Improved YOLOv8-based YOLO-DCNet for road surface damage detection

To address the issues of low detection accuracy for small road damage targets, poor adaptability to target deformations, and high missed detection rates in complex backgrounds, this paper proposes an improved road surface damage detection model, YOLO-DCNet, based on YOLOv8. First, the C2f module is integrated with deformable convolution (DCNv3) to enhance the model's ability to detect irregularly shaped road damage. Second, the CBAM attention mechanism is incorporated, combining spatial and channel attention to optimize feature extraction. Finally, the Dynamic Head is introduced to improve multi-scale feature fusion and detection capabilities, effectively enhancing the model's performance in detecting road damage at various scales. Experimental results on a road damage dataset show that the YOLO-DCNet model achieves a 2.7% improvement in mean Average Precision (mAP), a 2.6% increase in Recall (R), and a 3.2% increase in Precision (P) compared to the original YOLOv8n, resulting in more accurate detection.