Crack Segmentation Model Based on Deformable Convolution and Cross‐Stage Feature Fusion Network

ABSTRACT To address the issues of false positives and missed detections of multi‐scale cracks and small targets in complex environments, this paper proposes an enhanced YOLOv10 instance segmentation network named YOLO‐RCS (YOLOv10s road crack segmentation), specifically designed for segmenting surface cracks. YOLO‐RCS utilizes the DCNv4 module to enhance feature extraction in the backbone network, improving the accurate localization of surface crack segmentation. Additionally, we introduce a novel C3FB structure (an efficient fusion of the C3 module and FocalNextBlock structure) to replace the C2f module in YOLOv10's neck network, aiming to reduce the number of parameters while enhancing model accuracy. Finally, we improve the original loss function to the WIOU loss function, which increases the model's precision and mean average precision (mAP) for segmenting surface cracks. Experimental results show that our model achieves an mAP50 of 90.0% on the surface crack segmentation dataset Crackseg9k, a 5.0% improvement over the original algorithm, with a precision of 91.6%, demonstrating excellent segmentation performance. Compared to some mainstream object detection algorithms, our proposed method also exhibits certain advantages.