Scholarlink

<p>Handwritten signature verification remains a critical security challenge in financial transactions and document authentication, particularly due to increasing sophistication in forgery techniques. While recent deep learning approaches have shown promise, they still struggle with three key limitations: insufficient discriminative power for subtle signature variations, sensitivity to noise, and inconsistent performance across diverse signature styles. We address these challenges by introducing a novel verification framework that synergistically combines ArcLoss for enhanced feature discrimination with AutoEncoder for robust signature representation learning. Our framework integrates these components with ResNet50 and MobileNet backbones through a carefully designed fusion strategy. Comprehensive experiments demonstrate state-of-the-art performance across multiple metrics and datasets. On our extensive private test set, the framework achieves 99.7% accuracy with ResNet50 and 99.3% with MobileNet, while maintaining exceptionally low error rates (FAR: 0.01/0.09, FRR: 0.02/0.03). These results significantly outperform recent methods including SynSig2Vec and VAE-based approaches. Ablation studies validate the effectiveness of our ArcLoss–AutoEncoder fusion, showing consistent improvements over individual components and alternative combinations. The framework demonstrates robust generalization capability, maintaining superior performance across CEDAR (98%), GPDS (97%), and Chinese Signature Database (89%) benchmarks. Cross-dataset validation confirms its adaptability to diverse signature styles and forgery types. Through extensive comparison with five state-of-the-art methods, our approach shows consistent improvement in both accuracy and robustness, particularly in handling skilled forgeries and signature variations. This research advances signature verification technology towards practical deployment in security-critical applications. Code and implementation details are available at: <a href="https://shorturl.at/xzUkx">https://shorturl.at/xzUkx</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel fusion of ArcLoss and AutoEncoder for high-precision handwritten signature verification

  • 期刊论文

  • 发表日期：

    2025年12月27日

查看全文
  • Do Thanh Ton,
  • Duc-Tuan-Anh Nguyen,
  • Van-Hau Nguyen

摘要

Handwritten signature verification remains a critical security challenge in financial transactions and document authentication, particularly due to increasing sophistication in forgery techniques. While recent deep learning approaches have shown promise, they still struggle with three key limitations: insufficient discriminative power for subtle signature variations, sensitivity to noise, and inconsistent performance across diverse signature styles. We address these challenges by introducing a novel verification framework that synergistically combines ArcLoss for enhanced feature discrimination with AutoEncoder for robust signature representation learning. Our framework integrates these components with ResNet50 and MobileNet backbones through a carefully designed fusion strategy. Comprehensive experiments demonstrate state-of-the-art performance across multiple metrics and datasets. On our extensive private test set, the framework achieves 99.7% accuracy with ResNet50 and 99.3% with MobileNet, while maintaining exceptionally low error rates (FAR: 0.01/0.09, FRR: 0.02/0.03). These results significantly outperform recent methods including SynSig2Vec and VAE-based approaches. Ablation studies validate the effectiveness of our ArcLoss–AutoEncoder fusion, showing consistent improvements over individual components and alternative combinations. The framework demonstrates robust generalization capability, maintaining superior performance across CEDAR (98%), GPDS (97%), and Chinese Signature Database (89%) benchmarks. Cross-dataset validation confirms its adaptability to diverse signature styles and forgery types. Through extensive comparison with five state-of-the-art methods, our approach shows consistent improvement in both accuracy and robustness, particularly in handling skilled forgeries and signature variations. This research advances signature verification technology towards practical deployment in security-critical applications. Code and implementation details are available at: https://shorturl.at/xzUkx.