Rail Surface Defects Of Machine Vision Inspection System

Railway construction has been developed to harsh environmental areas with the rapid development of China's railway in recent years,Such as the Qinghai-Tibet Railway,as well as the Siberian region to build the railway to Russia.And with the increase in passenger and cargo traffic,train frequency and train speed are getting higher and higher.So the safety requirements of the train even more stringent.As an important part of the railway.The track will be with the use of the growth of time to be consumed.Its surface will appear cracks and other issues.If not timely maintenance,it will cause serious train accidents.At present,rail maintenance work mainly rely on manual inspection,completed by visual inspection.This inspection method is inefficient,and subject to light,subjective factors.The ultrasonic,eddy current detection means,does not apply to surface inspection,but for the internal track inspection.Due to the need to spray the magnetic powder in the surface of the track,and then after the magnetization of the tracks according to the magnetoresistance change,the magnetic particles show scars.So the magnetic particle method detection efficiency is low,it need a lot of magnetic powder for testing,not suitable for the need for rapid detection,long-distance rail detection.Therefore,the method based on machine vision detection has a lot of application prospects.First of all,this paper introduces the research background and necessity of rail scar testing,and the development trend and research status at home and abroad.And then,analysis of the causes of surface scars.The main reason is the track by the train in different directions and track the manufacture of materials containing impurities,or ingot cooling caused by the formation of inequality.The whole structure and experimental simulation platform are set up.Absolute photoelectric encoder counter output bit except the lowest bit,the other bit after "NOR" logic operation,and then the lowest logic "and" to ensure that the least significant bit for the output bit,so that only the binary count of 1 as the output.This signal acts on the camera shutter,verifying the waveform of this signal and synchronizing the camera's captured image.Then the article according to the illumination requirements,calculate the luminous flux range of different light sources,and finally select the LED light as a light source,The light source illuminates the rail surface in the form of a line light source.The camera parameters are determined by the line scan frequency(5952 Hz or more)and the number of pixels(107 or more),and the lens parameters are determined according to the imaging theorem and the camera imaging parameters.In the process of preprocessing,firstly,the image median filter is used to remove the salt and pepper noise,and then according to the track surface bright,dark background characteristics,statistics value of 255 pixels per line number after the threshold value of the track surface.Take 350 pixels as the vertical baseline.And the intersection point of the curve is the orbital surface boundary coordinates.So as to extract the surface of the track.In order to reduce the background of the computing process,improve operational efficiency.In the target rapid detection stage,first,statistical average pixel gray value in per line.According to the assumption that the product of the mean and the adjustment parameters is the uniform gray level(90),the adjusted parameter curve is obtained.And then multiply with corresponding of each row of pixels to obtain a gray-scale compensation image.In order to remove the shadows and reflective gray uneven,affect the follow-up treatment.Due to the difference of scar characteristics,direct threshold segmentation can not achieve the desired effect.After adding the bottom cap operation can get a clear scar map,that is,with the disc radius of 15 elements of the first corrosion,and then reducing the original map.It can solve the problem of not clear scratches,while stripping injury is not affected,but the processing of scars and gray background close.Then the image grayscale in the range of 0.1 to 0.3 to stretch to 0 to 1,and then 0.7 threshold;The target precise positioning stage utilizes the morphological opening operation,the structure element is the radius of the disc 1,agglutinate the misty area.After filling the holes by changing the background pixels,seed-filling method was used to mark the wound-connected region.According to the scar area of 2 square millimeters,calculated with the figure area pixels to reach 12 or more.Counting the acreage of wound-connected region and select the main scars.Finally,we use the Sobel algorithm to extract the edge of the scar.The edge coordinates are obtained.The red component of the coordinate is 255,and the green and blue components are 0 on the RGB image where the gray image is converted to the pseudo gray image,so as to achieve the red edge prompting effect,which in practice can quickly and accurately display and extract the scar.Finally,the parameters of the scar were extracted,and a BP neural network classifier was designed according to the regularity of the rectangularity and circularity parameters.After normalizing the training sample input,set the input to the implicit layer purelin,implicitly to the output layer logsig as the transfer function.Using the steepest gradient descent method to propagate the training function,training 5000 times,error precision 0.01 and training rate 0.01.Take another set of test samples,test after the normalized,and the accuracy of the classifier is obtained.