Application of Siamese Neural Network for Offline Signature Verification Based on Similarity Level

Abstract

The increasing demand for secure and accurate identity verification systems has led to the development of various biometric technologies, one of which is signature verification. Despite the rise of digital authentication methods, signatures remain a widely accepted and legally binding form of identity verification, especially in paper-based systems. This research explores the application of the Siamese Neural Network (SNN) method for offline signature verification based on image similarity levels. The study aims to reduce human error, speed up verification time, and increase accuracy in identifying genuine and forged signatures. The dataset used in this study consists of 210 signature images collected from 14 respondents, including 7 with genuine signatures and 7 with forged signatures (categorized as random, unskilled, and skilled forgeries). Preprocessing steps such as scanning, resizing, and CSV data generation were conducted to optimize input for the SNN model. The model was trained using contrastive loss to learn signature similarity representations and was evaluated using a confusion matrix. The training dataset included 147 image pairs, and the testing set contained 63 image pairs, resulting in 168 prediction possibilities. The SNN achieved an accuracy rate of 94%, correctly predicting 159 cases while misclassifying 8 due to image quality and unclear signature strokes. These results indicate that the Siamese Neural Network is effective for offline signature verification and demonstrates strong potential for real-world implementation in identity authentication systems. This research contributes to the field of computer vision, particularly in biometrics, by providing an efficient, learning-based approach to signature validation using deep learning techniques.