An improved randomized circle detection algorithm based on least square approach was proposed. Because of the incorrect selection of threshold, a circle is divided into some circles. This algorithm can solve the problem perfectly. First, a randomized algorithm for detecting circles randomly picked four edge pixels each time and calculated circle parameters. The set of pixels belonging to the circle were obtained. Second, these pixels can calculate new circle parameters, and new pixels belonging to these parameters were obtained. Third, until no new pixel was added, it returned these pixels into the set. At last, we checked whether or not the set of these pixels was real circle. Compared with the original algorithm, the proposed algorithm enhances the stability and accuracy of circles detection.