Pavement Crack Identification in UAV Images Based on Joint Context Information

Low-grade highway maintenance faces the challenge of high demand yet limited resources. Accurately identifying road damage is a key task to improve maintenance efficiency, which is crucial for addressing this demand–resource contradiction. To address this issue, YOLOv5s was selected as the foundational model due to its superior balance of detection accuracy, speed, and computational efficiency compared to other YOLO variants. Comprehensive optimizations were then implemented to further enhance its performance, including the development of a Global Context Squeeze (GS) module, a modified loss function, optimized Non-Maximum Suppression (NMS), and targeted image preprocessing strategies. The GS module is designed to effectively integrate contextual information, expand the receptive field, capture long-range dependencies, and strengthen feature extraction capabilities. A suburban road section in Shanghai with typical pavement damage was selected as the experimental site, where 8515 images were collected for model training and testing. Experiments demonstrated that the optimized YOLOv5s-G model achieved a mean average precision (mAP) of 90.7% for crack detection, a relative improvement of 18.6% over the original YOLOv5s. Furthermore, it outperformed models employing conventional optimization strategies, such as those with added small object detection layers or standard attention mechanisms. The superior performance of the YOLOv5s-G model significantly enhances pavement crack detection accuracy, offering technical support to improve low-grade highway maintenance efficiency and alleviate pressures from resource limitations.