Scholarlink

<p>Handwritten signature verification remains a challenging task due to the need for a large number of signature samples. In real-world applications, collecting extensive signatures from each individual for model training is impractical. To address this, we introduce a novel approach to few-shot signature verification using a “Double Siamese Network” model that integrates Convolutional Neural Networks (CNNs) and Data-efficient Image Transformers (DeiT). This architecture enables writer-independent verification without requiring individual-specific training while maintaining high performance. Our approach was evaluated on the CEDAR and ICDAR datasets, achieving high accuracies of 0.983 and 0.900, respectively, demonstrating its effectiveness in verifying signatures from familiar sources. A key finding is the impact of sample size on model performance: accuracy improved from 0.843 to 0.900 as the number of signatures per person increased from 5 to 9, emphasizing the crucial role of sample size in few-shot learning. However, cross-domain performance remains a challenge—when tested on the GPDS 1-150 dataset, accuracy significantly declined, indicating difficulties in transferring learned features across datasets.</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

Few-shot signature verification with Double Siamese Network

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Chalita Iamleelaporn,
  • Ranakorn Boonsuankergchai,
  • Tanwalai Yoongkieo,
  • Nattawut Intanai,
  • Patcharaporn Tuntino,
  • Thitirat Siriborvornratanakul

摘要

Handwritten signature verification remains a challenging task due to the need for a large number of signature samples. In real-world applications, collecting extensive signatures from each individual for model training is impractical. To address this, we introduce a novel approach to few-shot signature verification using a “Double Siamese Network” model that integrates Convolutional Neural Networks (CNNs) and Data-efficient Image Transformers (DeiT). This architecture enables writer-independent verification without requiring individual-specific training while maintaining high performance. Our approach was evaluated on the CEDAR and ICDAR datasets, achieving high accuracies of 0.983 and 0.900, respectively, demonstrating its effectiveness in verifying signatures from familiar sources. A key finding is the impact of sample size on model performance: accuracy improved from 0.843 to 0.900 as the number of signatures per person increased from 5 to 9, emphasizing the crucial role of sample size in few-shot learning. However, cross-domain performance remains a challenge—when tested on the GPDS 1-150 dataset, accuracy significantly declined, indicating difficulties in transferring learned features across datasets.