Signature Verification System: Real-Time Forgery Detection Using Deep Learning and Computer Vision

Abstract

Recent years have witnessed a significant rise in the application of deep convolutional neural networks (CNNs) in computer vision and pattern recognition tasks. Offline handwritten signature verification, which remains a challenging problem, is one of the most widely used biometric techniques in banking systems, administrative processes, and financial applications. This project explores CNN-based approaches for signature verification and evaluates the effectiveness of deep learning models as feature extractors using a transfer learning strategy. The proposed system presents a robust deep learning-based framework for detecting forged signatures by leveraging the ResNet50 architecture, a pre-trained convolutional neural network widely used for image classification tasks. ResNet50 incorporates residual connections, which enable efficient information flow across layers and prevent gradient vanishing, thereby improving the performance of deeper networks. In this approach, signature images are first preprocessed using OpenCV techniques such as resizing and normalization, and then passed through a fine-tuned ResNet50 model for feature extraction and classification into genuine or forged categories. To enhance usability, a web-based interface is developed using Django, allowing users to upload signature images and obtain real-time verification results. The system is designed to operate on standard computational resources, making it cost-effective, scalable, and suitable for real-world deployment in banking, legal documentation, and identity verification systems. Experimental evaluation on a dataset containing both genuine and forged signatures demonstrates that the proposed model achieves an average classification accuracy of approximately 95%, indicating strong robustness against variations in writing styles and image conditions. Furthermore, the system addresses key limitations of traditional methods, such as reliance on handcrafted features, poor generalization, and lack of real-time verification capability.