Bearings are very common mechanical components that are widely used in all kinds of machines. Currently, bearing surface defect detection is mainly completed by naked-eye observation. This manual detection mode has low reliability and is time consuming (especially for miniature bearings). In this paper, we present an automatic synthetic tiny defect (small defects that are difficult to find at vertical angles) detection system for bearing surfaces with self-developed software and hardware. Under two illumination modes in this system, thresholding segmentation, contour extraction, contour filtering, center location, region zoning and text recognition are successively implemented. Finally, four common defects (gap, stain, shrunken lid and scratch defects) are automatically detected. The experimental results show that this system has a high detection success rate and short time consumption. This method can provide technical support for real engineering applications involving bearing manufacturing and screening.