Research On Rail Surface Defect Detection Based On Machine Vision

China’ railway operation speed ranks first in the world and China has the largest railway network and transportation system in the world.With the rapid development of the railway industry,it also puts forward higher requirements for the safety of railway transportation.Steel rail is an indispensable part of the railway system.In all railway accidents,about 30% of the accidents are caused by rail defects,so the detection of rail defects is an important link to ensure the safety of railway hub.At present,defects detection methods such as manual detection,ultrasonic and electromagnetic detection have some limitations.In recent years,machine vision has made great achievements in defect detection.High-cost performance,high efficiency and high automation make more and more people use machine vision method for product detection.YOLO algorithm is popular because of its fast detection speed and good real-time performance.YOLOv4 is improved on the basis of YOLOv3.This paper designs a defect detection system based on YOLOv4 algorithm.In the actual detection process,it can be suspended at the rear of the flaw detection vehicle.Although the detection speed of YOLO algorithm is fast,the recognition rate of defects with small size on rail is low.Considering that the size of rail head is not large,this paper improves and optimizes the design based on YOLOv4.Specific improvements are as follows: The feature layer used for large target detection in the removed YOLOv4 algorithm is added with Dense Net module,which can use the obtained feature information for many times,and the detection effect is better for the smaller target;Considering that the improved feature layer is changed into two layers,the K-means algorithm can be used for the rail defect data set,and the size of the prior box can be obtained through cluster analysis.In this paper,the improved YOLOv4 algorithm is applied to the steel rail defect detection system.After the experimental platform is built,the rail crack defect is taken as the target to carry out the experiment.Experimental results show that the improved algorithm improves the effect of small target crack defect recognition.Compared with YOLOv4,the accuracy of the improved YOLOv4 network model reaches 95.1,which is improved by 0.6%.The average accuracy of the improved YOLOv4 network model is improved by 1.5%.Although the detection speed is 0.011 second slower,it can still meet the requirements of the system for real-time monitoring.