SigVer—A Deep Learning Based Writer Independent Bangla Signature Verification System

Abstract

Handwritten signature is one of the most popular biometric traits which have been addressed by researchers since the last few decades. Scientists have proposed disparate handcrafted feature-based techniques for signature verification in the past. Presently there has been an interest towards deep learning-based approaches which has demonstrated promising results across different avenues of pattern recognition problems. In this paper, light-weight CNN architecture named SigVer is proposed for Bangla offline writer independent signature verification. The proposed network was built with only six layers with no pooling in between convolution layers to minimize information loss. Experiments were performed on 2956 signatures from 57 writers and a highest accuracy of 99.49% was obtained. Further, comparative analysis was performed with established CNN architectures as well as handcrafted feature-based techniques and SigVer produced better results.