A Deep Feature Reliability-Based Framework for Offline Handwritten Signature Verification
Offline handwritten signature verification is widely employed as a biometric authentication technique in financial, legal, and administrative applications. However, accurately distinguishing genuine signatures from skilled forgeries remains a challenging task because of natural intra-writer variations and the inconsistent discriminative capability of deep feature representations. This paper presents a Deep Feature Reliability-Based Framework for Offline Handwritten Signature Verification that enhances verification performance by identifying and utilizing stable writer-specific deep features. Initially, signature images are preprocessed and represented using 2048-dimensional deep features extracted from a pre-trained ResNet50 network. A feature reliability estimation scheme is then introduced to evaluate the statistical consistency of individual feature dimensions across genuine signatures of each writer. Based on the estimated reliability scores, the most reliable features are selected and incorporated into a weighted cosine similarity measure for signature verification. The proposed framework is evaluated on the CEDAR offline handwritten signature dataset using three training–testing splits of 25–75, 50–50, and 75–25. Experimental results demonstrate that the proposed method achieves its best performance with the 75–25 split, obtaining an accuracy of 77.61%, precision of 83.77%, recall of 68.48%, and an F1-score of 75.33%. The findings indicate that incorporating feature reliability into the verification process improves the robustness of deep feature representations and provides an effective framework for offline handwritten signature verification.