KGC-RTDETR: a road damage detection method integrating kan-based nonlinear feature enhancement and global–local cross-scale attention

Objectives To improve road damage detection in vehicle-mounted images affected by irregular distress morphology, background interference, and scale variation.Methods KGC-RTDETR was developed by integrating three complementary modules into RT-DETR: BasicBlock-KAN for nonlinear deep-feature enhancement, global–local spatial aggregation for texture and context modelling, and cross-attention blocks for selective cross-scale feature fusion. Ablation and comparative experiments used 4,288 road images containing 19,376 annotated instances.Findings KGC-RTDETR achieved mAP@0.5 of 91.8% and mAP@0.5:0.95 of 64.7%, exceeding the baseline by 3.9 and 7.1 percentage points, respectively. Compared with RT-DETR-R34, these metrics increased by 0.8 and 1.4 percentage points, while parameters decreased from 124.3 million to 68.7 million and inference speed increased from 50.3 to 72.7 frames per second. Pothole average precision increased from 87.5% to 97.2% relative to the baseline, although broader generalisation requires further validation.Novelty The proposed architecture coordinates nonlinear feature representation, within-scale global–local aggregation, and selective cross-scale interaction before decoding, improving detection accuracy and localisation quality while reducing model size and increasing inference speed relative to the RT-DETR baseline.