SimCLR-YOLO:面向低标注依赖的公路病害检测算法OA
Sim CLR-YOLO:a road disease detection algorithm for low annotation dependence
针对公路病害检测中样本稀缺、弱纹理漏检率高及复杂背景干扰等难题,提出融合无监督预训练 SimCLR 与YOLOv8n的公路病害检测算法SimCLR-YOLO.首先,将SimCLR骨干替换为YOLOv8n的CSPDarknet,实现结构对齐与特征无缝迁移,解决网络结构不匹配导致的特征丢失问题.其次,将掩码重构融入对比学习,驱使编码器聚焦裂缝与坑槽的纹理轮廓,弥补对比学习对弱纹理特征的欠拟合缺陷.最后,嵌入CBAM注意力模块,引导模型在预训练阶段抑制路面纹理、阴影等背景噪声,增强关键区域的特征响应.在RDD2022公开数据集上的实验表明,SimCLR-YOLO算法的mAP@0.5达到89.4%,较基线模型提升1.7%,mAP@0.5:0.95提升1.1%,F1-score提升1.8%.不同标注比例下的验证结果显示,引入预训练权重在各梯度下均优于从头训练,在50%标注比例下,SimCLR-YOLO算法的mAP@0.5仍可达 87%,接近全量标注下的基线模型,显著降低了对标注数据的依赖.此外,改进仅在预训练生效,推理参数(3.01M)无额外开销,为样本匮乏条件下的公路病害实时检测提供了有效技术方案.
To address the challenges in road disease detection,such as sample scarcity,high miss rates for weakly textured targets,and complex background interference,a road disease detection algorithm named SimCLR-YOLO is proposed to integrate unsupervised pre-training SimCLR with YOLOv8n.Firstly,the SimCLR backbone is replaced with YOLOv8n's CSPDarknet to achieve structural alignment and seamless feature transfer,resolving the feature loss issue caused by network structure mismatch.Secondly,masked reconstruction is incorporated into contrastive learning,forcing the encoder to focus on the texture contours of cracks and potholes,thereby compensating for the under-fitting of contrastive learning to weakly textured features.Finally,a CBAM attention module is embedded to guide the model in suppressing background noise such as road texture and shadows during the pre-training phase,while enhancing feature responses in critical regions.Experimental results on the public RDD2022 dataset show that the SimCLR-YOLO algorithm achieves an mAP@0.5 of 89.4%,which is 1.7%higher than the baseline model,with a 1.1%improvement in mAP@0.5:0.95 and a 1.8%increase in F1-score.Validation under different annotation ratios demonstrates that the scheme incorporating pre-trained weights outperforms training from scratch across all data gradients.At a 50%annotation ratio,the SimCLR-YOLO algorithm still achieves an mAP@0.5 of 87%,approaching the baseline model under full annotation,significantly reducing the dependence on annotated data.Moreover,the improvements are only effective during pre-training,with no additional overhead in inference parameters(3.01M),providing an effective technical solution for real-time road disease detection under sample-scarce conditions.
向天明;燕芳;李嘉辉;孙瑞
内蒙古科技大学 自动化与电气工程学院,内蒙古 包头 014010内蒙古科技大学 自动化与电气工程学院,内蒙古 包头 014010内蒙古科技大学 自动化与电气工程学院,内蒙古 包头 014010内蒙古科技大学 自动化与电气工程学院,内蒙古 包头 014010
信息技术与安全科学
公路病害YOLOv8nSimCLR无监督预训练低标注依赖预训练
road diseaseYOLOv8nSimCLRunsupervised pre-traininglow annotation dependencepre-training
《液晶与显示》 2026 (8)
1158-1174,17
内蒙古自然科学基金(面上项目)(No.2025MS06002)内蒙古自治区重点研发和成果转化计划(No.2025SYFHH0062)Supported by Natural Science Foundation of Inner Mongolia Autonomous Region(No.2025MS06002)Key Research and Development and Achievement Transformation Program of Inner Mongolia Autonomous Region(No.2025SYFHH0062)
评论