文献类型：期刊文献

中文题名：基于生成对抗网络的隧道裂缝自动分割算法研究

英文题名：Research on Automatic Segmentation Algorithm of Tunnel Cracks Based on Generative Adversarial Network

作者：李子仡[1];饶志强[1];常惠[2];李益晨[1];丁璐[1];方建军[1]

第一作者：李子仡

机构：[1]北京联合大学城市轨道交通与物流学院,北京100101;[2]中国科学院自动化研究所,北京100190

第一机构：北京联合大学城市轨道交通与物流学院

年份：2023

卷号：45

期号：5

起止页码：136-142

中文期刊名：铁道学报

外文期刊名：Journal of the China Railway Society

收录：CSTPCD;;Scopus;北大核心:【北大核心2020】；CSCD:【CSCD2023_2024】；

基金：北京市自然科学基金(L221015)。

语种：中文

中文关键词：隧道裂缝分割;GAN网络;U-Net网络;自动分割

外文关键词：tunnel crack segmentation;GAN network;U-Net network;automatic segmentation

摘要：传统裂缝识别技术需要大量带标签的裂缝图像作为试验数据集,为减少裂缝图像标注的工作量,利用生成对抗网络图像分割的特性与优势,构建一种Crack-GAN网络用于隧道裂缝自动分割。Crack-GAN网络结构集成2个模块:融合残差化U-Net网络的生成器网络和利用全卷积网络生成置信图的判别器网络。首先U-Net模块使用密集的残差模块来生成保留细粒度信息的深层表示,然后判别器来判断输入真假,并以端到端的方式训练,再经过生成对抗模型之间不断迭代,使生成模型达到分割裂缝的最优状态。试验表明,Crack-GAN网络在自制铁路隧道裂缝数据集上的像素准确性为98.35%,精准率为71.23%,召回率为80.78%,F 1得分为75.98%,单次检测时间80 ms,综合表现优于U-Net和CrackSegNet。
Traditional crack recognition technology requires a large number of crack images with annotations as experimental data sets.In order to reduce the workload of crack image annotation,a crack-GAN network was constructed for automatic segmentation of tunnel cracks by using the characteristics and advantages of generative adversarial network.The Crack-GAN network structure integrated two modules:the generator network of U-NET network integrating residual modules and the discriminator network generating confidence graphs using full convolution network.Firstly,the U-NET module used dense residual modules to generate deep representations that retain fine-grained information.Secondly,the discriminator was used to determine the true or false input.Finally,through end-to-end training and continuous iteration between generative adversarial models,the generated model reached the optimal state of crack segmentation.The test results show that compared with existing networks under the same conditions,the Crack-GAN network exhibits 98.35%pixel accuracy,71.23%accuracy,80.78%recall rate and 75.98%F 1 score in self-made railway tunnel crack data set,with single detection time of 80ms and better overall performance than U-NET and CrackSegNet.