| The damage of the rail surface not only affects the comfort and stability of train operation,but also endangers the safety of the train.In recent years,with the speed increase of China’s wired railway and the development of heavy-haul railway and high-speed railway,the track line will be in service for a long time.Rail is prone to various damages due to improper operation of train driver,environmental factors and emergency braking of train.In addition,the surface damage of rail will only be more and more after long-term use.Therefore,it is particularly important for the railway to detect the damage of rail surface in time,identify and classify it,and take corresponding maintenance measures according to different damage categories,which can ensure the safety,comfort and continuous operation of the railway.Rail surface damage identification based on image processing has become an important method of rail surface damage identification because of its advantages of high precision,full automation and no contact.In this dissertation,the rail surface image is processed to detect and identify the damage in the surface area of the rail.Firstly,this dissertation expounds the research background and significance of the identification of rail surface damage based on image detection,a large number of literatures are studied,the research status of image processing technology in rail surface damage detection and recognition at home and abroad is analyzed,and the existing research results are summarized,and the specific work content of this dissertation is clarified.And then,when the surface image of the rail is collected in the actual environment,it will be affected by environmental factors,equipment factors,light and other factors,so it is very important to preprocess the image.The principles of histogram equalization,single-scale Retinex and multiscale Retinex is researched,simulation experiments are carried out on these three image enhancement methods.The enhancement effects of the three methods are compared,and choose the histogram equalization method to enhance the rail image.The principles of Gaussian filter and bilateral filter are studied.In order to ensure that the most noise is removed and more effective edges are retained in the filtering process,an adaptive median guided filter is proposed to filter and reduce the noise of the rail image.The experimental simulation of three filters is carried out,and the effect of the three filters is tested by the objective evaluation index of filters PSNR.Secondly,it is necessary to extract the rail surface area,which will affect the subsequent detection and identification work.The filtered Canny edge detection algorithm is used to roughly extract the preprocessed rail surface image,so as to eliminate the sundries outside the rail surface area as much as possible.The Hough transform is applied to detect the straight line of the rail after the rough extraction,and the surface area of the rail is proposed.After the rail surface area is extracted,the damaged part in the image is detected.This dissertation studies the traditional Canny algorithm,analyzes its advantages and disadvantages,and improves the traditional Canny algorithm Gaussian filter noise reduction and manually setting double threshold.The Gaussian filter is replaced by adaptive median guided filter to achieve better filtering effect and retain more edge information.Use the improved Otsu method,the traditional Canny edge detection algorithm can adaptively select the threshold at the end,give up the method of manually setting double threshold.Finally,the method of deep learning is used to recognize and classify the detected damage images.Deep learning can minimize the complexity and computation of recognition.In this dissertation,the improved Faster R-CNN is improved by replacing the VGG-16 feature extraction network with Mobile Net V2 light network,so that the improved Fast R-CNN can achieve faster detection rate and identify and classify the rail surface damage images without changing the recognition effect.Compared with SSD and YOLOv3,the improved Faster RCNN has better recognition effect.The research of rail surface damage identification based on image detection proposed in this dissertation can quickly and accurately locate the rail area and its damage.Accurate damage identification can meet the accuracy and speed of track inspection image,and has a high application prospect. |
