Sample Paper in ACM Format
Pay Attention on Highlighted Part and format your paper accordingly, Please refer the ACM Template also
Signature Recognition using Vector Quantization
H B Kekre V A Bharadi
Department of Computer Science Department of Computer Science Department of Computer Science
Mukesh Patel School of Technology Mukesh Patel School of Technology Mukesh Patel School of Technology
Management & Engineering, NMIMS Management & Engineering, NMIMS Management & Engineering, NMIMS
University University
Mumbai, India Mumbai, India
+91-9323557897 +91-9819125676
hbkekre@yahoo.com vinu_bharadi@rediffmail.com tanuja_sarode@rediffmail.com
ABSTRACT
Handwritten Signatures are one of the widely used biometric traits
for document authentication as well as human authorization.
Various techniques have been implemented for Automatic
Signature Recognition. In this paper we discuss the application of
vector quantization to the problem of signature recognition.
Vector quantization based methodology is used here to detect
intra and inter-class variations in signatures. Here we discuss a
method for the codebook generation; this method is fast and
simple. We use the codebook to generate a codevector histogram
specific to the signature template. The spatial moments related to
the codevectors are also calculated. These parameters are used to
classify the signature. Feasibility of this technique for signature
recognition is discussed in his paper.
Categories and Subject Descriptors
I.4.7 Image Processing and Computer vision
General Terms
Algorithms, Measurement, Performance, Design,
Experimentation, Security, Verification.
Keywords
Biometrics, Static Signature Recognition, Vector Quantization,
Histogram.
1. INTRODUCTION
The biometrics is most commonly defined as measurable
psychological or behavioural characteristic of the individual that
can be used in personal identification and verification. The driving
force of the progress in this field is, above all, the growing role of
the Internet and the requirements of society. Therefore,
considerable applications are concentrated in the area of electronic
commerce and electronic banking systems and security
applications of vital installations [2]. Signature verification is an
important research area in biometrics which has applications in
person authentication [2][5]. We can generally distinguish
between two different categories of verification systems: online,
for which the signature signal is captured during the writing
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
ICWET’11, February 25 - 26, 2011, Mumbai, Maharashtra, India.
Copyright 2011 ACM 978-1-60558-812-4…$10.00.
T K Sarode
University
Mumbai, India
+91-9820122805
process, thus making the dynamic information available, and
offline for which the signature is captured once the writing
processing is over and, thus, only a static image is
available[1][3][4][6][24][25].
In this paper we deal with Offline signature Verification System.
Here we discuss a new set of parameters that and be used in any
of the signature verification system for classification of the
signatures. This feature set is based on Vector Quantization.
Vector Quantization is a clustering technique mainly used for
lossy image compression [23]. Where we first generate a
codebook having codevectors that represents the image segments
and then the list of codevectors describes the data to be
compressed. For encoding and decoding the codebook serves a
very important role [9-15]. Here we use this approach to classify
the signatures. We have built a codebook that is specific to the
application and then we have used this codebook to generate a
codevector histogram which is used to classify the signatures, we
also used the spatial information related to the codevector.
The paper is organized as follows; section 2 is dealing with the
Vector Quantization, Signature template and Codebook
prerequisites. Section 3 will discuss the codebook generation
process. In Section 4 we discuss the process of formation of
signature codevector Histograms and spatial information of the
codevectors. Section 5 is focused on the classification process.
We discuss the performance analysis & results in section 6 and
conclusion and future scope is discussed in section 7.
2. VECTOR QUANTIZATION
Due to the extensive technology expansion and usage of internet
in all kinds of information processing, and the advancement of
multimedia applications, the need for storage and transmission of
images and videos is increased. The problem with images and
videos is that they require large amount of bandwidth for
transmitting and storage for archiving. A data compression
technique reduces the amount of data needed to represent an
image, so that they can be economically transmitted or archived.
Vector Quantization (VQ) is an efficient technique for data
