Fast mean-shift algorithm for image segmentation

Mean-shift is a nonparametric clustering method based on kernel density estimation for feature space analysis. In many computer vision tasks, such as image segmentation and tracking, it has acquired promising performance. However, the slow convergent speed limits the practical applicability. In this paper, an improved fast mean shift algorithm is proposed by enlarging the shift step length and combining the Aitken accelerating iterative algorithm. Depending on the physical meaning of the mean-shift, the accelerated shift procedure with enlarged mean-shift vector was analyzed in detail. The convergence condition of discrete data has also been deduced under the new iterative formula. We use the concept of scaling factor to represent the expansion degree of the enlarged mean-shift vector and the performance of the algorithm with different fixed scaling factors was also investigated. In order to optimize the oscillation caused by the enlarged mean-shift vector, we introduced the technique of Aitken accelerating iterative algorithm to speed up the convergence. Experimental results on various challenging images have demonstrated the superior performance of our proposed methods with several famous and state-of-the-art algorithms. 1553-9105/Copyright © 2014 Binary Information Press