Scholarlink

<p>Edge detection is a fundamental task in computer vision, essential for various applications such as image segmentation, object detection, and 3D curve reconstruction. Traditional edge detection methods often struggle with semantic relationships, resulting in edges with meaningless parts. Deep learning-based edge detectors, while more effective, rely heavily on pixel-level labels, which are prone to errors and subjective variations. To address these issues, we propose a novel dual-branch attention fusion edge detection model with Canny-assisted supervision (DAF-C). The model integrates the dual attention mechanisms of SimAM and CBAM in parallel, effectively enhancing its ability to recognize complex edge structures. By deeply fusing edge features, it reduces information loss. The designed interactive attention fusion module allows for dynamic recalibration of features, thereby generating edge maps with improved continuity and reduced noise. Furthermore, we introduce Canny-based compensatory supervision during training, which increases label diversity and enables the model to receive more comprehensive and effective supervision. This approach ensures the alignment of semantic comprehension with geometric precision by incorporating Canny edge priors as auxiliary supervision. Extensive experiments on the BSDS500 and Multicue datasets demonstrate that our DAF-C model outperforms existing multi-scale edge detection models, achieving competitive performance and robust edge detection results. The code for this project is available at <a href="https://github.com/sfg-oss/daf-c">https://github.com/sfg-oss/daf-c</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

Enhanced edge detection via Dual-branch attention fusion with Canny-assisted supervision

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月10日

查看全文
  • Hengtao Wang,
  • Peishun Liu,
  • Quanjie Dou,
  • Yibao Song,
  • Mengqi Luo,
  • Rongjia Han,
  • Boning Zhang

摘要

Edge detection is a fundamental task in computer vision, essential for various applications such as image segmentation, object detection, and 3D curve reconstruction. Traditional edge detection methods often struggle with semantic relationships, resulting in edges with meaningless parts. Deep learning-based edge detectors, while more effective, rely heavily on pixel-level labels, which are prone to errors and subjective variations. To address these issues, we propose a novel dual-branch attention fusion edge detection model with Canny-assisted supervision (DAF-C). The model integrates the dual attention mechanisms of SimAM and CBAM in parallel, effectively enhancing its ability to recognize complex edge structures. By deeply fusing edge features, it reduces information loss. The designed interactive attention fusion module allows for dynamic recalibration of features, thereby generating edge maps with improved continuity and reduced noise. Furthermore, we introduce Canny-based compensatory supervision during training, which increases label diversity and enables the model to receive more comprehensive and effective supervision. This approach ensures the alignment of semantic comprehension with geometric precision by incorporating Canny edge priors as auxiliary supervision. Extensive experiments on the BSDS500 and Multicue datasets demonstrate that our DAF-C model outperforms existing multi-scale edge detection models, achieving competitive performance and robust edge detection results. The code for this project is available at https://github.com/sfg-oss/daf-c.