compression and has been successfully used in various
applications such as index compression [15][18]. VQ has been
very popular in a variety of research fields such as speech
recognition and face detection. VQ is also used in real time
applications such as real time video-based event detection [15]
and anomaly intrusion detection systems, speech data
compression, and image segmentation [9-15].
In vector quantization we segment the image to form a set of
codevectors and then we find the best match for the codevector
from the codebook. The compressed data will consist of the
codevectors index. The better the codebook better is the
compression and error (MSE) [12][18].This technique is as shown
in the Figure 1. Here we show a very crude outline actual
technique has complex codevectors and may have
multidimensional vector space.
Figure 1. Encoding and Decoding of Image using Vector
Quantization
VQ can be defined as a mapping function that maps k-
dimensional vector space to a finite set CB = {C1, C2, C3, ..….,
CN}. The set CB is called codebook consisting of N number of
codevectors and each codevector Ci = {ci1, ci2, ci3, ……, cik} is of
dimension k. The key to VQ is the good codebook. there are
various codebook generation algorithms available in literature [9-
15,18].
In Encoding phase image is divided into non overlapping blocks
and each block is then converted to the training vector Xi = (xi1,
xi2, ….., xik ). The codebook is searched for the nearest codevector
Cmin by computing squared Euclidean distance which is
Minkowaski distance for r =2 as given by equation (1) between
vector Xi and all the codevectors of the codebook CB. This
method is known as exhaustive search (ES).
d ( X i , C min )  min 1 j  N {d ( X i , C j )}
k 3.1 Codevector generation
Where d ( X i , C j )  ( xip  c jp ) 2 (1)
p 1
It is obvious that if the codebook size is increased to reduce the
distortion the searching time will also increase.
This paper is focussed on detection of intra and inter-class
variations in signatures using vector quantization. We have used a
feature vector extraction mechanism based on vector quantization.
Vector Quantization is used here for feature vector generation
rather than the traditional application of data compression. Vector
Quantization is applied to Signature template to generate all the
codevectors, Signature verification is a problem coming under the
area of pattern recognition. The codebook is having the list of
codevectors that can be used to describe best possible match for
the image segments. We use these codevectors to build a
codevector histogram which is a feature vector.
Signature template that is to be segmented is pre-processed
[5][6][8][16][17][20][22]. We are having binary image that is
obtained after pre-processing as shown in Fig. 2.
Figure 2. Original Scanned Signature. & normalized
signature template.
The normalized signature template is having following properties
1. It is having only black and white pixels
2. The signature is only one pixel thick, it having only
single pixel thick runs.
3. The size of template considered here is 200*160 Pixel.
For generation of codevectors we segment the image into 4*4
pixel blocks, total 2000 (50*40) such blocks are generated and
then these blocks are used to generate the codevector histogram.
The codevector histogram is used as a classifying feature; we
match the codevector histograms to evaluate the similarity. This is
vector quantization used for signature template. In the next section
we discuss the codebook design process.
3. CODEBOOK DESIGN
The codebook design process is divided in to three steps
codevector generation, codebook optimization & codevectors
grouping. Here we design codebook for the later use in the
histogram generation. The codebook is designed for all
combination of possible codevectors, it is then optimized by
removing unwanted codevectors and then the codevector
clustering is performed, for this process we have developed
application specific clustering method, which is based on Grey
Code.
Here we have a binary image as inputs, which is the normalized
signature template. We segment the signature template to for 4*4
pixel blocks. Total 216 i.e. 65536 combinations are possible for
this block size; since the signature is thinned it is a single pixel
thick hence we must neglect the blocks which correspond to
thickness of more than one pixel.
Initially we generate a codebook having all the 65536
codevectors. These codevectors are all possible combinations of a
16 bit binary sequence, i.e. it starts for 0 to 65535. For generation
of codevector Histogram we need to form groups of codevectors
which are similar or having minimum difference. This process is
illustrated in Fig. 3. Codevector Clusters or groups are formed by
codebook optimization and grouping process.
 Figure 3. Vector Quantization of a signature template.

