Implementation and Performance Analysis of CNN Architectures for Offline Signature Recognition: A Reproduction Study

Project Description: This research presents a reproduction study of deep learning-based architectures for offline handwritten signature recognition. Developed as part of the Machine Learning curriculum at BRAC University, the project addresses the challenges of biometric authentication by implementing a Convolutional Neural Network (CNN) to distinguish between genuine and forged signatures from static images. Methodology: The study utilizes a custom CNN architecture featuring multiple convolutional layers, max-pooling for spatial reduction, and dense fully connected layers for classification. The model was trained using the Adam optimizer and binary cross-entropy loss, focusing on capturing intricate spatial patterns and structural nuances in handwritten strokes that are often missed by traditional handcrafted feature extraction methods. Key Results: The reproduced model achieved a high test accuracy of 95.15%, demonstrating the robustness of CNNs in handling intra-class variations and signature forgeries. The findings validate the effectiveness of the original research’s architecture and provide a benchmark for future iterations in biometric security systems.