YOLO-RCS: A Road Crack Segmentation Model with Deformable Convolution and Cross-stage Feature Fusion

To address the challenges of false positives and missed detections in multi-scale and small crack segmentation under complex environments by proposing an enhanced instance segmentation model, YOLO-RCS (YOLOv10s Road Crack Segmentation), tailored for surface crack detection. The model incorporates the DCNv4 (Deformable Convolutional Networks v4) module into the backbone to improve feature extraction and enhance localization accuracy. Additionally, a novel C3FB module, combining the C3 module with the FocalN extBlock structure, replaces the original C2f module in the YOLOv10 neck, reducing parameter count while boosting segmentation performance. WIOU loss function with a dynamic focusing mechanism is also introduced to better guide bounding box regression, improving precision and mean Average Precision (mAP). Experimental results on the CrackSeg9k dataset demonstrate that YOLO-RCS achieves a precision of 91.60/0, a recall of 89.3%, an F1 score of 90.5% and an mAP of 90.0%. This outperforms the YOLOv10s baseline by a significant margin. The model demonstrates robust segmentation performance across diverse crack types and scales. On the Crack-BPHDR dataset, YOLO-RCS maintains its strong generalisation ability, achieving a precision of 87.4%, a recall of 91.3%, an F1-score of 89.3% and an mAP of 86.0%. YOLO-RCS outperforms the most advanced crack segmentation models in terms of accuracy, robustness, and generalisation capabilities.