Improved YOLOv10 for crack detection in road maintenance operations

In road maintenance, accurate and efficient crack detection is crucial, but existing technologies struggle with large-scale applications. This paper proposes an improved YOLOv10 crack detection algorithm. Specifically, a triple attention mechanism is added after the C2F module in the feature extraction network to enhance crack feature extraction in complex backgrounds and reduce missed detections. A dynamic snake convolution module is embedded in the deep network to improve the model's ability to identify irregular cracks by focusing on complex geometric features. An adaptive low-pass filter generator is used in the upsampling part of the neck fusion network to optimize prediction box accuracy and reduce positioning errors. Additionally, a SIoU loss function is introduced during regression to accelerate model convergence by quickly aligning prediction boxes. Experiments show that the proposed method achieves an mAP50 of 83.70%, mAP50- 95 of 45.40%, and recall of 60.50%. On the RDD dataset, it reaches an mAP50 of 53.60%, mAP50-95 of 22.80%, and recall of 58.20%. This method maintains YOLOv10's high-speed detection while improving crack detection accuracy and recall.