Offline Signature Verification System Using Siamese Networks with Four Distinct Backbone Architectures: MobileNetV2, VGG16, InceptionV3 and a Custom CNN

This document presents an in-depth examination of offline signature verification utilizing Siamese neural networks. Our study tackles the issue of advanced forgeries by assessing a Siamese network design with four different backbone architectures: MobileNetV2, VGG16, InceptionV3, and a custom CNN. For model evaluation, we employed the CEDAR dataset, a reputable benchmark known for offering a realistic testing framework. Our custom CNN surpassed all other models with an outstanding accuracy of 99.81%. This marks a notable advancement compared to conventional techniques and exceeded the performance of the pre-trained models, which also yielded strong results. Our system effectively accommodates natural variations in signatures while providing robust protection against forgeries. The results highlight the benefits of a specially designed CNN for this purpose,…