Depth Image Segmentation Of Train Wheel Based On Deep Learning

As a key component of the railway track,the wheel alignment(wheel set)is directly related to the safety of train operation,and the railway department has strict maintenance standards for the wheel pair.At present,the detection of wheel treads in China is still mainly based on manual testing,with large workload and low detection efficiency.In order to adapt to the intelligent development of the railway,it is necessary to further study the wheel set tread segmentation.Thanks to the rapid development of hardware,the improvement of computing power and the continuous development of deep learning algorithms,image segmentation based on deep learning algorithm has a qualitative leap in performance and effect compared with traditional algorithms.In this paper,the algorithm of train wheel tread image segmentation based on deep learning is designed.This paper first introduces the research on the wheel tread detection method at home and abroad,and the development and advantages of deep learning.It introduces the basic theory of wheel and rail,the basic theory of image processing and deep learning.It laid the foundation for the research work in the subsequent chapters.Then,in order to highlight the target area in the image,the MSRCR algorithm is used to enhance the image of the original image,and the contour of the area to be detected is more obvious.Then,Canny edge detection is used to extract the edge information.The improved ellipse detection algorithm using arc selection and gradient information reuse is proposed.And it is used to pre-segment the wheel set area so that the tread area occupies the main part of the image,and then a tread area marking procedure based on Qt is used to mark the segmented image.These work is prepared for subsequent wheel set tread image segmentation.The U-Net algorithm has been improved for the segmentation of the wheel set image.In order to solve the problem of insufficient data in wheel tread segmentation,data augmentation is adopted to expand data size.The public data set is used to pretrain the model,and using migration learning and introducing L2 regularization in deconvolution layer is to prevent model over-fitting.The CRF is introduced to improve the accuracy of tread segmentation and it can more accurately realize feature localization.Finally,through the comparison with the traditional watershed algorithm,the experimental verification of segmentation in the tread segmentation dataset,the segmentation algorithm based on deep learning can reach 93.49% in accuracy.The experimental results show that the algorithm has good segmentation effect on the wheel tread.