Improved YOLOv5 for Road Disease Detection

Guangfu Wu,Longxin Liang,Hao Liu,Yun Li

DOI: https://doi.org/10.1109/docs63458.2024.10704249

2024-01-01

Abstract:Rapid detection and high accuracy are crucial for effective road disease detection in road maintenance. This study proposes an improved YOLOv5 algorithm to address the issues of poor efficiency and low accuracy in traditional road disease detection methods. First, an enhanced attention module, Mixed Attention Squeeze-and-Excitation (MASE), is integrated into the backbone network. This module improves feature extraction in complex backgrounds by enhancing foreground and background information discrimination. It significantly refines detail processing in scenarios where diverse and intricate background elements obscure or confuse disease features. Second, the original PANet feature fusion framework is enhanced with a cross-layer enhancement network (CLEN), improving the fusion of small-scale features. This enhancement makes it more efficient at processing feature information across different scales, thereby addressing the issue of tiny disease features disappearing during multiple downsampling stages. A new target detection bounding box loss function, Hybrid IoU Loss (HIoU), is also designed to provide a more comprehensive loss calculation. This function effectively addresses the challenge of detecting irregularly shaped diseases. Experimental results demonstrate that the improved algorithm significantly outperforms the original algorithm, with mAP values increased by 10.4 % on the CWNU(China West Normal University) dataset and 2.3 % on the RDD2022 dataset.