Evaluation of a training method and of various rejection criteria for a neural network classifier used for off-line signature verification

This paper addresses the problems related to the design of a neural network classifier used in the first stage of an automatic handwritten signature verification system. We used the directional probability density function as a global shape vector, and its discriminating power was enhanced by a pretreatment. The training phase of the backpropagation network (BPN) was conducted by using the global classification error in memorization and in generalization. To improve the global performance of the BPN classifier, various rejection criteria were evaluated and the number of hidden neurons optimized by means of experimental protocols. The BPN classifier is better than the threshold classifier, and compares favourably with the k nearest neighbour classifier