基于YOLOv8-MCMA模型在道路缺陷检测的应用研究

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

关键词:  计算机视觉  目标检测  道路缺陷检测  MobileViT  MIRA  YOLOv8-MCMA

Key words:  computer vision  object detection  road defect detection  MobileViT  MIRA  YOLOv8-MCMA

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-MCMA）。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，AKConv），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%。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。

Abstract:

The detection precision for road defects, which come in various sizes has often been insufficient. This paper proposes a lightweight multi-scale convolutional mobile attention model（YOLOv8-MCMA） for road defect detection to tackle this issue. Firstly, the model integrates the MobileViT structure, maintaining high recognition accuracy even with a reduced parameter count. Second, it employs the Content-Aware Reassembly of Features (CARAFE) as an up-sampling module, focusing on the detection of small-scale crack images. Additionally, a Multi-scale Inverted Residual Attention (MIRA) module is introduced to enhance the model's sensitivity to features across different scales. Finally, the traditional convolution in the model's neck is replaced with an Alterable Kernel Convolution (AKConv), which better captures irregular crack information, thus reducing detection errors. The experimental results show that in the Road Damage Detection Dataset, RDD2022_China and Crack-forest Dataset, the value of mAP@0.5 increased by 3.7%, 1.4% and 2.6%, respectively, compared with the YOLOv8n model. Compared to the YOLOv8 model, there is a 23% reduction in parameter count. The YOLOv8-MCMA model demonstrates significant advantages over other algorithms and shows strong adaptability to detecting multi-scale road defects.