Road pothole detection algorithm based on improved YOLOv10n

Road pothole detection is crucial for traffic safety, road maintenance efficiency, service life extension, and intelligent transportation system development. This paper proposes DDW-YOLOv10n, an advanced algorithm to address the accuracy and irregular pothole detection issues. Firstly, the Dynamic Sampling (DySample) module is integrated to enhance the neck's feature fusion from a point-sampling perspective. Secondly, a new detection head fuses DCNv3 with Dynamic Head, combining attention mechanisms and leveraging DCNv3's dynamic sampling for better irregular defect prediction. Finally, the MPDIoU loss is optimized via Wise-IoU v3, using a dynamic, non-monotonic mechanism with gradient enhancement for precise localization, improving detection accuracy and model convergence. Compared with the original YOLOv10n, DDW-YOLOv10n achieves 1.8% increase in mAP@0.5, reaching 83.4%, and 5.7% improvement in accuracy, attaining 88.2%, while decreasing the number of parameters. Compared with mainstream object detection algorithms (Faster R-CNN, SSD, YOLO series and RT-DETR.), DDW-YOLOv10n exhibits superior performance in terms of both accuracy and model weight, demonstrating the optimal overall detection capabilities in complex environments, which exhibits significant potential for application in automated road condition monitoring.