Although randomized Hough transform and randomized circle detection are two fast algorithms for circle detection in image, there exists deficiency of speed and accuracy while practicing them. In this paper, we generally summarized two problems in the above algorithms. First, sampling distribution, accumulation distribution and number of consecutive sampling were concluded as the problem of sampling constraints. Second, the bias between parameters that are only determined by the three agent pixels and true ones were regarded as the problem of refinement. Based on the analysis of the two problems, we operated improved randomized circle detection algorithm and randomized Hough transform as a fast recognition method and a refinement scheme, respectively. Thus, a new circle detection method which is in the framework of recognition-refinement was proposed. Results from applying our method to images with noise and inferior boundaries demonstrate that this framework manages to balance well the tradeoff between speed and accuracy of detection, and show the effectiveness of the proposed algorithm.