Abstract: Aiming at the surface defect problem of sky railway track beam, the method of image analysis is adopted to inspect. Firstly, for the image information of steel plate surface which is collected by CCD camera, the normal samples are obtained by means of a rough inspection method based on mean and variance. The mean and variance of surface images are calculated by sliding windows of a certain size, and the samples are divided into suspected defect samples and normal samples after the threshold is determined. Then, on the basis of rough inspection, Bayes fine inspection method based on integral graph is adopted to improve the accuracy. Using the compressed sensing algorithm, the characteristic samples of defects are obtained, and the normal samples which are wrongly divided into defect samples in the rough inspection process can be removed. The experimental results show that the detection rate and accuracy of this method for surface defects of steel plate are greater than 98% and 95% respectively.