| Over the past decade, the problem of handwritten signature verification has been solved through real-time measurement of signature dynamics, such as pressure and acceleration. These on-line systems offer excellent performance even when tested against practiced forgeries. However, special equipment required at the point of signature generation limits their applicability. Furthermore, non-practiced, or casual, forgeries are more abundant in the credit and banking world where centralized verification is possible. In this thesis, new methods to distinguish casual forgeries from valid signatures, based solely on images of the signatures, are presented.; The slope histogram approach is presented first. This method exploits the regularity of length and curvature of a signature. Overall signature content at various angles is evaluated to form the histogram. Histograms are then passed to a classifier constructed from a small number of valid signatures. Performance of the classifier on a data pool of 1000 signatures is evaluated. In particular, the equal error rate of this approach is shown to average 7% across 9 different subjects.; Next, the synthetic discriminant function (SDF) approach is presented. This approach selects a linear filter which produces a specified output for each image of a training set. Performance of this approach with a small number of valid signatures in the training set is examined, and substantial improvement is demonstrated when forgeries are included in the set. In particular, the equal error rate for the SDF with forgeries included is shown to average approximately 4%. The effects of image pre-processing on SDF performance are examined. Alternatives to the inclusion of forgeries in the training set are proposed. Results are presented that establish equality between the SDF and a single-layer neural network under certain conditions, as well as equality between the SDF and certain eigenvector decomposition techniques.; Finally, it is shown that valid signatures produce decisions of authenticity that are roughly independent for each of these systems. The methods can therefore be combined to attain improved performance, and an integrated system with an average equal error rate of approximately 1% is demonstrated. The performance of the combined system is superior to that of existing off-line methods. Additional tests are performed on a set of less-casual forgeries that show a 4-5% increase in equal error rate when efforts are made to duplicate the target signature. |
