Abstract: According to the detection demand of bearing steel ball surface defects with high precision and efficiency, a detection algorithm is proposed. To begin with, the algorithm of redrawing external rectangles of rounded contours was applied to segregate the area of targeted steel ball from the background area so as to pick up regions of interest; then, with image subtraction, a difference image could be got by subtraction between images of regions of interest and standard images; finally, binary the band area right in the middle of the ball for defect features. Experiment and evaluation on algorithm performances based on VC++ 6.0 development environment and Open CV computer vision library demonstrate that the algorithm is suitable for various defects, with the average processing time less than 30 ms for a single image, the recognition rate more than 97%, the mistake examining rate lower than 3% and the detection efficiency as high as 12 000 per hour for the diameter of 5.953 mm. Thus, it meets the demand of industrial assembly, enjoying a sound application prospect.