A step-by-step asymptotic method of image corner detection is proposed in high speed vision detection for quantities of micro two-dimensional workpieces. In this method, the inherent geometric relationship in the single pixel contour is taken as a priori information, based on which, the region where corners exist is separated from full contour. After that, we select suitable algorithm to detect corners and get their positions accurately along with false corners being eliminated in time. Then, we can lock region again ... This process continues circularly, until all the corners have been detected. Experimental results show that, for the same workpiece, whose images are getting in different position and attitude, the standard deviation of distance between points is within 0.5 pixels. Compared with Harris corner detection algorithm and CSS corner detection algorithm, the proposed method is scarcely affected by foreign substance in the field. Due to the accuracy, high speed and high repeatability, this method has obvious advantage in high speed vision detection for quantities of micro two-dimensional workpieces.