19-07-2013, 04:56 PM
OFFLINE SIGNATURE VERIFICATION
OFFLINE SIGNATURE.pdf (Size: 581.95 KB / Downloads: 97)
INTRODUCTION
The problem of signature verification is nothing new – it has been a topic of intense
research and experimentation for many years now. But this idea of signature verification
has received a fresh new boost only in the recent years after the advent of computers and
the internet. It finds application in a large number of fields starting from online banking,
passport verification systems to even authenticating candidates in public examinations
from their signatures (e.g. in JEE).
Broadly speaking, signatures are either verified online or offline. In online signature
verification, the event of putting the signature is recorded as a function of time. In this
case, since the signature is being tracked as it is being generated, it is possible to generate
more information from it such as the pressure being applied on the pen or paper, the
speed at which the signature is being put, and so on. However, this method has an
obvious drawback – not all signatures we need to compare can be recorded as they are
being signed. On the other hand, offline signature verification compares two signatures
after they have been put, using the scanned images of the two signatures as input.
Although some information may be lost, the wide number of applications to which this
method can be put more than compensates for its drawback.
COLLECTION OF SIGNATURES AND PRE-PROCESSING
The first phase in this project was the collection of signatures and their preprocessing.
Signatures were collected from twenty (20) persons from IIT Guwahati on A4 sheets.
Each person was requested to sign twelve times on the sheet. These sheets were later
scanned and all the twelve signatures for each person were cropped and separated. The
twelve signatures thus obtained for 20 people were stored in separately labeled folders,
with the signatures in each folder being numbered from s1 to s12.
Introduction to VPP and HPP
VPP (Vertical Projection Profile) is nothing but the sum of all pixels at each x
coordinate, plotted versus x. On the other hand, HPP (Horizontal Projection
Profile) is the sum of all pixels at each y coordinate, plotted versus y.
It must be kept in mind that after the pre-processing described before, each pixel
containing a signature has the logical value of 1. Thus, for a particular x
coordinate, whenever the values of all pixels along y are summed up, it gives an
estimate of the intensity of the signature at that x coordinate. Thus, by similar
reasoning, we can conclude that the VPP and HPP profiles qualitatively describe
the distribution of the signature along the horizontal (x) and vertical (y) axes
respectively.
False Rejection
Suppose that it is known for a fact that a given signature has been signed by a
particular person, that is, it is genuine. However, if the system refutes this clam
and rejects this signature as not belonging to that particular person, such cases of
rejection are termed as false rejection. In this case, a total of 10 signatures from
each user were used to build the feature vector. The remaining five signatures of
every user (that were collected at a later date) were used for testing the
performance of the system of grounds of false rejection. The five signatures were
tested based on the feature model that had been developed for that particular user
and any rejection was counted as a false rejection. Thus, on the whole for 37
users, there were a total of 185 comparisons made for testing for false rejection.
Finally, false rejection ratio was calculated by dividing the total number of false
rejections by the total number of comparisons (185).
FUTURE WORK
This project has been conducted using a database of 37 users with 10 specimens per user.
Increasing the number of specimen signature per user would increase the accuracy of the
results, as the feature models grow more accurate. Furthermore, since we have used a
probabilistic measure i.e. the coefficient of correlation to compare the DCTs, increasing
the number of specimens would have a salutary effect on the accuracy of the most
important test. In this study, we have used VPP and HPP to gather information in the
spatial domain and DCT to do the same in the transform (frequency domain). Further
refinements could include finding and isolating the individual letters in a signature and
then comparing these, as well as checking the tilt (i.e. rotation of the signature).