改进YOLOv8的路面缺陷检测算法  

The pavement defect detection algorithm of YOLOv8 was improved

作　　者：周建新 张媛 贾梓涵 何洋 Zhou Jianxin;Zhang Yuan;Jia Zihan;He Yang(School of Electrical Engineering,North China University of Science and Technology,Tangshan 063200,China)

机构地区：[1]华北理工大学电气工程学院,唐山063200

出　　处：《电子测量技术》2024年第19期146-154,共9页Electronic Measurement Technology

基　　金：河北省自然科学基金(F2018209201)项目资助。

摘　　要：为了减少背景干扰对路面缺陷检测的影响,解决对小尺寸细长裂缝能够提取的特征十分有限的问题,本文基于YOLOv8模型进行了改进。首先,将网络中的C2f融合动态蛇形卷积设计了C2f-Dysnake模块改善对目标形状和边界的敏感性,增强了对细长裂缝的特征提取能力;其次,将重参数化泛化特征金字塔网络RepGFPN与动态上采样器DySample结合构成新的颈部网络RDFPN,增加了对低层特征图的关注度,提升了对小目标的检测能力;最后在主干网络中加入MPCA注意力机制,捕捉不同尺度上的位置关系,提高主干网络的特征提取能力。实验结果表明,改进算法在RDD2022数据集上与原算法相比mAP50提高了2.3%,检测速度达到了98 fps,综合考量较其他算法有明显优势,验证了该方法的有效性和优越性。In order to reduce the influence of background interference on pavement defect detection and solve the problem that the features that can be extracted from small-sized slender cracks are very limited,this paper is improved based on the YOLOv8 model.Firstly,the C2f-Dysnake module was designed by fusing the C2f in the network with dynamic serpentine convolution,which improved the sensitivity to the shape and boundary of the target and enhanced the feature extraction ability of the slender cracks.Secondly,the reparameterized generalization feature pyramid network RepGFPN and the dynamic upsampler DySample were combined to form a new neck network RDFPN,which increased the attention to the low-level feature map and improved the detection ability of small targets.Finally,the MPCA attention mechanism is added to the backbone network to capture the position relationship at different scales and improve the feature extraction ability of the backbone network.Experimental results show that the improved algorithm improves mAP50 by 2.3%and reaches 98 fps on the RDD2022 dataset,and the detection speed reaches 98 fps,which has obvious advantages over other algorithms and verifies the effectiveness and superiority of the proposed method.

关 键 词：YOLOv8 动态蛇形卷积 注意力机制 路面缺陷检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN919.8[自动化与计算机技术—计算机科学与技术]