Research On Classification Of Railway Track Surface Defects Based On Computer Vision Technology

Railway transport plays an irreplaceable role in promoting the development of the national economy.With the speed and weight of trains continuing to increase,the orbital at this stage,is needed to be guaranteed by using more advanced orbital detection methods and higher quality orbit detection equipment,as an important part of railway transportation.The general orbital defects are mostly caused on the surface,then spread to inside.Therefore,the track surface defect detection becomes imperative.This paper chooses the computer vision technology to study the track surface defect classification,the research content includes the following aspects:(1)This paper analyzes the necessity of computer vision technology to deal with orbital defects through the elaboration of computer vision processing flow.The track defects are divided into cracks and scars,and the classification method is described in details.Image acquisition is the initial step of defect detection system,therefore describing of the image input processing lays the foundation to subsequent track defect image processing.(2)Image preprocessing includes image enhancement and image denoising.Image enhancement can increase image clarity and reduce the difficulty of computer recognition in late defect classification.The image enhancement methods contain Gamma gray scale transformation and histogram equalization.Those two transformation processes are discussed.By contrasting the histogram of the transformed image,the gray scale transformation is selected with contrast increasing and the histogram trend similar to the original image.Noise can be classified into salt and pepper noise and Gaussian noise,and three typical denoising methods of spatial domain filtering are used.Meanwhile,the median filter is chosen as the denoising method in line with the evaluation index of the defect image of PSNR.(3)The purpose of segmentation is to separate the orbital defect area from the track non-defective area,which includes three steps: background segmentation,target segmentation and morphological processing.The principle of segmentation is expounded in detail,and the image must meet the requirement of obvious defect and no fragmentation.(4)The purpose of edge detection is to extract the defect profile from segmentation.The segmentation process of four kinds of edge detection operators are described respectively.The Canny algorithm is selected as the edge detection method according to its function of preserving the original defect features and the image noise suppression ability after the edge detection.(5)BP neural network is selected as the defect image classifier,BP neural network structure can achieve a more complex classification due to its many layer achieves and close connection of layers you.In order to distinguish the crack and scar defect,roundness rate,rectangular rate,slenderness ratio as classification feature are used to design BP neural network.There are three input layer nodes in the network,corresponding to the selected orbital defect eigenvalues;seven hidden layer nodes,which constitute the middle convergence;two output layer nodes,corresponding to crack and scar 2 defects.Training the training through the training samples and input the identification of samples in the training of the network,classification results accurately.