基于改进U-Net网络的半监督裂缝分割方法  

A Semi-supervised Crack Segmentation Method Based on Improved U-Net Network

作　　者：罗杨 万黎明 李理[1] 刘知贵[1,2] LUO Yang;WAN Li-ming;LI Li;LIU Zhi-gui(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621000,China;School of Information Engineering,Southwest University of Science and Technology,Mianyang 621000,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621000 [2]西南科技大学信息工程学院,四川绵阳621000

出　　处：《计算机技术与发展》2022年第12期179-184,共6页Computer Technology and Development

基　　金：国家自然科学基金(U21A20157)。

摘　　要：裂缝反映了结构的受力状态,是结构健康检测的重要关注对象之一。基于数字图像利用深度学习方法进行结构表面裂缝自动识别具有速度快、精度高等优势,不过深度学习方法严重依赖像素级标注信息,为此,提出一种基于半监督学习的改进U-Net方法。使用特征提取能力更佳的残差网络作为主干特征提取网络代替U-Net中由卷积层和池化层进行简单堆叠而成的下采样部分;在主干网络中插入池化窗口长且窄的条带池化注意力辅助下采样进行特征的细化,增强特征提取能力;针对裂缝图像中裂缝区域的亮度普遍暗于背景区域的情况,网络中的池化操作均采用平均池化使网络能更好地处理裂缝图像;利用半监督学习,在训练时同时训练两个网络并利用其分割结果相互监督从而使深度学习分割方法降低对标签数据的依赖度。改进的U-Net分割方法在自建裂缝数据集上进行了对比实验,结果表明,相较于原始U-Net网络,改进方法具有更高的分割精度,训练时可使用更少的标签数据。Cracks reflect the stress state of a structure and are one of the important objects of concern for structural health inspection.The automatic recognition of cracks based on digital images using deep learning methods has the advantages of high speed and accuracy,but the deep learning methods rely heavily on pixel-level annotation,so an improved U-Net method based on semi-supervised learning is proposed.A residual network with better feature extraction capability is used as the backbone feature extraction network instead of the down-sampling part of U-Net,which consists of simple stacking of convolutional and pooling layers.A strip pooling attention with long and narrow pooling window is inserted into the backbone to assist down-sampling for feature refinement and enhance feature extraction capability.For the situation that the brightness of crack region in a crack image is generally darker than that of the background region,all of the pooling operation in the network is average pooling so that the network can better handle the crack images.Using semi-supervised learning,two networks are trained at the same time and their segmentation results are used to supervise each other so that the deep learning segmentation method can reduce the dependence of labeled data.The improved U-Net segmentation method is tested on a self-built crack dataset,the experimental results show that the proposed method has higher accuracy than the original U-Net network and fewer label data needed for training.

关 键 词：裂缝分割 半监督 注意力机制 深度学习 U-Net 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]