Thinning operation is illustrated in Fig. 5 & Fig. 6. This shows a

3.2 Codebook Optimization codevector block and output of thinning module.
The codebook is having all the combinations of a 16 bit binary
word. Two consecutive codevectors may have variations in bit
positions, hence to arrange the codebook in a manner that
codevectors are arranged with minimum bit difference in the
consecutive positions, so that the consecutive blocks are similar,
we rearrange the codebook. The rearrangement is done by sorting
the codebook according to the GREY coding sequence. According
to the property of GREY code two consecutive binary words are
having difference in only one bit. Hence we have arrangement of
codevectors, where consecutive codevectors have minimum Figure 5. Thinning of a 4 X 4 pixel block
hamming distance. Fig. 4 Shows a codevector group arranged as
per the GREY coding and the Intra group distance, which is the
Hamming distance of each codevector with the first codevector.

Figure 6. Graphical representation of thinning of pixel

Figure 4. Codevector Group showing arrangement of
codevectors according to GREY Coding and their Intra
Group Distance (Snapshot of Final Codebook)
The codebook is having all the combinations that correspond to
pixel runs of thickness of more than one pixel. Hence we must
remove such codevectors. The invalid codevectors can be
removed by thinning of the codevector blocks.
We pass each block to the thinning module the output is the
thinned block having only single pixel thick runs, as these are the
possible blocks that can be a part of signature template, since we
have normalized and thinned binary signature template. The
signature is thinned by morphological thinning [7] process. The
block
After thinning operation the invalid blocks are discarded and we
get a set of valid codevectors, we get total 11755 valid
codevectors. Next step is to form group the codevectors to form
vector groups or clusters.
3.3 Codevector Clustering (Grouping)
In order to generate the codevector histogram we group similar
codevectors together to form a group, various combinations are
tried in program code. Here we present grouping of 12
codevectors to form total 980 groups. In the first step of the
process we have sorted the codebook in GREY coding sequence
and thinned the codevectors. The next step is to cluster these final
codevectors, the algorithm is as follows. We are using hamming
distance for the clustering of codevectors; aim is to group
codevectors together so that they have minimum hamming 4. GENERATING CODEVECTOR
distance in a group.
HISTOGRAM
The Clustering Algorithm is as follows: The Histogram generation process is shown in Fig. 3. This shows
the vector quantization process. To generate the codevector
1. Start with first codevector in the codebook
histogram from a signature template we follow the following steps
2. Set initial distance to 1.
1. Segment the signature template into 4*4 pixel Blocks.
3. Search for the codevector in the codebook for match; if
Here we have a 200*160 pixel size template hence we
it is already grouped then go for next vector.
get a total 2000 (50*40) signature segments
4. If codevector satisfying distance criteria is found then make
(Codevector Blocks)
the grouped flag for that vector to 1 and find next
2. For each segment which is 2 dimensional, generate a 1
codevector for matching.
Dimensional codevector, just by rearranging the pixel
5. This process is continued until one group is filled, and then
values in a row, This forms a corresponding codevector
start a new group.
for the segment
6. If full codebook is covered and no vector satisfying the
3. Now for each codevector find its participation group ID
distance criteria is found then increase the distance and
and store this information in an array from the codebook
search again.
groups previously formed.
7. In case of no vector found until you reach upper limit of
4. Now we have total 980 Groups, so start scanning the
distance, start a new group.
group ID of codevector and count the number of
8. This process is continued until all codevectors are covered.
codevector (Frequency) for each Group ID, store the
information in an Array
The participants of group are codevectors with minimum
5. This will form the frequency distribution of the
intergroup hamming distance and hence they represent a set of
codevector in each groups
similar pixel blocks and hence similar template segments. We use
6. To add the spatial Information for Each codevector
this codevector groups to generate codevector Histogram. A
group we find the spatial moments, center of gravity and
typical codevector Group entry is shown in Fig. 7, which lists a
centre of inertia, Gx and Ix respectively, given by Eq. 2
set of 12 codevector groups. We use two dimensional array of the
& 3. This will add the spatial information of codevector
size of (980, 12) to store codevector groups. Next step is to use
in each group [21].
this codebook for generation of codevector Histogram. This is
M
1 1 N
 xi
discussed in detail in next section.
Gx 
M i 1
Gy   yi
N i 1
(2)

