基于增强多尺度融合YOLOv8的道路病害检测算法  

Road Disease Detection Algorithm Based on Enhanced Multi-scale Fusion YOLOv8

作　　者：高雪豪 吴建平[1] 韦杰 何旭鑫 余咏 GAO Xue-hao;WU Jian-ping;WEI Jie;HE Xu-xin;YU Yong(School of Information Science&Engineering,Yunnan University,Kunming 650504,China)

机构地区：[1]云南大学信息学院,云南昆明650504

出　　处：《计算机技术与发展》2024年第11期140-147,共8页Computer Technology and Development

基　　金：国家自然科学基金项目(62172354)。

摘　　要：在资源有限的户外,针对道路病害检测精准度和实时性不高的问题,提出一种多尺度融合的YOLOv8的道路病害检测算法。在主干网络中使用C2iAFF多尺度特征融合模块,缓解背景特征与目标特征之间分辨难的问题,提高对细长裂痕目标的检测能力;构造特征融合模块RFB_(3×3),对多尺度目标特征信息进行聚合提取,提高特征的表达能力;加入通道注意力机制SE,让模型学习重要的特征,提高模型的精准度;最后采用更优的归一化Wasserstein距离度量损失函数,使用NWD帮助小尺寸检测物体定位。实验结果表明,改进后的道路病害检测模型在仅增加0.3M参数量和0.4GFLOPs计算量的情况下,mAP50提高了2.4百分点,F1-Score提高了2.4百分点,达到了道路养护工作要求的检测精度和速度。In the outdoors with limited resources,a multi-scale fusion YOLOv8 road disease detection algorithm is proposed to solve the problem of low accuracy and real-time performance of road disease detection.The C2iAFF multi-scale feature fusion module is used in the backbone network to alleviate the problem of difficult discrimination between background features and target features,and improve the detection of slender crack targets;the feature fusion module RFB_(3×3) is constructed to aggregate multi-scale target feature information extraction to improve the expression ability of features;the channel attention mechanism SE is added to allow the model to learn important features and improve the accuracy of the model;finally,a better normalized Wasserstein distance measurement loss function is used,and NWD is used to help locate small-sized detection objects.Experimental results show that the improved road disease detection model increased mAP50 by 2.4 percentage points and F1-Score by 2.4 percentage points while only increasing the number of parameters by 0.3M and the calculation amount by 0.4GFLOPs,reaching the detection accuracy and speed required for road maintenance work.

关 键 词：道路病害检测 目标检测 YOLOv8 多尺度融合 注意力机制 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]