In response to the current issues of insufficient automation and low efficiency in road inspection,this study aims to enhance the efficiency and accuracy of UAV-based road inspection.Based on YOLOv8s,an improved model,AC-YOLO,is proposed specifically for UAV aerial scenarios.First,the model integrates a dynamic large-kernel convolu-tional attention mechanism,LSK-attention,into the backbone network to expand the receptive field and improve the model's accuracy in detecting road crack areas.Second,a multi-scale feature fusion strategy is introduced in the neck structure by incorporating the BiFPN network,enhancing the model's ability to detect fine cracks.Finally,the loss func-tion is replaced with WIoUv3,optimizing the gradient allocation strategy to enable the model to focus more precisely on crack regions.Experimental validation on the UAV-PDD2023 dataset demonstrates that the improved AC-YOLO achieves an accuracy of 0.895,representing a 0.128 increase compared to the original model.The mAP50 reaches 0.791,reflecting an improvement of 0.071,while the Fl score increases by 0.051.Moreover,the model's size is reduced by 8.5％,and the FPS reaches 129,showing a 4％improvement.The model's generalization performance is verified across multiple tasks,and the experimental results confirm that AC-YOLO offers superior detection capabilities,making it highly effective for UAV-based road crack detection.
万方数据
维普
