摘要
针对路面病害检测存在的病害尺度差异大、细小病害特征提取难、病害在图像中占比小等问题,提出一种基于改进YOLOv8的路面病害检测方法。以YOLOv8s网络结构为基础,通过引入通道注意力机制和无跨步卷积网络结构,构建了一种无信息丢失的下采样网络模块,在剔除背景冗余信息的同时更多地保留了病害纹理特征;通过构建基于PANet的多尺度自适应特征融合网络,增强了网络浅层特征捕获能力,实现了不同尺度特征信息的高效融合;采用Focal Loss损失函数,对各样本赋予对应的权重,缓解了正负样本不平衡问题。实验表明:所提方法在RDD2020和RDD2022数据集上的平均精度分别达到57.1%和52.8%,与YOLOv8s模型相比分别提升3.2和0.6个百分点,整体性能优于YOLOv5等其他检测网络。
To overcome big differences in pavement damage scale,difficulties in feature extraction of small damage and the small proportion of defects in images in pavement disease detection,this paper proposes a pavement damage detection method based on improved YOLOv8.First,based on the YOLOv8s network structure,a down-sampling network module without information loss is built by introducing the channel attention mechanism and the no step convolutional network structure,removing the background redundant information and retaining more pavement disease texture features.Second,the ability of the network to capture shallow features is enhanced by building a multi-scale adaptive feature fusion network based on PANet,and the efficient fusion of feature information at different scales is realized.Finally,the Focal Loss function is employed to assign corresponding weights to each sample,which alleviates the imbalance between the positive and negative samples.Our experiments show the proposed method achieves an average precision of 57.1%and 52.8%on the RDD2020 and RDD2022 datasets,up by 3.2 and 0.6 percentage points respectively compared with those of the YOLOv8s model,and it performs better than other detection networks such as YOLOv5.
作者
邓天民
李亚楠
李庆营
李宇航
王含笑
DENG Tianmin;LI Yanan;LI Qingying;LI Yuhang;WANG Hanxiao(College of Traffic and Transportation,Chongqing Jiaotong University,Chongqing 400074,China;Shandong High-Speed Engineering Test Co.,Ltd.,Jinan 250000,China)
出处
《重庆理工大学学报(自然科学)》
CAS
北大核心
2024年第4期138-145,共8页
Journal of Chongqing University of Technology:Natural Science
基金
重庆市技术创新与应用发展项目(CSTC2020JSCX-CYLHX007,cstc2020jscx-cylhX0005)
山东省交通运输科技计划项目(2020-MSI-041)。
关键词
路面病害检测
多尺度特征融合
无跨步卷积网络
ASFF
pavement damage detection
multi-scale feature fusion
no step convolutional network structure
ASFF