摘要
针对路面裂缝形态复杂、易受环境干扰,且检测存在精度与轻量化不平衡等问题,本文提出一种自适应特征的轻量化路面裂缝检测方法。首先,根据裂缝狭长且跨度大的特点,设计了裂缝高效注意力机制,通过压缩特征维度,以捕捉其长距离空间依赖。其次,构建动态采样金字塔进行自适应采样和提取目标特征,以增强对异构裂缝特征的表示能力。然后,改进HGNet_GS轻量化主干网络,并提出了轻量化检测头,显著降低了计算冗余;采用Powerful IoU损失函数解决框锚膨胀问题并提升小模型的收敛速度。此外为验证模型泛化性,自建了民用路面缺陷数据集,其中包含不同光照条件下路面缺陷共计2985张。实验结果表明,与基准模型YOLOv8n相比,本文模型参数量和计算量分别减少了50%和52%。在自建数据集上,mAP50和mAP95分别提升了5.4%和4.1%;在公开的RDD2022数据集上,mAP50和mAP95分别提升了2.1%和1.5%。该模型已应用于边缘设备并完成工程作业测试,验证了其能够满足轻量化路面裂缝检测的工程应用需求,为自动化道路维护提供了技术方案。
To address the issues of complex crack morphologies,environmental interference,and the imbalance between detection accuracy and model lightweight requirements in road surface inspection,this paper proposed a lightweight road crack detection method with adaptive feature extraction.First,a Crack Efficient Attention(CEA)module was designed based on the slender shape and large span characteristics of cracks,compressing feature dimensions to capture long-distance spatial dependencies.Second,a Dynamic Sampling Feature Pyramid Network(DSFPN)was constructed for adaptive sampling and target feature extraction,enhancing representation capability for heterogeneous crack features.Third,the HGNet_GS lightweight backbone network was improved,and a CEA Group Head(CGHead)was proposed,significantly reducing computational redundancy;the PIoU(Powerful IoU)loss function was adopted to solve anchor box expansion problems and improve convergence speed for small models.Additionally,a civilian road defect dataset containing 2985 images under various lighting conditions was established to validate model generalization.Experimental results show that compared with the baseline YOLOv8n model,the proposed method reduces parameters and computational cost by 50%and 52%,respectively;on the selfbuilt dataset,mAP50 and mAP95 increase by 5.4%and 4.1%;on the public RDD2022 dataset,these metrics improve by 2.1%and 1.5%.The model has been deployed on edge devices and verified through engineering tests,demonstrating its capability to meet practical requirements for lightweight road crack detection and providing a technical solution for automated road maintenance systems.
作者
刘媛媛
朱凯
顾志辉
岳猛
王靖智
朱路
LIU Yuanyuan;ZHU Kai;GU Zhihui;YUE Meng;WANG Jingzhi;ZHU Lu(School of Information and Software Engineering,East China Jiaotong University,Nanchang 330013,China)
出处
《光学精密工程》
北大核心
2026年第2期336-351,共16页
Optics and Precision Engineering
基金
国家自然科学基金项目(No.62366015,No.61963016,No.61967007)。
