Scholarlink

<p>Rockfalls in mountainous transportation corridors are highly unpredictable and pose significant risks, which require accurate and timely detection to ensure safety. However, such detection is often hindered by environmental factors like illumination changes, rain, and fog, which degrade the visibility of rockfall outlines. In addition, many high-accuracy detection models are computationally intensive and impractical for real-time monitoring. To address these challenges, we constructed and open-sourced a large-scale Mountain Rockfall Dataset (MRDataset), which comprises 3,921 annotated images. We also propose a lightweight Edge-Enhanced Network (EENet), which incorporates an Edge Spatial Stem (ESStem) to capture preliminary edge features and a Global Edge Fusion Network (GEFNet) to combine multi-scale information. The model uses dilated convolutions to preserve contextual continuity in boundaries and employs pruning and knowledge distillation to reduce size while maintaining detection accuracy. Experimental results demonstrate that EENet surpasses YOLO11, improving Precision by 5.2%, Recall by 1.5%, mAP50 by 3.3%, and mAP50<sub>:95</sub> by 2.9%, while reducing the number of Parameters and FLOPs by 73.3% and 52.4%, respectively. Further tests show that EENet preserves edge integrity under complex conditions, substantially decreasing both missed and false detections. This study offers an accurate and efficient approach for rockfall monitoring, with promising potential for real-world 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

EENet: An edge-enhanced network for robust rockfall detection in complex mountainous environments

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Hongbing Liu,
  • Mingjin Zhang,
  • Shenghan Zhuang,
  • Kunping Chen,
  • Jinxiang Zhang

摘要

Rockfalls in mountainous transportation corridors are highly unpredictable and pose significant risks, which require accurate and timely detection to ensure safety. However, such detection is often hindered by environmental factors like illumination changes, rain, and fog, which degrade the visibility of rockfall outlines. In addition, many high-accuracy detection models are computationally intensive and impractical for real-time monitoring. To address these challenges, we constructed and open-sourced a large-scale Mountain Rockfall Dataset (MRDataset), which comprises 3,921 annotated images. We also propose a lightweight Edge-Enhanced Network (EENet), which incorporates an Edge Spatial Stem (ESStem) to capture preliminary edge features and a Global Edge Fusion Network (GEFNet) to combine multi-scale information. The model uses dilated convolutions to preserve contextual continuity in boundaries and employs pruning and knowledge distillation to reduce size while maintaining detection accuracy. Experimental results demonstrate that EENet surpasses YOLO11, improving Precision by 5.2%, Recall by 1.5%, mAP50 by 3.3%, and mAP50:95 by 2.9%, while reducing the number of Parameters and FLOPs by 73.3% and 52.4%, respectively. Further tests show that EENet preserves edge integrity under complex conditions, substantially decreasing both missed and false detections. This study offers an accurate and efficient approach for rockfall monitoring, with promising potential for real-world applications.