Mean shift, mode seeking, and clustering-英文文献.pdfVIP

790 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 17, NO. 8, AUGUST 1995 Mean Shift, Mode Seeking, and Clustering Yizong Cheng Abstract-Mean shift, a simple iterative procedure that shifts A relation among kernels called “shadow” will be defined in each data point to the average of data points in its neighborhood, Section 111. It will be proved that mean shift on any kernel is is generalized and analyzed in this paper. This generalization makes some k-means like clustering algorithms its special cases. It equivalent to gradient ascent on the density estimated with a is shown that mean shift is a mode-seeking process on a surface shadow of its. Convergence and its rate is the subject of Sec- constructed with a “shadow” kernel. For Gaussian kernels, mean tion IV. Section V shows some peculiar behaviors of mean shift is a gradient mapping. Convergenceis studied for mean shift shift in cluster analysis, with application in Hough transform. iterations. Cluster analysis is treated as a deterministic problem Section VI shows how, with a twist in weight assignment, the of finding a fixed point of mean shift that characterizes the data. deterministic mean shift is transformed into a probabilistic Applications in clustering and Hough transform are demon- strated. Mean shift is also considered as an evolutionary strategy evolutionary strategy, and how it becomes a global optimiza- that performs multistart global optimization. tion algorithm. Index Terms-Mean shift, gradient descent, global optimiza-