In view of the influence of different illumination angles, color and intensity of light sources (LS) on the accuracy of crack detection, this text took rail surface detection (RSD for short here) as an example, constructed an optical model to describe the interaction process between light and surface, and realized the improvement of mechanical surface detection accuracy with the help of machine vision technology. Under the circumstances of different LS irradiation angle, LS color and LS intensity, the rail surface image was acquired by using the linear array CCD (charge coupled device) camera. In this text, Karpathy was used to divide the data set, preprocess the image, reduce image noise data and enhance image quality. In this text, gray co-occurrence matrix, Canny edge detection and color moment were used to extract rail surface features, and Convolutional Neural Network (CNN) model was used for feature detection. The results show that the difference between the gray-scale value (GSV) at the crack and the GSV of the normal rail is the largest at a light angle of 15°. Under the environment of white light irradiation and LS intensity of 5000 lumens, the detection accuracy of the CNN model was 98.3% when the illumination angle of the LS was 15°. The optical model of RSD using CNN can effectively improve the detection accuracy.
