signatures collected
measures to quantify class separation. A logical follow-up over 15 sessions. The signatures were recorded
would be to use a neural network classifier directly on the dynamically using a Wacaim digitizing tablet that provided
autoregressive parameters (avoiding distances measures a uniformly time-spaced (x,y) coordinate sequence of points
entirely). along the signature at a sampling rate of 205
Furthermore, all the earlier work depended on a non- samples/second and a spatial resolution of 1200 points/inch.

929
0-8186-8821-1198 $10.00 0 1998 IEEE
- t V i
Figure 3. Perpendicular distance sequence

4. Segmentation through DTW


~ ~ ~~~

Dynamic time warping is an algorithm which estimates


Figure 1. Sample digitized signatures
the best geometrical associations between the points of two
similar sequences. It does this by evaluating various
In order to avoid the effects of muscle fatigue, no more than permitted pairings between the points of the two sequences,
20 signatures were recorded in one sitting. Some sample and selecting the best alignment path through these points
digitized signatures are shown in Figure 1. (the allowed pairings form a grid) based on some optimality
criteria and search constraints [ 5 ] .
3. Preprocessing A basic version of the DTW algorithm is explained
below using Figure 4. In this figure the two axes represent
In addition to (x,y) coordinates of points traced along the the sequence indices of the signatures. There are many
signature, the digitizing tablet also indicates whether each
point corresponds to a pendown or penup point. Each
signature is considered to be characterized by all the points
between the first and last pendowns in the execution of the
signature.
The (x,y) coordinate sequence corresponding to each
signature is converted into a 1-D sequence using the
following process. First, the signature contour is re-
sampled using just 256 points that are re-positioned along
it at equal-arc-length intervals, so that we have uniform
spatial sampling for each signature. The choice of 256
points was based on a power spectral analysis of signature
sequences which led to the conclusion that 256 samples
were sufficient to represent all salient curve characteristics
in the worst case. The best fit straight line is then found for
the 256 interpolated samples as shown in Figure 2. Next, a i I
1-D sequence s(n) is computed by finding the perpendicular Figure 4. Finding the optimal warping function using
distance of each point to the base line resulting in the Dn/v
sequence shown in Figure 3, which was used in subsequent
processing.
possible pairings of these indices subject to certain
restrictions of monotonicity, continuity, boundary
alignment, and search window width, as expressed by the
following conditions:
a) monotonicity condition
i ( k - 1) I i ( k ) a n d j ( k - 1) I j ( k ) , (1)
b) continuity condition

i ( k ) - i ( k - 1) I 1 and j(k) - j ( k - 1) I 1, (2)

930
c) boundary condition chosen segmentation points. The optimal warping function
that establishes the best correspondence between points of
i(1) = 1, j(1) = 1, & the “misaligned” signature with the “master” signature is
(3) indicated. The segmental points of the “misaligned”
i ( K )= I , j ( K ) = J , signature are established by finding points corresponding to
the pre-selected segmentation points of the master signature
d) search window condition through the optimal warping function. An appropriate
master signature is selected for each writer and uniform
spatial segmentation is used to define its segment
boundaries; these are the sequence points 32, 64,96, etc.

256
where “r” is the maximum permitted search window width. 0 -
The cumulative distance g(ij) between the two
sequences from the beginning of the signatures to point (ij)
is then calculated as
M
A
g ( i - 1, j ) + d ( i , j ) S
T
E
R

g(i - 1, j - 1) + d ( i , j )

256
Here d(i,j) is the distance between the ithpoint of the first
signature sequence and the jthpoint of the second signature
sequence. The initialization of the above recursion is done
by
0 MISALIGNED

Figure 5. Segmentation using the DTW


g(1,l) = d(L1). (6)

The overall distance (also known as distortion) between the 5. Feature Extractioin through AR Modeling
two sequences is then given by
Second order AR model coefficients are estimated from
each segment of the signatures aligned through the DTW
procedure described above using the covariance method.
The AR model for section ‘i’ may be written as:
The optimal warping function (or path) is then found
recursively by starting at point (1,J) and backtracking to the ( n )- ai,
5 ( n ) = s,; (9)
beginning of the signatures. This path is indicated by the
dark line in Figure 4. The above description of the
algorithm does not include nuances such as choice of the
“slope” of the path and a “weighting” function, w(k), which
scales d(ij) in equation (5). We have chosen a “ slope” of
zero and a “weighting” function given by
where ‘p’ denotes the AR imodel order, ai is the local mean,
w(k)=(i(k)-i(k-l))+(j(k)- j(k-1)). (8) ri(n) is a zero mean sequence obtained by removing ai
from si(n),cik are the model coefficients for section ‘i’,
For more details on the significance of the above DTW and w(n) is a random noise sequence with zero mean and
features, readers are referred to [6]. variance.
The basic concept of the segmentation technique
employing DTW can be understood from Figure 5. Here 6. Verification using,a Neural Network
the vertical and horizontal axes represents the spatial axis of
a chosen “master” signature and “misaligned” signature Verification was carried out using a neural network
respectively, and black dots on a master signature indicate classifier. The topology used was that of a multi-layer

931
perceptron as shown in Figure 6. A single hidden layer
made up of 16 neurons was chosen on the basis of trial-and-
error runs. The input to the network consists of 16 nodes
resulting from the use of 8 signature segments and a 2”d FROM WRIER”1’

order AR model corresponding to each segment (8 X 2=16). NN(WRITER “I”)


