基于CAN-YOLOv11n的道路缺陷检测方法

Road Defect Detection Method Based on CAN-YOLOv11n

摘要: 针对道路缺陷检测中裂缝目标细长、尺度变化显著、背景纹理复杂以及小目标定位不稳定等问题，本文在YOLOv11n的基础上提出CAN-YOLOv11n道路缺陷检测模型。首先，在P4分支引入CAA模块，增强缺陷区域的方向性上下文响应，并抑制路面阴影、标线和纹理噪声干扰；其次，采用ASFF对多尺度特征进行自适应融合，提升细小裂缝、坑洞和修补区域在不同尺度下的表达能力；最后，在边界框回归中引入NWD-CIoU，使CIoU的几何约束与NWD对小目标位置偏移的鲁棒性形成互补。实验结果表明，在RDD2022中国区域数据集上，CAN-YOLOv11n的R、mAP50和mAP_50-95分别达到84.0%、88.6%和59.4%，较YOLOv11n基线分别提升1.8、2.4和0.3个百分点。所提方法能够减少复杂路面场景中的缺陷漏检，并提高道路缺陷的整体检测性能。

Abstract: To address the challenges of elongated cracks, large-scale variation, complex pavement textures, and unstable localization of small defects in road defect detection, this paper proposes CAN-YOLOv11n road defect detection model based on YOLOv11n. First, CAA is introduced into the P4 branch to strengthen directional contextual responses and suppress interference from pavement shadows, lane markings, and texture noise. Second, ASFF adaptively fuses multi-scale features to improve the representation of small cracks, potholes, and repaired areas at different scales. Finally, NWD-CIoU is adopted for bounding-box regression, combining the geometric constraints of CIoU with the robustness of NWD to positional deviations of small targets. Experiments on the China subset of RDD2022 show that CAN-YOLOv11n achieves 84.0% recall, 88.6% mAP50, and 59.4% mAP_50-95, outperforming the YOLOv11n baseline by 1.8, 2.4, and 0.3 percentage points, respectively. The proposed method reduces missed detections in complex pavement scenes and improves overall road defect detection performance.