n
1 n
1
Ix 
M
x
i 1
2
i
Iy 
M
y
i 1
2
i (3)

7. We have the signature Codevector Histogram, Spatial

information of each codevector, That is specific to the
signature and can be compared
Fig. 8 Shows how the codevectors and their spatial information is
related [21].

Figure 7. Codevector clusters formed after grouping process Figure 8. Codevector Histogram and Spatial information.
 Figure 9. Frequency Vs Codevector Group (Codevector Histogram)

5. CLASSIFICATION PROCESS We consider signature from same as well as different users and
We have codevector histograms and their associated spatial then perform the codevector histogram generation and distance
moments, to evaluate the distance between them we used an calculation. Fig. 10 shows standard test signature 1, test 1, test 2
Euclidian distance based formula described in [18]. This is and Fig. 9 shows codevector histogram standard and test signature
described below, given an encoded image having similar 1.
representation as a text document, image features can be extracted
based on codevectors frequency. The feature vector for signature
6. RESULTS
template I1 and the feature vector for test signature I2 are given We use test signatures from other persons also; these signatures
below, are shown in Fig. 11. Now for Each Signature in test signature
For I1, It is given by I1= {W11, W21, … WN1} (4) group we match the histogram with standard signature shown in
For I2, It is given by I2= {W12 , W22 , … WN2 } (5) Fig 10. We calculate the S-Score by Equation 6 & 7.We also
calculate The Euclidian Distance d [7] for frequency. We present
In the histogram model, Wij = Fij , where Fij is the frequency of Euclidian distance for frequency of codevectors and the moments
group Ci appearing in Ij .Thus, the feature vectors I1 and I2 are of inertia Ix, Iy. The result is shown in Table I
the codevector histograms. The similarity measure is defined in the Table1. Similarity score, Euclidian distance for Frequency,
histogram model, Wij = Fij , where Fij is the frequency of Ci and Vertical and Horizontal Spatial moments of inertia
appearing in Ij . The similarity measure is defined in [18],[19] is
Sign S-Score Euclidian Ix Iy
called as S-score, give by
Distance
1
s ( I 2, I1)  (6) T1 3.2316 15.4596 6418.207 2668.473
1  dis ( I 2, I1)
T2 3.3409 18.8149 5477.747 2687.285
Where the distance function is D1 2.4871 47.7700 7770.988 3255.156
N
| Wi1  Wi 2 |
dis ( I 2, I 1)  
D2 2.5723 19.1572 5806.319 3283.935
(7)
i 1 1  Wi1  Wi 2
D3 2.6023 21.8860 8233.13 3504.646

This formula is used to evaluate the similarity between two D4 2.1200 33.3766 7820.745 3753.719
codevector Histograms, to evaluate the similarity between spatial D5 2.6731 22.2261 7541.738 3199.428
information we use simple Euclidian distance. The testing results
D6 2.8159 19.3132 8500.541 2938.301
are discussed in next section.
D7 2.7948 26.7207 7902.172 3157.305

One thing should be noted is that the similarity Score is Higher is

Figure 10. Standard signature and test Signature T1 &T2
better and the Euclidian distance is lower is better. The results
clearly indicate that the similarity score for the codevector
histograms is higher (S > 3) for signature from same user (T1 and
T2) and it is low (S < 2.8) indicating dissimilarity for signatures
from different user (D1 to D7).
We use the codevector histogram of the signature as a classifying
feature of the signature. To evaluate similarity we use the S-Score
as discussed above well as the Euclidian Distance between the
histograms.
 Figure 11. Signatures used for testing.

