Overview

As an increasing number of transactions, especially financial, are being authorized via signature, methods of automatic signature verification must be developed if authenticity is to be verified on a regular basis. Approaches to signature verification fall into two categories according to the acquisition of the data: on-line and off-line. On-line data records the motion of the stylus while the signature is produced, and includes location, and possibly velocity, acceleration and pen pressure, as functions of time. Off-line data is an image of the signature. Although on-line verification is of great interest for point of sale'' applications, our research focuses primarily on off-line verification which is of interest to check clearing houses, tax processing centers and other locations where a hard copy of the signature is obtained.

Progress on the problem of signature verification using on-line data has far surpassed progress on the processing of off-line signature images, partially because the features such as precise position, velocity, acceleration and pressure are not explicit in the off-line data. One difficulty lies in the fact that it is hard to segment signature strokes due to highly stylish and unconventional writing styles. Many features essential for detection of forgeries, however, are embedded in the stroke level and can be extracted robustly.

Our approach relies on a combination of methods by using on-line type models used to segment and process images of signatures obtained off-line. We assume that the model consists of an image of the signature, as well as the temporal ordering'' of stroke segments, including retraced and other strokes fully or partially occluded in the image. Although the model will typically be gathered during an enrollment process from a tablet device, it can also be traced manually or obtained from a system which extracts temporal ordering from document images.

Given a model, we are working on a system to detect random, simple and traced forgeries. We have developed algorithms to segment the strokes into stylistically meaningful segments.

Our algorithms pre-process the signature by extracting an edge map from the grey level image. This edge map preserves enough structure so that we can obtain information from the signature such as local stroke direction and width. From the local features, we attempt to establish a correspondence between the model and the signature. A cost function is used to measure the deviation between the two signatures. If we cannot establish such a correspondence, the signature is ruled out as a random forgery. Results of our preliminary implementation show that we can reject approximately 94% of random forgeries, while accepting over 98% of genuine signatures.

Based on the existence of a correspondence, the system is being extended to use local shape features to detect simple forgeries and stroke intensity profiles to detect traced forgeries.