Scholarlink

<p>Using unmanned aerial vehicles (UAVs) for road crack inspection significantly enhances efficiency. It also broadens coverage, which is important for ensuring traffic safety. However, due to the high visual similarity between road textures and defect targets, coupled with the complex and irregular morphology of cracks, missed detections are likely to occur. Most current models adopt standard convolutional structures during feature encoding, which limits their ability to capture fine-grained texture details. Furthermore, high-level semantic features often lack adequate spatial resolution. To overcome these limitations, this paper proposes an efficient multi-scale dual-attention network (EMDANet). First, in the feature extraction stage, we design an enhanced global shuffle convolution module with a parallel-branch structure. It preserves fine-grained information and improves discrimination between defects and similar backgrounds. Second, high-level features possess strong semantic representation but lack sufficient spatial detail. To overcome this issue, an attention mechanism is employed to enhance spatial perceptual localization, thereby improving the recognition of irregular defect patterns. Finally, in the detection stage, four detection heads are constructed to enhance the detection of objects at different scales. In addition, the SlideLoss function is introduced to address the class imbalance problem in crack categories. Experimental results show that the AR and mAP50 reach 66.9% and 69.3%, respectively, on the&#xa0;RDD2022-China-Drone dataset. The detection speed achieves 41 frames per second on self-collected 4K road videos, demonstrating excellent real-time performance and strong potential for engineering applications.</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

EMDANet: real-time road crack detection in UAV remote sensing images via multi-scale dual-attention network

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Lanxue Dang,
  • Gaowei Zhang,
  • Huiyu Mu,
  • Yi Xie

摘要

Using unmanned aerial vehicles (UAVs) for road crack inspection significantly enhances efficiency. It also broadens coverage, which is important for ensuring traffic safety. However, due to the high visual similarity between road textures and defect targets, coupled with the complex and irregular morphology of cracks, missed detections are likely to occur. Most current models adopt standard convolutional structures during feature encoding, which limits their ability to capture fine-grained texture details. Furthermore, high-level semantic features often lack adequate spatial resolution. To overcome these limitations, this paper proposes an efficient multi-scale dual-attention network (EMDANet). First, in the feature extraction stage, we design an enhanced global shuffle convolution module with a parallel-branch structure. It preserves fine-grained information and improves discrimination between defects and similar backgrounds. Second, high-level features possess strong semantic representation but lack sufficient spatial detail. To overcome this issue, an attention mechanism is employed to enhance spatial perceptual localization, thereby improving the recognition of irregular defect patterns. Finally, in the detection stage, four detection heads are constructed to enhance the detection of objects at different scales. In addition, the SlideLoss function is introduced to address the class imbalance problem in crack categories. Experimental results show that the AR and mAP50 reach 66.9% and 69.3%, respectively, on the RDD2022-China-Drone dataset. The detection speed achieves 41 frames per second on self-collected 4K road videos, demonstrating excellent real-time performance and strong potential for engineering applications.