Research On Detection Algorithm For Rail Surface Defects

With the rapid growth of modern railway operation mileage, train speed and loading capacity, it becomes more and more important to efficiently detect the rail surface defects.The traditional detection methods of rail surface defects have been unable to adapt the trend of high speed and accurate automatic detection, So there is an urgent need for a new detection technology,which is more efficient. This paper designs a set of detection and recognition algorithm of rail surface defects based on the digital image processing, realized automatic detection and recognition of rail surface defects.Firstly, the rail images, which collected from the scene are affected by non rail area, noise, uneven illumination and different surface reflectance of rail, so must preprocess the rail images. The paper proposes an adaptive projection algorithm by analyzing the gray level difference of rail region and non rail region, the proposed algorithm can cut the rail region accurately. By analyzing some characteristics of rail surface image, contrast measurement principle and Weber contrast, the paper proposes a local Weber contrast cutting method, the experimental results show that the rail image enhancement effects of the proposed algorithm is better than some traditional enhancement algorithm. Through contrasting and analyzing the denoising effects of threshold method smooth linear filtering, median filtering and multistage median filtering, the multilevel median filter is selected to denoise the rail image.Secondly, Through analyzing the gray probability distribution curve of preprocessed rail image, the target part curve and the background part curve of the between class variance,the paper proposes a proportion emphasized maximum between class variance method for rail image segmentation, this method is a improved maximum between class variance method,the experimental results show that the rail image segmentation effects of the proposed algorithm is better than some traditional segmentation algorithm. The segmented rail images are processed by morphological, and the most small holes and isolated points of rail image are eliminated.Then, By combining the run length encoding method and the recursion marking method,the paper proposes recursion marking method based on run length encoding to mark the rail surface defects, the proposed algorithm takes less time than some commonly used marking algorithm. The marked defects are extracted and numbered, and the precision rate and recall rate of two evaluation criteria are used for defect detection performance evaluation.The features of rail surface defects are extracted, and the length-width ratio and the compactness are selected to the input of following pattern recognition.Finally, the paper designs a sizeable LVQ neural network for rail defect classification,and the neural network is trained, the experimental results show that the crack defects and the scar defect are identified accurately by using the trained LVQ neural network.