【作者】 刘俊；

【导师】 王卫东； 杨敏捷；

【作者基本信息】 中南大学 ， 工程（专业学位）， 2022， 硕士

【摘要】 扣件系统作为轨道结构的重要零部件,其工作状态直接影响轨道结构的稳定性。本文研究地铁扣件图像采集与预处理、扣件弹条像素级分割、扣件状态判识等,创建基于深度学习的地铁扣件断裂状态判识系统,主要工作内容如下:（1）地铁扣件图像预处理。对比分析均值、高斯、中值3种滤波算法,选择5×5中值滤波模板对图像进行降噪处理;通过直方图均衡化和拉普拉斯锐化提高图像扣件区域与背景的对比度,减弱环境对图像的影响;通过多种数据增强方式扩充扣件图像负样本,构建地铁扣件图像样本库（包含2000张正常扣件图像、400张前拱断裂扣件图像和400张后拱断裂扣件图像）。（2）基于DeepLabv3+语义分割模型实现地铁扣件弹条的像素级分割。在语义分割模型实验中,Deep Lab v3+的训练集损失收敛速度、收敛值和验证集MIo U都优于U-Net模型。通过鲁棒性实验证明,基于Deep Lab v3+语义分割模型对判识亮度改变范围-25+75和高斯模糊σ5的图像具有良好鲁棒性。（3）分别针对扣件原始图像和语义分割图像,对比分析Mobile Net V2、Shuffle Net V2和Res Net50 3种深度学习分类模型的判识性能。对原始图像,Res Net50是最优判识模型,对正常扣件和后拱断裂扣件能较好判识,但存在将前拱断裂扣件误判为正常扣件的问题。对语义分割图像,各模型的判识效果均优于原始图像,且Mobile Net V2模型的判识效果在准确率、精确率、召回率和₂F分数4个评价指标均优于其他模型,能有效判识正常扣件、前拱断裂扣件和后拱断裂扣件3种伤损类型。（4）创建了基于深度学习的地铁扣件断裂状态判识系统,包括扣件图像采集、图像预处理、扣件弹条像素级分割、扣件状态判识和人机交互5个模块,实现了地铁扣件断裂状态判识系统集成。图42幅,表19个,参考文献121篇更多还原

【Abstract】 The fastener system is an important part of the track structure,and its working state has a direct impact on the stability of the track structure.This thesis studies the image acquisition and pre-processing,the pixel-level segmentation of fastener clips and the identification of fastener status.An identification system of subway fastener fracture state is thus established.The main contents are as follows:（1）Image pre-processing of subway fastener.Three filtering algorithms,namely the mean,Gaussian and median,are compared and analyzed,and 5×5 median filtering template is selected for noise reduction.The contrast between the image fastener area and background is improved using histogram equalization and Laplacian sharpening to weaken the influences of environment on the images.The negative samples of fastener images are expanded using various data enhancement methods to build the sample library of subway fastener images,including2000 normal fastener images,400 front arch fracture fastener images and400 rear arch fracture fastener images.（2）The pixel-level segmentation of subway fastener clips is realized based on the Deep Lab v3+semantic segmentation model.In the semantic segmentation model experiments,the loss convergence speed and convergence value of the training set and the MIo U of the validation set using Deep Lab v3+outperform those of U-Net model.The robustness experiments demonstrate that the Deep Lab v3+semantic segmentation model has high robustness to images with discerning brightness range-25+75 and Gaussian blurσ5.（3）Based on the original image of the fastener and the semantically segmented image,the recognition performance of three deep learning classification models,namely the Mobile Net V2,Shuffle Net V2 and Res Net50,is compared and analyzed.For the original image,Res Net50 is the optimal recognition model and can identify the normal fasteners and the rear arch fractured fasteners better.However,there is a problem of misjudging the front arch fractured fasteners as normal fasteners.For semantically segmented images,the recognition effect of each model is better than that of the original image.It is concluded that the recognition results using the Mobile Net V2 model are better than that of other models using four evaluation indicators,namely the accuracy,precision,recall and F₂ score.The model thus can effectively identify three damage types,namely the normal fasteners,front arch fracture fasteners and rear arch fracture fasteners.（4）A recognition system of subway fastener fracture state based on deep learning is built,including five modules of fastener image acquisition,image pre-processing,fastener clips pixel-level segmentation,fastener state recognition and human-computer interaction.It thus realizes the recognition system integration of subway fastener fracture state.更多还原