A Prototype System For O -Line Signature Veri cation Using Multilayered Feedforward Neural Networks

This research investigated the suitability of using multi-layered feedforward neural networks for the task of oo-line signature veriication. The input to the network is a binary bitmap of size 160x35 pixels. Three classes of forged signatures were used in our experiments: skilled, traced and casual. The performance was evaluated based on a database of 480 signatures collected for 4 subjects. The networks with no hidden layer had the best overall clas-siication performance, indicating that the problem is linearly separable. Further, it was found that the networks are sensitive to diierent signers. The performance of the feedforward neural networks was the best for the casual forgeries where most networks were able to achieve 0% false acceptance rates. The feedforward neural networks ability to deal with skilled forgeries was limited and its ability to deal with the traced signatures was the worst.