DEVELOPMENT OF AN INTELLIGENT OFFLINE SIGNATURE VERIFICATION SYSTEM USING DEEP RESIDUAL NEURAL NETWORKS (ResNet)

Abstract

Offline signature verification remains a critical biometric authentication challenge due to the inherent variability in genuine signatures and the increasing sophistication of forgery techniques. This study develops an intelligent offline signature verification system leveraging Deep Residual Neural Networks (ResNet) to address these challenges. The proposed system employs a ResNet-50 architecture with transfer learning to automatically extract discriminative features from static signature images, eliminating the need for handcrafted feature engineering. The model was trained and evaluated on benchmark signature datasets, including CEDAR and GPDS, with comprehensive image preprocessing and augmentation techniques applied to enhance generalization. The system achieved superior performance with a verification accuracy of 98.2%, significantly outperforming conventional CNN architectures (91.7%). The False Acceptance Rate (FAR) and False Rejection Rate (FRR) were recorded at 0.015 and 0.012 respectively, demonstrating the model's effectiveness in distinguishing genuine signatures from skilled forgeries. The residual learning framework effectively mitigates the vanishing gradient problem, enabling deeper network training and more robust feature representation. This research contributes a reliable, scalable solution for automated signature verification applicable in banking, legal, and administrative domains where document authentication is paramount.