Identification and Detection of Road Damages Based on Improved YOLOv5
TL;DR
This work proposed an improved YOLOv5- based model for accurate road damage identification by introducing three key improvements: a P2 detection head to improve the detection of small and distant road damages, the CBAM attention mechanism to reduce background interference and highlight important damage features, and a BiFPNinspired feature fusion structure to strengthen multi-scale feature integration and improve information flow across network layers.
Abstract
Road damage significantly affects transportation safety, vehicle condition, and road maintenance efficiency. Traditional inspection methods are often costly and time-consuming, leading researchers to explore image processing and deep learning techniques for automated road damage detection. This work aims to investigate the application of an object detection approach for damage identification and detection on road surfaces. This work proposed an improved YOLOv5- based model for accurate road damage identification. The proposed approach enhances YOLOv5s by introducing three key improvements: (1) a P2 detection head to improve the detection of small and distant road damages, (2) the CBAM attention mechanism to reduce background interference and highlight important damage features, and (3) a BiFPNinspired feature fusion structure to strengthen multi-scale feature integration and improve information flow across network layers. Experimental results on the GRDD2020 dataset demonstrate the effectiveness of the proposed YOLOv5sP2-CBAM-BiFPN model.