Scholarlink

<p>This paper proposes a hybrid ensemble framework for dynamic signature verification called DeepCrossSign, with the aim of increasing classification accuracy and cross-dataset robustness. The proposed method is based on four deep learning models (Convolutional Neural Networks combined with Bidirectional Long Short-Term Memory, Deep Pyramidal Residual Networks, Parallel Temporal Convolutional Neural Networks with Transformers, and Convolution-and-Attention Network) to capture the spatial, temporal, and sequential dynamics of online signature data. In addition, a soft-voting approach and a class-weighted loss function are used to cope with unbalanced data and improve decision making. The system was tested on two benchmark datasets. On the SIGNDUMP dataset, the framework provided an accuracy of 96.35% and Equal Error Rate of 0.20%. On the SignatureFeatures dataset, it showed an accuracy of 97.76% and an Equal Error Rate of 3.54%. In a cross-dataset evaluation setup, the system showed superior performance as compared to similar models as it generated 90.49% accuracy and an Equal Error Rate of 5.25% when trained with SIGNDUMP and tested on SignatureFeatures. These results demonstrate the capabilities of the framework in both intra-domain and cross-domain scenarios.</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

DeepCrossSign: a robust ensemble framework for cross-dataset dynamic signature verification

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Atiya Kazi,
  • Vinayak Bharadi,
  • Kaushal Prasad

摘要

This paper proposes a hybrid ensemble framework for dynamic signature verification called DeepCrossSign, with the aim of increasing classification accuracy and cross-dataset robustness. The proposed method is based on four deep learning models (Convolutional Neural Networks combined with Bidirectional Long Short-Term Memory, Deep Pyramidal Residual Networks, Parallel Temporal Convolutional Neural Networks with Transformers, and Convolution-and-Attention Network) to capture the spatial, temporal, and sequential dynamics of online signature data. In addition, a soft-voting approach and a class-weighted loss function are used to cope with unbalanced data and improve decision making. The system was tested on two benchmark datasets. On the SIGNDUMP dataset, the framework provided an accuracy of 96.35% and Equal Error Rate of 0.20%. On the SignatureFeatures dataset, it showed an accuracy of 97.76% and an Equal Error Rate of 3.54%. In a cross-dataset evaluation setup, the system showed superior performance as compared to similar models as it generated 90.49% accuracy and an Equal Error Rate of 5.25% when trained with SIGNDUMP and tested on SignatureFeatures. These results demonstrate the capabilities of the framework in both intra-domain and cross-domain scenarios.