In order to real-time on-line detection of surface defects of high-speed rail hub, a superpixel adaptive detection method based on visual saliency attention mechanism is proposed. Firstly, homomorphic filter was used to preprocess defect image to remove the uneven distribution of image brightness caused by ambient light pollution noise, and a spectral residual visual attention model of defect image on high-speed rail hub surface was constructed. Then, superpixel segmentation algorithm was used to segment the defect saliency image by adaptive threshold. The two-dimensional space location of surface defects of high-speed rail hub was marked, and the boundary detection and shape estimation of surface defects of on high-speed rail hub were realized. The experimental verification of the method was carried out on the experimental platform for surface defect detection of high-speed railway hub. The results show that the method can effectively restrain over-segmentation in image segmentation, extract the boundary information of defects accurately and have good robustness.