Tracing the Truth: Offline Signature Forgery Detection with Deep Learning Methods

Abstract

Offline signature verification has become increasingly critical in today's digital age, where document authentication and fraud prevention are paramount concerns. This thesis presents a comprehensive deep learning approach to offline signature forgery detection, implementing multiple state-of-the-art architectures including MobileNet, ResNet50, InceptionV3, VGG19, and an ensemble model. The research addresses the challenging problem of distinguishing between genuine and forged signatures using the BHSig260 dataset, which contains Bengali and Hindi signatures. The proposed system incorporates advanced preprocessing techniques, data augmentation, and transfer learning to achieve high accuracy in signature verification. Key contributions include the development of an ensemble model that combines the strengths of multiple deep learning architectures, comprehensive evaluation metrics, and a robust preprocessing pipeline that handles real-world signature variations. The system demonstrates significant improvements in accuracy, precision, recall, and F1-score compared to individual models, making it suitable for practical applications in banking, legal documentation, and security systems. This research contributes to advancing automated signature verification technology and provides a scalable solution for combating signature forgery in various domains.