Scholarlink

<p>To address the challenges of missed detection, false detection, and low efficiency in road damage detection, this paper proposes an enhanced road damage detection model YOLO-R based on improved YOLOv8. First, we construct a collaborative module combining Switchable Atrous Convolution (SAC) with C2f. This module achieves precise multi-scale global feature extraction through dynamic adjustment of receptive fields. Second, an Adaptive Downsampling module (ADown) is designed to reduce computational complexity while maintaining feature representation capability. Furthermore, a Slim-neck module based on lightweight Group Shuffle Convolution (GSConv) and GSbottleneck is developed using VoV-GSCSP architecture to significantly improve feature fusion efficiency. Finally, the Shape-IoU bounding box regression loss function is introduced to improve localization accuracy through geometric prior optimization. Experiments on the cross-regional CRDDC2022 dataset (covering China and the US) demonstrate that YOLO-R achieves 76.3% mAP@0.50 while maintaining high inference speed, outperforming the baseline model by 3.7%. The model parameters and computational costs were reduced by approximately 12% and 33%, respectively. Comparative experiments confirm the model’s effectiveness and superiority in real-time road damage detection in complex scenarios.</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

YOLO-R:road damage detector based on dynamic receptive fields and lightweight design

  • 期刊论文

  • 发表日期：

    2025年7月18日

查看全文
  • Jinghong Mao,
  • Mei Wang,
  • Zhenghong Liu

摘要

To address the challenges of missed detection, false detection, and low efficiency in road damage detection, this paper proposes an enhanced road damage detection model YOLO-R based on improved YOLOv8. First, we construct a collaborative module combining Switchable Atrous Convolution (SAC) with C2f. This module achieves precise multi-scale global feature extraction through dynamic adjustment of receptive fields. Second, an Adaptive Downsampling module (ADown) is designed to reduce computational complexity while maintaining feature representation capability. Furthermore, a Slim-neck module based on lightweight Group Shuffle Convolution (GSConv) and GSbottleneck is developed using VoV-GSCSP architecture to significantly improve feature fusion efficiency. Finally, the Shape-IoU bounding box regression loss function is introduced to improve localization accuracy through geometric prior optimization. Experiments on the cross-regional CRDDC2022 dataset (covering China and the US) demonstrate that YOLO-R achieves 76.3% mAP@0.50 while maintaining high inference speed, outperforming the baseline model by 3.7%. The model parameters and computational costs were reduced by approximately 12% and 33%, respectively. Comparative experiments confirm the model’s effectiveness and superiority in real-time road damage detection in complex scenarios.