We evaluate FAR and FRR for both the methods simultaneously. VQ codevector histogram. Here we can see that the EER for FAR
Besides this we find the performance of the spatial moments of & FRR is 21%.
the Codevectors.
For all these features we have performed total 140 tests each. We
present the summary for the test for
1. VQ Codevector Histogram S-Score - (VQS) (Threshold -3.0)
2. VQ Histogram Euclidian Distance- (VQED) (Threshold – 22)
3. Spatial Moment of Gravity- (SPMG) (Threshold -6400)
4. Spatial Moment of Inertia-(SPMI) (Threshold – 3090)
Table 2. Signature Verification Results for VQ based features
Case VQS VQED SPMG SPMI
Cases that Should be 70 70 70 70
Accepted
Cases that actually 51 61 43 45
Accepted
Figure 12 FAR, FRR Plot for VQ S-Score Distance.
Cases that Should be 70 70 70 70
Rejected
Cases that actually 45 47 41 44
Rejected

Table 3. Performance Metrics for VQ based features

Sr. Parameter VQS VQED SPMG SPMI
1 FAR 20.00 32.85 41.43 37.15
2 FRR 24.14 12.86 38.57 35.72
3 TAR 72.85 87.14 61.43 64.28
4 TRR 64.28 67.15 58.57 62.85

5 CCR 68.57 77.14 60.00 63.57

6 FCR 31.43 22.86 40.00 36.42
6.1 Performance Metrics for VQ Codevector
features
We have evaluated the performance of VQ-Codevector Histogram
based module with metrics such as FAR, FRR [18] [19] [21] for
S-Score, Euclidian distance, the moments Gravity & Inertia. The
results are presented below. Fig. 12 shows FAR, FRR plots for S-
Score. Equal Error Rate (EER) of 22% is reported for FAR FRR.
Fig.13 shows the above mentioned plots for Euclidian Distance of
Figure 13 FAR, FRR Plot for VQ-Euclidian Distance.
The VQ codevector moments are also used as classifying feature
in the signature recognition system. We evaluate the performance
of these features separately the results are as shown in Fig.13 &
15. Figure 16 shows their comparison. The results indicate that the
Codevector Histogram distance based features S-Score and
Euclidian Distance give better classification rates, FAR & TAR.
In the final program we use both the metrics for decision making.
These features have high EER. The codevector moment of gravity
has EER of 60% for FAR, FRR, the moment of inertia has EER of
36% for FAR, FRR.
 using these feature and a classifier as discussed in [4] we have
achieved accuracy up to 95%. The test bed consists of 1000
signatures collected from 120 users, including genuine signatures
and around 200 forged signatures. On this test bed we have
performed total 353 tests for verification mode and 257 tests for
recognition mode. The system is having decision threshold of
60% for both, the signature verification and signature recognition
mode. Out of 353 verification tests 152 tests were for genuine
signatures and 201 tests were for forged signatures. For the
recognition mode we made 135 tests for genuine signatures and
122 tests were for forged signatures. Fig. 16 shows FAR-FRR
plot for signature recognition system for verification mode and the
EER is 3.29%. At selected threshold level of 60 %. We have
achieved final FAR of the system as 2.5% and accuracy 95.08
%.
Figure 14. FAR, FRR Plot for VQ-Moment of Gravity.

