The Design Of Railway Track Deformation Detection System Based On Deep Learning

China’s railway development has become the lifeblood of the nation and an indispensable way of transportation.The railway track will be deformed due to long-term thermal expansion and contraction and natural disasters.Therefore,railway deformation detection has become an important part of railway inspection.This topic is the research on railway track deformation detection based on deep learning.A railway track deformation detection system based on improved YOLOv4 model is designed and completed,which improves the detection process of traditional YOLOv4 model,reduces the complexity of trunk feature extraction network,and improves the speed of the whole YOLOv4 model on the premise of ensuring the detection accuracy.The main content of this paper includes the following four parts.(1)Data set production and data enhancement:in the process of rail deformation image classification,considering that there are few original rail deformation image data sets,data enhancement operations are carried out on the original rail deformation image data.Through rotation,translation,mirror image,noise and color space transformation,the rail deformation data set is expanded.The rail deformation data is expanded from 2178 originally collected pictures to 10890,It provides data support for the design and improvement of YOLOv4 model.(2)Optimization of backbone feature extraction network:Based on the research of backbone feature extraction network of YOLOv4,the CSPDarkNet network is analyzed,the heavyweight CSPDarkNet network is changed into lightweight networks Mobilenetvl,Mobilenetv2,Mobilenetv3,and the optimization of PANet feature enhancement network is completed by using deep separable convolution instead of basic convolution process.After testing,the FPS is increased by about twice when the accuracy decreases little.(3)K-means clustering algorithm:in target detection,in order to obtain better IOU values between the prediction box and the real box,so as to locate the target more accurately,it is necessary to calibrate the anchor box in advance to cluster the target.In this paper,K-means clustering algorithm will be used to optimize the original data set,and the distance between the initial clustering centers should be as far as possible.The recognition rate is improved by 4.68%.(4)Addition of attention mechanism:according to the basic principles of Attention and Self-Attention and the process of focusing the input local information by the attention mechanism,the training process of YOLOv4 network prediction of rail deformation is improved.Through the comparative analysis of the addition of SE module,ECA module and CBAM module,it is concluded that the addition of CBAM attention mechanism improves the average accuracy by 0.38%.The production of rail deformation detection data set and network improvement are the core links of the design of rail deformation detection system in this paper.A feasible improvement scheme is proposed to improve the prediction speed and accuracy of YOLOv4 model,which can also be used as a reference for the improvement of other YOLOv4 detection models.