A novel road damage detection model with efficient attention and Dynamic Snake Convolution

Road damages detection is crucial for ensuring traffic safety, optimizing maintenance costs, and extending the service life of roads. However, it faces three key challenges: (1) damages and background often have similar pixel intensities, making them hard to distinguish; (2) damage types vary greatly in shape and size, increasing the difficulty of robust feature extraction; and (3) road interferences such as water stains, shadows, or markings can easily cause false detections. To address these problems, we propose Bi-level Routing Attention, Snake Convolution, and Wise Intersection over Union enhanced You Only Look Once version 8 (BSW-YOLO), which integrates three targeted modules into the You Only Look Once version 8 (YOLOv8) framework. First, the Bi-level Routing Attention with DropKey (BRA-DropKey) highlights true damage features and suppresses background noise, solving the similarity in pixel intensities between damages and surroundings. Second, dynamic Snake Convolution (SnakeConv) captures geometric contours for fine-grained features, improving adaptability to diverse shapes and sizes. Third, Wise Intersection over Union (Wise-IoU) loss refines anchor box quality, reducing false detections from road interferences such as water stains and shadows. Experiments conducted on the Road Damage Dataset (RDD) 2022, a benchmark dataset for road damage detection, demonstrate that BSW-YOLO achieves a mean Average Precision at an intersection over union threshold of 0.5 ([email protected]) of 90.5%, significantly outperforming other baseline models and road damage detection methods.