Figure 15. FAR, FRR Plot for VQ-Moment of Inertia. Fig. 16. FAR-FRR plot for Signature Recognition System in Verification
Mode
FAR FRR for VQ
7. CONCLUSION
45 In this paper we discussed application of vector quantization
40 technique for the problem of pattern recognition. Here we
35
Parameter Value

30 consider the problem of signature recognition. The normalized

25 FAR signature template is segmented to form a codevector histogram.
20 FRR We use a new codebook generation and optimization approach to
15 design the codebook. We form codevectors groups for the purpose
10
5
of generation of histograms.
0 We have added the spatial information of the codevectors, i.e. the
VQS VQED SPMG SPMI spatial moments to the histograms. We have used existing
Feature distance measurements approach based on Euclidian distance to
evaluate the similarity between the histograms.
The results clearly indicate that the S-Score for the signature from
Figure 16. Performance Metrics calculated for VQ Features same user is higher and the Euclidian distance is lesser as
VQS- VQ S-Score, VQED- VQ Euclidian Distance, SPMG- compared to the signatures from different user. Hence the patterns
Spatial Moments of Gravity, SPMI – Spatial Moments of are classified. We can use this technique to classify the signatures.
Inertia We have achieved the Correct Classification Ratio (CCR) in the
At individual level the VQ based features have low accuracy, this range of 60 to 77%. Here we have not used any training
is mainly because no training was performed and only VQ based mechanism we are just analyzing the feature set over a database
features are considered for classification. These features when collected from 100 persons. This accuracy is improved to 95% by
combined with other feature vector give high accuracy. In [8] using training & combining VQ based features with cluster based
authors have combined these features with cluster based global global features.
feature set, which include,
1. Grid & Texture Information features.
2. Walsh Coefficients of Horizontal and vertical pixel
distributions.
3. Vector Quantization based-Codeword histogram.
4. Geometric Centers of Signature Template.
8. REFERENCES
[1] Abdullah I. Al-Shoshan, “Handwritten Signature Verification
Using Image Invariants and Dynamic Features”, Proceedings
of the International Conference on Computer Graphics,
Imaging and Visualization (CGIV'06) ,March 2006, 0-7695-
2606
[2] A. K. Jain, A. Ross, S. Prabhakar, “An Introduction to
Biometric Recognition”, IEEE Transactions on Circuits and
Systems for Video Technology, Vol. 14, No. 1, January 2004
[3] A. K. Jain, A. Ross, and S. Prabhakar, “On Line Signature
Verification”, Pattern Recognition, vol. 35, no. 12, pp. 2963-
2972, Dec 2002.
[4] B. Majhi, Y. Reddy, D. Babu, “Novel Features for Off-line
Signature Verification”, International Journal of Computers,
Communications & Control Vol.I (2006), No. 1, pp. 17-24.
[5] Bai-ling Zhang, “Off-line Signature Recognition and
Verification by Kernel Principal Component Self-
regression”, Proceedings of the 5th International Conference
on Machine Learning and Applications (ICMLA'06), 0-7695-
2735-3/06, 2006, 4 - 6
[6] B. Fang, Y.Y. Wang, “A Smoothness Index Based Approach
for Off-line Signature Verification”, Document Analysis and
Recognition, 1999, ICDAR '99. Proceedings of the Fifth
International Conference on 20-22 Sept.1999, PP 785 - 787,
10.1109/ICDAR.1999.791905
[7] Jayraman S, Esakkirajan S, Veerakumar T, “Digital Image
Processing”,tata McGraw Hill Publication, 2009, ISBN –
978-0-07-014479-8
[8] H B Kekre, V A Bharadi, “Signature Recognition using
Cluster Based Global Features”, IEEE International
Conference (IACC 2009), Thapar University, Patiala-
Punjab, India. March 2009,
[9] H.B.Kekre, Tanuja K. Sarode, “New Fast Improved
