Advancing Offline Signature Verification with Bidirectional Siamese Deep Learning: A Writer-Independent Approach

Abstract

Offline handwritten signature verification is critical for secure authentication in legal, financial, and administrative domains. Traditional systems struggle with intra-user variability and skilled forgeries, while writer-dependent approaches lack scalability. Unlike digital data validation, verifying physical signatures is a complex task due to variations in writing styles, intra-writer inconsistency, and skilled forgery attempts. This research introduces a writer-independent deep learning approach using a Bidirectional Recurrent Convolutional Siamese Network (BRCSN) to address these challenges. The proposed model combines convolutional neural networks (CNNs) to extract spatial features and bidirectional gated recurrent units (Bi-GRUs) to learn sequential stroke patterns, allowing the system to capture both local and global dependencies from static signature images. A triplet loss function is employed to optimize the network by minimizing the distance between genuine signature pairs and maximizing the distance between forged ones in the embedding space. The system is evaluated on benchmark multi-lingual signature datasets and achieves high classification accuracy, recall, and F1-score, while maintaining low false acceptance and rejection rates. Importantly, the BRCSN model performs consistently across diverse users without requiring writer-specific training, making it well-suited for real-time deployment in legal, financial, and administrative applications. Inference time is kept under 100 milliseconds, supporting practical use cases. By eliminating the need for retraining and leveraging spatial- temporal learning through BRCSN, this research contributes to the advancement of secure and scalable biometric systems capable of robust forgery detection.