How effective are radial basis function neural networks for offline handwritten signature verification?

Abstract

The objective of this project was to investigate the effectiveness of totally radial basis function neural network (RBFNN) single-layer architecture for offline handwritten signature verification. An RBFNN, initialised by supervised clustering, was adopted for each author’s signature samples. RBFNNs are quite new in this domain, and are well-known for the robustness in eliminating outliers and for the relatively simple computations required to be trained. These were the main motivator factors that challenged the author of this project to investigate the effectiveness of RBFNNs in the field of offline handwritten signature verification. A signature database was collected for the scope of this study as no international public database is available. Professional recommendations by J. Gaffiero who is a Maltese graphologist and personal recommendations by H. Baltzakis helped to acquire a signature database with as much intrapersonal variations as possible. Three groups of signature features namely global, grid and texture features were used to evaluate the system in different scenarios. The grid and texture features were extracted from a superimposed grid of 12×8 segments, where a vector quantisation (VQ) technique was required to cluster the respective column feature vectors. In this case, two VQ approaches were investigated; an adaptively sized codebook VQ and a fixed size codebook VQ of 50 codewords. The entire system was extensively tested with random signature forgeries and the high recognition rates obtained show that the proposed architecture is effective in this field. Surprisingly, the fixed size codebook VQ performed at least twice as good as the adaptively sized codebook VQ. In fact, the best results where obtained when global and grid features where combined producing a feature vector of 592 elements. In this case a Mean Error Rate (MER) of 2.04% with a False Rejection Rate (FRR) of 1.58% and a False Acceptance Rate (FAR) of 2.5% were achieved. The mentioned results were found to rank better than some other published studies.