基于YOLOv8-MCMA模型的道路缺陷检测应用研究  

Application Research of Road Defect Detection Based on YOLOv8-MCMA Mode

作　　者：徐克圣[1] 孙蓉 XU Kesheng;SUN Rong(School of Rail Intelligent Engineering,Dalian Jiaotong University,Dalian,Liaoning,116052,China)

机构地区：[1]大连交通大学轨道智能工程学院,辽宁大连116052

出　　处：《广西科学院学报》2025年第1期33-44,共12页Journal of Guangxi Academy of Sciences

基　　金：辽宁省教育厅科研经费项目(LJKMZ20220838)资助。

摘　　要：道路缺陷具有多尺度特征,导致其检测准确度不高。为改进这一问题,本文提出一种面向道路缺陷检测的轻量级多尺度卷积移动注意力模型(YOLOv8 Multi-scale Convolutional Mobile Attention,YOLOv8-MC-MA)。首先,采用MobileViT网络,可以使模型在减少参数量的同时保持较高的识别准确率;其次,使用内容感知的特征重组(Content-Aware Reassembly of Features,CARAFE)模块为上采样模块,以提升细小裂缝的检测能力;再次,设计多尺度倒置残差注意力(Multi-scale Inverted Residual Attention,MIRA)模块,增强模型对多尺度特征的敏感性;最后,将颈部的普通卷积替换为可变核卷积(Alterable Kernel Convolution,AK-Conv),以更好地捕捉不规则的裂缝信息,从而降低检测误差。实验结果表明,与YOLOv8n模型相比,本文提出的模型在Road Damage Detection Dataset、RDD2022_China和Crack-forest Dataset上的平均精确度均值@0.5(mAP@0.5)分别提高了3.7%、1.4%和2.6%,参数量减少了23.3%。与其他模型相比,该模型展现出明显优势,并对多尺度道路缺陷具有较强的适应能力。Road defects have multi-scale characteristics,resulting in low detection accuracy.In order to improve this problem,this article proposes a lightweight multi-scale convolutional mobile attention model for road defect detection(YOLOv8 Multi-scale Convolutional Mobile Attention,YOLOv8-MCMA).Firstly,the MobileViT network can make the model maintain a high recognition accuracy while reducing the number of parameters.Secondly,the Content-Aware Reassembly of Features(CARAFE)module is used as the up-sampling module to improve the detection ability of small cracks.Thirdly,a Multi-scale Inverted Residual Attention(MIRA)module is designed to enhance the sensitivity of the model to multi-scale features.Finally,the ordinary convolution of the neck is replaced by an Alterable Kernel Convolution(AKConv)to better capture irregular crack information,thereby reducing the detection error.The experimental results show that com-pared with the YOLOv8n model,the average accuracy@0.5(mAP@0.5)of the proposed model on Road Damage Detection Dataset,RDD2022_China and Crack-forest Dataset is increased by 3.7%,1.4%and 2.6%respectively,and the parameter amount is reduced by 23.3%.Compared with other models,this model shows significant advantages and has strong adaptability to multi-scale road defects.

关 键 词：计算机视觉 目标检测 道路缺陷检测 MobileViT网络 MIRA模块 YOLOv8 MCMA模型 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]