GMSA-YOLO for multiscale road defect detection in UAV-based maintenance

To tackle the difficulties of recognizing multi-scale targets, dealing with complex backgrounds, and adapting to illumination changes in UAV-based road monitoring tasks, this paper presents an improved detection model based on the YOLOv13 framework, named GMSA-YOLO (Global-grouped and Multi-Scale Attention YOLO). The proposed model integrates a Global Grouped Coordinate Attention (GGCA) module into the backbone network, which captures global spatial correlations in both height and width dimensions, thereby enhancing feature representation and contextual modeling abilities. In the feature fusion phase, a Multi-Scale Dilated Fusion Attention (MDFA) module is designed to adaptively combine multi-scale features through multi-dilation convolutions and attention-weighted fusion. Experimental results on the public UAVROAD dataset show that GMSA-YOLO achieves remarkable performance improvements while maintaining low computational complexity of 5.7 Gflops. The model attains mAP@0.5 of 92.9% and mAP@0.5:0.95 of 63.4%, outperforming mainstream detection methods. Ablation studies further confirm the effectiveness and complementarity of the GGCA and MDFA modules in enhancing global features and fusing multi-scale information. This research offers a high-accuracy and high-efficiency solution with strong engineering applicability for UAV-based automated road inspection.