30-07-2013, 04:57 PM
Offline Signature Verification Using Critical Region Matching
Offline Signature.pdf (Size: 630.81 KB / Downloads: 69)
Abstract
Signature can be seen as an individual characteristic of a person which, if modeled
with precision can be used for his/her validation. An automated signature verification
technique saves valuable time and money. The paper is primarily focused on skilled
forgery detection. It emphasizes on the extraction of the critical regions which are more
prone to mistakes and matches them following a modular graph matching approach.
The technique is robust and takes care of the inevitable intra- personal variations. The
results show significant improvement over other approaches for detecting skilled
forgery.
Introduction
Signature Verification is the process of recognizing an individual’s handwritten
signatures. Signatures have been by far the most popular means for establishing the
authenticity of individuals. Signature authentication offers a quick, simple and cost
effective means for validating the authenticity of a document by determining the
difference between an original signature and a counterfeit one.
The two most widely used approaches for the verification of the signatures are static
and dynamic. Dynamic or Online Verification analyzes the behavioral biometric of the
handwritten signature while it is being written with the aim of verifying its authenticity,
thus establishing the identity of the user. The user’s presence is vital for the real time
verification of his/her signatures. In contrast, Static or Offline verification obviates the
necessity of user’s presence as it compares the various characteristics of a pre-recorded
signature image on order to reach a desired conclusion
Preprocessing:
In the following, we adopt the following notations. Sets are denoted in bold and Cap
face. Scalar quantities are denoted in regular (non-bold) font. Members of a set are
denoted in small and bold letters of their respective sets, and are indexed. In addition,
the set of x-coordinates and y-coordinates corresponding to a set of points are
represented by subscripting the name of the set, e.g. XA represents the set of X-
coordinates of the point set A
Rotation of Signatures
The accuracy of the results is largely dependent on the rotation algorithm used for
orientation correction. The rotation algorithm should be robust and must produce the
same results for images taken from the same user. The rotation algorithm rotate-image
is given in Figure 3. The binarized and noise cleaned signature image is input to the
algorithm. We use the bottom pixels of a signature image as a template to fit an
orientation line through them using the polyfit function of Matlab® (Mathworks Inc
Ltd) (lines 1-2). The polyfit function is further explained in Section 3.1. Finally, the
cp2tranform () function produces a projective transformation of the input image, using
the slope of the orientation line as a guiding parameter (line 3).
Experimentally, we found that the above algorithm showed excellent parity between
the rotated-corrected transformations of the sample signature and new input signature
from the same user, when the rotation angle varied between -30 to +30 degrees.
Thinning of Signatures
The signatures are thinned in order to reduce the computations required by the graph
matching algorithm. We employed the technique of Thinning by LOCAL coupling
Points [5]. This algorithm works very well on signature images, as it is able to preserve
their intricate details and other geometrical properties. Figure 6 shows the images
before and after for a sample signature.
Extraction of critical points:
A contour based approach is followed to extract the critical points. In this approach
the contour is traversed and any sharp change in the curve is marked as a critical point.
Critical points can be best described as the set of points which model the basic structure
of the signature. They are a minimum set of points to represent the shape of a signature.
A contour can be described as the outer boundary of a signature. To extract the same,
the disconnected components in the signature are joined and the ‘holes’ inside the
signature are filled. The set of all four-connected boundary pixels define a contour of
the signature. The process undergoes the following steps. Figure 12 gives a pictorial
illustration of the sequence of steps in the process.
As depicted in figure 12a, the contour image obtained is first thickened using a 5X5
morphological filter followed by thinning [5]. This is done to eliminate any sharp
changes and to bring about uniformity in the curve. A unique point on the contour
image (must occur in the same region for the same user set) is then selected as a starting
point and the contour is traversed in the clockwise direction.
Discussions and Conclusions
Traditionally, graph matching is used on the whole image and each pixel is compared with
every other pixel in the other image, thus incurring a large computational overhead. For
images of resolution of P × Q and A x B, P, Q, A and B being typically in the range 300-1200,
we compute a non-negative n×m matrix, where the element in the i-th row and j-th column
represents the cost of assigning the i-th pixel to the j-th pixel in considered images 1 and 2
respectively. The complexity of the algorithm amounts to be O(n3) where (n~=m).