Scholarlink

<p>Visual recognition is a promising approach for detecting sudden rockfalls. However, environmental factors such as rain, dust, and dense vegetation can significantly degrade recognition accuracy, increasing the likelihood of false positives and missed detections. In this study, we propose a lightweight rockfall detection approach, REGM-YOLO, which leverages receptive field and attention mechanisms to enhance performance in complex environments. Our approach begins with constructing a rockfall dataset augmented by adding noise and plant shapes with green opacity to simulate challenging conditions. The method incorporates a receptive field module to extend the perceptual range of rockfall features, enhancing detection accuracy even in the presence of occlusions. Additionally, a channel attention mechanism refines feature expression by emphasizing rockfall characteristics and minimizing irrelevant background interference. A lightweight backbone network accelerates detection speed, while an advanced activation function improves the identification of rockfalls with intricate features. Experimental results indicate that REGM-YOLO improves the <i>P</i>-value by 7%, <i>R</i>-value by 20.5%, MAP@0.5 by 16.7%, and MAP@0.5:0.9 by 11.7%, while reducing the number of parameters by 38.2%, compared to the original model, thereby validating its effectiveness. This study is expected to support the development of more efficient rockfall monitoring systems.</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

A lightweight rockfall detection method in complex environments based on receptive field and attention mechanism: REGM-YOLO

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Hui Chen,
  • Lu Zhang,
  • Shuaixing Yan,
  • Xiaopeng Li,
  • Dongpo Wang

摘要

Visual recognition is a promising approach for detecting sudden rockfalls. However, environmental factors such as rain, dust, and dense vegetation can significantly degrade recognition accuracy, increasing the likelihood of false positives and missed detections. In this study, we propose a lightweight rockfall detection approach, REGM-YOLO, which leverages receptive field and attention mechanisms to enhance performance in complex environments. Our approach begins with constructing a rockfall dataset augmented by adding noise and plant shapes with green opacity to simulate challenging conditions. The method incorporates a receptive field module to extend the perceptual range of rockfall features, enhancing detection accuracy even in the presence of occlusions. Additionally, a channel attention mechanism refines feature expression by emphasizing rockfall characteristics and minimizing irrelevant background interference. A lightweight backbone network accelerates detection speed, while an advanced activation function improves the identification of rockfalls with intricate features. Experimental results indicate that REGM-YOLO improves the P-value by 7%, R-value by 20.5%, MAP@0.5 by 16.7%, and MAP@0.5:0.9 by 11.7%, while reducing the number of parameters by 38.2%, compared to the original model, thereby validating its effectiveness. This study is expected to support the development of more efficient rockfall monitoring systems.