Clustering Algorithm for Codebook Generation for Vector
Quantization”, International Conference on Engineering
Technologies and Applications in Engineering, Technology
and Sciences, Computer Science Department, Saurashtra
University, Rajkot, Gujarat. (India) January 2008.
[10] H. B. Kekre, Tanuja K. Sarode, “New Fast Improved
Codebook Generation Algorithm for Color Images using
Vector Quantization,” International Journal of Engineering
and Technology, vol.1, No.1, pp.: 67-77, September 2008.
[11] H. B. Kekre, Tanuja K. Sarode, “Fast Codebook Generation
Algorithm for Color Images using Vector Quantization,”
International Journal of Computer Science and Information
Technology, Vol. 1, No. 1, pp.: 7-12, Jan 2009.
[12] H. B. Kekre, Tanuja K. Sarode, “An Efficient Fast Algorithm
to Generate Codebook for Vector Quantization,” First
International Conference on Emerging Trends in Engineering
and Technology, ICETET-2008, held at Raisoni College of
Engineering, Nagpur, India, pp.: 62- 67, 16-18 July 2008.
Avaliable at IEEE Xplore.
[13] H. B. Kekre, Tanuja K. Sarode, “Fast Codebook Generation
Algorithm for Color Images using Vector Quantization,”
International Journal of Computer Science and Information
Technology, Vol. 1, No. 1, pp.: 7-12, Jan 2009.
[14] H. B. Kekre, Tanuja K. Sarode, “Fast Codevector Search
Algorithm for 3-D Vector Quantized Codebook”, WASET
International Journal of cal Computer Information Science
and Engineering (IJCISE), Volume 2, No. 4, pp.: 235-239,
Fall 2008. Available: http://www.waset.org/ijcise
[15] H. B. Kekre, Tanuja K. Sarode, “Fast Codebook Search
Algorithm for Vector Quantization using Sorting
Technique”, ACM International Conference on Advances in
Computing, Communication and Control (ICAC3-2009), pp:
317-325, 23-24 Jan 2009, Fr. Conceicao Rodrigous College
of Engg., Mumbai. Available on ACM portal.
[16] H. Baltzakis, N. Papamarkos, “A new signature verification
technique based on a two-stage neural network classifier”,
Engineering Applications of Artificial Intelligence 14 (2001)
95±103, 0952-1976/01/$ - PII: S 0 9 5 2 - 1 9 7 6 (0 0) 0 0 6
[17] J Edson, R. Justino, F. Bortolozzi and R. Sabourin, "An off-
line signature verification using HMM for Random, Simple
and Skilled Forgeries", Sixth International Conference on
Document Analysis and Recognition, pp. 1031-1034,
Sept.2001.
[18] [3] L. Zhu, A. Rao & A. Zhang, “Theory of Keyblock-based
Image Retrieval”, ACM Journal, Volume V, No. N, March
2002, PP 1-32.
[19] [4] L. Zhu, A. Rao & A. Zhang, “Using Thesaurus To Model
Keyblock Based Image Retrieval”, IEEE International
Conference on Multimedia and Expo. 2001, ISBN 0-7695-
1198-8/01
[20] Miguel A. Ferrer, Jesu´s, B. Alonso, and Carlos M. Travieso,
“Offline Geometric Parameters for Automatic Signature
Verification Using Fixed-Point Arithmetic”, IEEE
transactions on pattern analysis and machine intelligence,
vol. 27, no. 6, June 2005
[21] [1] Ordonez, G. Cazuguel, J. Puentes B. Solaiman, C. Roux,
“Spatial-textural Medical Image Indexing based on Vector
Quantization” Proceedings of the 25th Annual International
Conference of the IEEE EMBS, 2003,0,802-805,A,A,GB.
[22] P. Deng, H. Yuan Mark Liao & H. Tyan, “Wavelet Based
Off-line Signature Recognition System”, Proceedings 5th
Conference on Optical Character Recognition and Document
Analysis, 1996, Beijing, China.
[23] R. M. Gray, “Vector quantization”, IEEE ASSP Mag., pp.: 4-
29, Apr. 1984
[24] R. Plamondon, G. Lorette, ‘‘Automatic Signature
Verification and Writer Identification – The State of the Art”,
Pattern Recognition, vol. 4, no. 2, pp. 107–131, 1989.
[25] R. Plamondon, “The design of an On-line signature
verification system”, Theory to practice, International journal
of Pattern Recognition and Artificial Intelligence, (1994)795-
811.