融合注意力机制的混凝土桥梁多类病害语义分割  

Semantic Segmentation of Concrete Bridge Multiple Defects Combined with Attention Mechanism

作　　者：黄彩萍 余子行 李辉 HUANG Cai-ping;YU Zi-hang;LI Hui(School of Civil Engineering,Architecture and Environment,Hubei University of Technology,Wuhan 430068,China)

机构地区：[1]湖北工业大学土木建筑与环境学院,武汉市430068

出　　处：《公路》2024年第12期414-424,共11页Highway

基　　金：国家自然科学基金资助项目,项目编号51708188。

摘　　要：对混凝土桥梁病害图像进行语义分割时常存在精度不足、移动端设备计算能力有限等问题。建立了混凝土桥梁病害数据集(包括剥落、裂缝和露筋),在构建的多个语义分割模型中,分别采用深层卷积网络和轻量化卷积网络作为主干特征提取网络,引入不同的注意力机制模块进行多角度对比研究。结果表明,在对混凝土桥梁多类病害图像进行语义分割时,VGG16作为U—Net的主干网络时,其识别精度最高,平均交并比为80.37%,类别平均像素准确率为90.03%;轻量化卷积网络MobileNetV2—DeeplabV3的参数量显著减少,具有更快的检测速度,图像处理的速度达到了71.87帧/s,适用于病害实时检测;引入SE、CBAM、CA这3种注意力模块后,VGG16—U—Net和MobileNetV2—DeeplabV3的识别精度均得到了提高,其中CA模块能更好地引导模型识别出混凝土细微病害。Aiming at the problems in semantic segmentation of concrete bridge defects images,such as insufficient precision,and limited computing power of mobile devices,the concrete bridge defects data sets(including spallation,cracks and exposed reinforcement)is established.In the constructed multiple semantic segmentation models,deep convolutional network and lightweight convolutional network are used as the backbone feature extraction network,and different attention mechanism modules are introduced to carry out multi-angle comparative research.The comparison of experimental results shows that for the semantic segmentation of multi-class concrete bridge defects images,when VGG16 is used as the backbone network of U-Net,it achieves the highest recognition accuracy with a Mean Intersection over Union(MIoU)of 80.37%and a Mean Pixel Accuracy(MPA)of 90.03%.The lightweight convolutional network MobileNetV2-DeeplabV3+significantly reduces the number of parameters,resulting in faster detection speed of 71.87 frames/s,making it suitable for real-time defects detection.After introducing the SE,CBAM,and CA attention modules,both VGG16-U-Net and MobileNetV2-DeeplabV3 have achieved higher recognition accuracy,of which,the CA module can better guide the model to identify the subtle concrete defects.

关 键 词：桥梁工程 混凝土病害 注意力机制 深度学习 语义分割 

分 类 号：U445.71[建筑科学—桥梁与隧道工程]