Scholarlink

<p>Recently, U-shaped networks have been widely explored for pavement crack segmentation and have achieved promising performance. However, despite these advances, pavement crack detection remains a critical and challenging task for road maintenance and traffic safety due to the complex and fine-grained nature of cracks. Specifically, current frequency domain methods do not effectively capture the relationships between high- and low-frequency features, resulting in suboptimal segmentation performance. In addition, most of the crack segmentation methods overlook the attenuation of information during the encoder-decoder process, which is crucial for preserving crack integrity. To address these issues, a Frequency Multi-angle Compensation and Low-resolution Guided Network (FCLG-Net) is constructed. In particular, a Frequency domain Multi-angle Compensation (FMAC) module is proposed that effectively leverages low-frequency information as global compensation to learn spatial weights and residual information of multi-angle high-frequency features, thereby enhancing the capability of crack segmentation. Meanwhile, a Low-resolution Guided Adaptive Fusion (LGAF) module is designed to refine the original low-resolution features and incorporate them into the adaptive fusion process with encoder-decoder layers, ultimately improving the completeness of crack segmentation. Experimental results on the CrackTree260, CrackLS315, and Crack760 datasets demonstrate that FCLG-Net outperforms several state-of-the-art segmentation methods. Specifically, the proposed method achieves F1 Scores of 85.62%, 68.89%, and 84.42%, and MIoU of 75.06%, 51.95%, and 73.38% on the three datasets, respectively. The source code is publicly available at <a href="https://github.com/zyy32/FCLG-Net.git.">https://github.com/zyy32/FCLG-Net.git.</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

FCLG-Net: frequency multi-angle compensation and low-resolution guided network for pavement crack segmentation

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Wen Yang,
  • Yingying Zheng,
  • Hang Sun,
  • Chao Liang,
  • Lei Fang

摘要

Recently, U-shaped networks have been widely explored for pavement crack segmentation and have achieved promising performance. However, despite these advances, pavement crack detection remains a critical and challenging task for road maintenance and traffic safety due to the complex and fine-grained nature of cracks. Specifically, current frequency domain methods do not effectively capture the relationships between high- and low-frequency features, resulting in suboptimal segmentation performance. In addition, most of the crack segmentation methods overlook the attenuation of information during the encoder-decoder process, which is crucial for preserving crack integrity. To address these issues, a Frequency Multi-angle Compensation and Low-resolution Guided Network (FCLG-Net) is constructed. In particular, a Frequency domain Multi-angle Compensation (FMAC) module is proposed that effectively leverages low-frequency information as global compensation to learn spatial weights and residual information of multi-angle high-frequency features, thereby enhancing the capability of crack segmentation. Meanwhile, a Low-resolution Guided Adaptive Fusion (LGAF) module is designed to refine the original low-resolution features and incorporate them into the adaptive fusion process with encoder-decoder layers, ultimately improving the completeness of crack segmentation. Experimental results on the CrackTree260, CrackLS315, and Crack760 datasets demonstrate that FCLG-Net outperforms several state-of-the-art segmentation methods. Specifically, the proposed method achieves F1 Scores of 85.62%, 68.89%, and 84.42%, and MIoU of 75.06%, 51.95%, and 73.38% on the three datasets, respectively. The source code is publicly available at https://github.com/zyy32/FCLG-Net.git.