Handwritten Signature Verification (Offline) using Neural Network Approaches: A Comparative Study

Forgery detection has been a challenging area in the field of biometry, e.g., handwritten signatures. Signature verification is a bi-objective optimization problem. The two crucial parameters are accuracy and time of computation. In this work, a comprehensive study on application of Adaptive Resonance Theory (ART) Nets (Type 1 and 2) and Associative Memory Net (AMN) has been conducted. To decrease the time complexity a corresponding parallel version using OpenMP is developed for each algorithm. The algorithms are trained with the original/genuine signature and tested with a sample of twelve very similar-looking forged signatures. The study concludes that ART-1 detects fake signatures with an accuracy of 99.89%; whereas, ART-2 and AMN detect forgery with accuracies of 99.99% and 75.68% respectively which are comparable to other methods cited in this paper.