The output layer was made up of 1 neuron corresponding to 16byl
SEGMWTAL
________
accepting (target “1 ”) or rejecting (target “0”)the claimed
registered writer class. mD WEIGHT AND BIAS

CLAIM OF WRITER’I”
Figure 7. Training and testing procedure

7. Results and Discussion


The segmentation performance of uniform spatial and
DTW-based methods are presented and compared. The
uniform spatial segmentation technique divides the
signature into eight even segments, with each segment
W
INPUT FIRST HIDDEN modeled by an AR model. Examples of these are provided
LAYER LAYER in Figures S(a) and S(b) for two signature samples. In these
Figure 6. Multilayer perceptron with single hidden figures, the black dot denotes the eight points corresponding
layer to the segmentation boundaries of the signatures. As can be
seen the segment boundaries are inconsistent between the
The database was divided into three groups (labeled as two signatures.
A, B, & C) made up of signatures from sessions 1-5,6-10,
and 11-15 respectively; each group thus consisted of 800
signatures (50 signaturedwriter X 16 writers). A flow chart
explaining the training and testing procedure is shown in
Figure 7. Training was carried out using the
backpropagation algorithm using 250 signatures per writer.
These were made up of 100 authentic signatures and 150
random forgeries (the genuine signatures of other writers) 1 0

from groups A & B. A total of 16 neural networks are m m m ya) m m gmo I”

required for the verification task - each writer needs his


(a): Uniform spatial segmentation point (AMR 06/01)
own network. The training procedure results in a unique
set of weights and biases for each writer.
The network was subsequently evaluated using 200 test
signatures per writer from group C - made up of 50 x 10
1 8
authentic signatures and 150 random forgeries - that were I
not utilized in the training process. As shown in Figure 7,
the verification process took place as follows. When an
identity claim was made, the feature vector of the claimant
is input to the neural network corresponding to the claimed
identity. If the output of the network is above the threshold
of 0.5, the identity claim is accepted, otherwise declared as
a forgery. The choice of the threshold determines the 1.m I
1 m 2 m ym m 5om m 7rm gm m 1 m
I

relative magnitudes of FA and FR types of errors - it can be (b): Uniform spatial segmentation point (AMR 06/02)
altered to tradeoff one for the other depending on the
application.

932
xm 4

I I Table 1. Experimental results


2 -

Distances input

I”
t am
U
3x0 m 5 x c eaa ma mn sm 1
-
segmentation)

AR coeff. input
(c): DTW based segmentation points (AMR 06/02) to NN (uniform

Figure 8. Comparison of segmentation results

Now consider the results of DTW-based segmentation segmentation)


using a search window of 20 points. The signatures shown
in Figures 8(a) and 8(c) were the “master” and “misaligned” References
signatures respectively. As can be seen, there is a far
superior correlation between the segmentation points of the [ 11 “Special Issue on Automated Biometric Systems”,
two signatures as compared to uniform segmentation. Proceeding of the IEEE, September 1997.
A window width of 20 (established on the basis of trial [2] Benjamin Miller, “Vital Signs of Identity”, IEEE Spectrum,
and error runs) was found to work very well in our February 1994, pp. 22-30.
experiments in aligning genuine signatures over the entire [3] N. Mohankrishnan,Mark 11. Paulik, and Mohamad Khalil, “On-
database. The optimal window length is related to the Line Signature Verification Using a Nonstationary Autoregressive
Model Representation”, Proceedings of the IEEE International
consistency of the writers. Parizeau and Plamondon [7] Symposium on Circuits and Systems, Chicago, Illinois, May 1993,
mentioned that the effect of window length (within reason) pp. 2303-2306.
is not critical to DTW performance. [4] N. Mohankrishnan, Wan-Suck Lee, and Mark J. Paulik,
The verification results obtained are shown in Table 1. “Multi-Layer Neural Network Classification of On-Line
In this table the first result is reproduced from earlier work Signatures”, Proceedings of the IEEE Midwest Symposium on
in which conventional distance measures were used as Circuits and Systems, Ames, Iowa, August 1996.
inputs to the neural network for purposes of comparison. [5] Makoto Yasuhara and Masatomo Oka, “Signature Verification
Experiment Based on Nonlinear Time Alignment: A Feasibility
The remaining two results pertain to the work discussed in
Study”, IEEE Transactions on Systems, Man, and Cybemetics,
this paper. As can be seen from the first two results (both March 1977,pp. 212-216.
based on uniform segmentation), the use of AR coefficients [6] H. Sakoe and S. Chiba, “Dynamic Programming Algorithm
as the input to the neural network classifier improves FR Optimization for Spoken Word Recognition”, IEEE Transactions
error rates from 2.52% to 0.88% and FA error rates from on Acoustics, Speech, and Signal Processing, February 1978,pp.
1.17% to 0.67% respectively. This clearly indicates the 43-49.
advantages of using the AR coefficients directly as a feature 171 M. Parizeau and R. Plamlondon,“A ComparativeAnalysis of
set. The last two results while both using AR coefficient Regional Correlation,Dynamic Time Warping, and SkeletalTree
Matching for signature Verification”, IEEE Transactions on
inputs to the neural network, are based on uniform and
Pattern Analysis and Machine Intelligence, July 1990, pp. 710-
DTW segmentation respectively. When improved 717.
alignment between signatures is implemented with the latter
technique, the FR and FA error rates decrease even further
to 0.25% and 0.25% respectively. It is quite evident that the
segmentation technique using DTW has significant
benefits.

933

You might also like