Research On Identification Methods Of Rail Surface Defects In Image-based Rail Inspection System

Rail is an important component of railway line,its damage status directly affects the safety of railway transportation.With the high-speed,heavy-duty,and high-density operation of trains,rail surface defects caused by wheel-rail rolling contact fatigue are becoming more and more common.Timely detection of rail surface defects and take maintenance measures can prevent it from further developing into internal defects,thus ensuring the safe operation of the railway.Machine vision technology,with its advantages of high speed,high precision and non-contact,provides a new direction for the non-destructive testing of rail surface defects.Therefore,the machine vision recognition method for rail surface defects is studied in this paper by analysing the characteristics of rail images obtained by the image-based rail inspection system.First,according to composition and working process of the machine vision system,and combining with the acquisition requirements of track inspection images and the imaging principle of the camera,a portable image-based rail inspection system used to collect high quality rail images is constructed through the parameters calculation and selection of light source,industrial camera,lens,and industrial computer and other hardware devices.Secondly,the pixel grayscale statistical feature of each row in the rail image is analyzed,and the rail surface area in the image is extracted by combining the distribution curve of gray mean and standard deviation for each row,and an adaptive median filtering algorithm is used to filter the rail surface image.Then the formation mechanism and imaging characteristics of the rail surface features are analyzed,and by analyzing the single-scale and multi-scale Retinex methods for the problems existing in rail image enhancement,z-score standardization link of the reflection component is introduced to improve the traditional Retinex image processing framework,rail image is enhanced.Finally,through theoretical derivation and experiments under different conditions,the effectiveness of the enhancement method is verified.Then,the grayscale and gradient features of different regions in the rail image are analyzed,and the background smoothing filter is designed based on the idea of bilateral filtering.By smoothing different feature region of original image with corresponding scale which is adaptively adjusted by using local gray scale and gradient change information,the background image is obtained.Then image difference between original image and background image is made,and by setting dynamic threshold for the differential image,the segmentation of rail surface defects is realized.Finally,the effectiveness of segmentation method is verified by the rail images collected in different orbit environments.Finally,defect attributes are described by extracting the geometry and grayscale features of the defects region.And by using the Relief algorithm,defect features are selected and the features that are not related to the classification are filtered out.Then by using AdaBoost multi-classifier combination method and taking CART decision tree as a weak classification algorithm to design a combined classifier,rail surface defects classification is realized.Finally,statistics and analysis of experimental data for defect classification are made.The results show that the defects identification method in this paper can be applied to recognition of rail surface defects under low speed conditions.