Improved YOLOv8 Method for Multi-scale Pothole Detection

The timely detection and repair of road potholes are crucial for maintaining road safety. The variability in size of potholes, along with their difficulty to distinguish from the road surface, poses significant challenges to conventional object detectors. To tackle the issues of multi-scales and low localization precision, we introduced an innovative algorithm called RAW-YOLOv8, which is based on YOLOv8n. Firstly, this paper designs the C2f_RB module and integrates Dattention (deformable attention), allowing the module to grasp the spatial dependencies between the target area and the global to reduce the interference from the road surface background in pothole detection. Secondly, an AFPN-DBB-C2 (asymptotic feature pyramid network with diverse branch block and C2 layer) structure is utilized to reconstruct the neck of the YOLOv8n model, enhancing the model’s capability for feature extraction and integration. Finally, Wise-ShapeIoU is employed as the bounding box loss to assist the model in achieving more precise target localization and quicker convergence. The effectiveness of the proposed improvements is validated through comparisons with other existing models. Experimental outcomes indicate that this model exhibits commendable detection performance, with an increase in the average precision by 2.2% at mAP@0.5 and a 30% reduction in parameter count compared to the baseline model. This proposed method facilitates deployment on resource-constrained edge devices, providing valuable insights into road pothole detection technology and highlighting its potential in the field of road maintenance.