05-07-2012, 03:52 PM
OFFLINE SIGNATURE VERIFICATION AND IDENTIFICATION
USING DISTANCE STATISTICS
OFFLINE SIGNATURE VERIFICATION.pdf (Size: 1.24 MB / Downloads: 77)
Introduction
Research on online signature verication systems is widespread while those on of-
ine are not many. Oine signature verication is not a trivial pattern recognition
problem when it comes to skilled forgeries. This is because, as opposed to the online
case, oine signatures lack any form of dynamic information. Also, the performance
of the two systems is not directly comparable due to this dierence. In the past
some authors have worked on simple forgeries while others have dealt with the ver-
ication of skilled forgeries. Our present work deals with the verication of skilled
forgeries.
Data acquisition
We performed experiments on two dierent databases, namely databases A and
B. Database A was built by us at CEDAR. This consisted of 55 writers with 24
genuine signature samples per writer. To obtain forgeries, we asked some other
arbitrary people to skillfully forge the signatures of the writers in our database.
In this fashion, we collected 24 forgery samples per writer from about 20 skillful
forgers. The input signatures were scanned at 300 dpi in 8-bit gray scale (that
is, 256 shades of gray) and stored in Portable Network Graphics (PNG) format.
Writers were asked to sign in a predened space of 22 inches. Figure 1 shows one
sample for each of the writers from the database A. Figures 2(a) and 2(b) show 5
genuine and 5 forgery samples of a writer (writer 34) respectively.
Feature extraction
Features for oine signature verication using scanned images can be divided into
three types7;13:
(i) Global features that are extracted from every pixel that lies within a rectangle
circumscribing the signature. These features do not re
ect any local, geometri-
cal, or topological properties of the signature, but include transformations,6;18
series expansions,15 image gradient analysis23 etc. Although global features
are easily extractable and insensitive to noise, they are dependent upon the
position alignment and highly sensitive to distortion and style variations.
(ii) Statistical features that are derived from the distribution of pixels of a signa-
ture, e.g. statistics of high gray-level pixels to identify pseudo-dynamic charac-
teristics of signatures. This technique includes the extraction of high pressure
factors with respect to vertically segmented zones (for example, upper, middle
and lower zones)2 and the ratio of signature width to short- or long-stroke
height.17 The statistical features take some topological and dynamic informa-
tion into account and consequently can tolerate minor distortions and style
variations.
Comparison process
Classier design
A methodology for handwriting verication and identication has been recently
described24 and a complete system for handwriting examination known as CEDAR-
FOX system has been recently developed.29 On the same lines, the present signature
identication and verication system can be used in two modes of operation:
(i) Verication, where the goal is to provide a level of condence as to whether a
questioned document and a known document are from the same writer.
(ii) Identication, where the goal is to identify the writer of a questioned document
given a repository of writing exemplars of several known writers.