Akademik Veri Yönetim Sistemi
Pattern Recognition Letters, cilt.198, ss.93-100, 2025 (SCI-Expanded)
The ability to detect forged signatures is crucial for maintaining the integrity of legal and financial documents. While traditional forensic methods – guided by expert judgment – have proven effective in many cases, they may face challenges in consistency and scalability, especially in complex or high-volume scenarios. This study explores deep learning architectures – including CNNs and Vision Transformers – enhanced with attention mechanisms and explainable AI (XAI) to improve accuracy, objectivity, and interpretability in forensic signature verification. Using the CEDAR dataset, Xception and ResNet50 achieved accuracies of 93.19% and 91.04%, respectively (p<0.001). When increasing reference samples from two to five, ResNet50 reached 94.10% accuracy. To evaluate interpretability, we compared multiple XAI methods, including Integrated Gradients, DeepLIFT, and SHAP. A validation study with a forensic handwriting examiner confirmed alignment between model attributions and expert-marked regions. Integrated Gradients yielded the highest agreement (F1@5 = 0.807) and the most faithful explanations (Insertion AUC = 0.86). These findings suggest that XAI-enhanced models can offer repeatable and interpretable decision support for forensic document analysis, potentially increasing their acceptability in legal and regulatory settings without compromising expert oversight.