A neural network-based on-line signature verification system

This dissertation investigates the development of a neural network based system for automated signature authentication that relies on an autoregressive characterization for the segments of a signature. The primary contributions of this work are two-fold: (a) the development of the neural network architecture and the modalities of training it, (b) adaptation of the dynamic time warping algorithm to formulate a new method for enabling consistent segmentation of multiple signatures from the same writer.; The validity of the use of autoregressive modeling of signatures has been demonstrated by earlier researchers. However, in their work, they used classical pattern recognition techniques. Here the next step is taken through the application of neural network classifiers to the problem. A multilayer perceptron trained through the backpropagation algorithm is formulated and its performance is investigated using an extensive signature database. While a signature is nonstationary signal in its totality, it is modeled as a concatenation of stationary segments in order to capture its evolution. It is extremely important to be able to segment multiple samples of a writer's signature consistently. The dynamic time warping algorithm is used in this work as the basis for developing a new method for segmenting multiple signatures from a writer consistently.; The performance of the signature verification system developed has been tested using a sizable database that includes a comprehensive set of simulated and real forgeries. The accuracies obtained are very promising in terms of the potential for practical deployment of system.