Scholarlink

To address pseudo-changes caused by background variations in remote sensing change detection, a dual-branch difference-enhanced network (DDECNet) with an efficient cross-attention mechanism is proposed. The Dual-Temporal Interaction Augmentation Module (DTIAM) enhances semantic consistency during bi-temporal feature interaction through temporal state alignment. The Difference Feature Enhancement and Supplementary Module (DFESM) reduces information loss in differential feature extraction using dual-branch feedback mechanisms for bidirectional temporal state correction. Differential features and bi-temporal interaction features are integrated via the Efficient Cross-Attention (ECA) mechanism, which establishes adaptive attention weighting through multi-scale aggregation to enhance key region identification. Evaluations on WHU-CD, SYSU-CD, and LEVIR-CD demonstrate superior performance, achieving F1-score improvements of 0.48% and 0.78% over existing methods on WHU-CD and SYSU-CD, respectively, while maintaining competitive accuracy on LEVIR-CD. The method achieves state-of-the-art performance, with code available at: https://github.com/decoder0112/DDECNet .

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

DDECNet: Dual-Branch Difference Enhanced Network with Novel Efficient Cross-Attention for Remote Sensing Change Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Wei Wang,
  • Qing Su,
  • Xin Wang

摘要

To address pseudo-changes caused by background variations in remote sensing change detection, a dual-branch difference-enhanced network (DDECNet) with an efficient cross-attention mechanism is proposed. The Dual-Temporal Interaction Augmentation Module (DTIAM) enhances semantic consistency during bi-temporal feature interaction through temporal state alignment. The Difference Feature Enhancement and Supplementary Module (DFESM) reduces information loss in differential feature extraction using dual-branch feedback mechanisms for bidirectional temporal state correction. Differential features and bi-temporal interaction features are integrated via the Efficient Cross-Attention (ECA) mechanism, which establishes adaptive attention weighting through multi-scale aggregation to enhance key region identification. Evaluations on WHU-CD, SYSU-CD, and LEVIR-CD demonstrate superior performance, achieving F1-score improvements of 0.48% and 0.78% over existing methods on WHU-CD and SYSU-CD, respectively, while maintaining competitive accuracy on LEVIR-CD. The method achieves state-of-the-art performance, with code available at: https://github.com/decoder0112/DDECNet .