Development of an Attention-Based Deep Metric Learning Method for Offline Signature Verification in a Limited-Data Environment

Abstract

This study aims to develop an attention-based Deep Metric Learning method for offline signature verification in environments with limited data availability. The main challenge in offline signature verification lies in the high intra-class variation and the similarity between genuine and forged signatures, particularly when the amount of training data is limited. The proposed method employs a Siamese Convolutional Neural Network architecture combined with an attention mechanism to enhance discriminative feature extraction capabilities. The dataset used in this study was obtained from offline sources and Kaggle, consisting of genuine and forged signature images. The research process includes preprocessing, signature pair generation, feature extraction, embedding generation using Deep Metric Learning, and optimization using Contrastive Loss. Experimental results demonstrate that the proposed method achieved an Accuracy of 91.12%, Precision of 92.27%, Recall of 92.43%, F1-score of 90.75%, and an Equal Error Rate (EER) of 4.88%. These results indicate that the integration of the attention mechanism and Deep Metric Learning effectively improves the system's capability to recognize signature patterns under limited data conditions.