Modified K-means algorithm for vector quantizer design

被引:80
作者
Lee, D
Baek, S
Sung, K
机构
[1] Dept. of Electronics Engineering, Seoul National University
关键词
D O I
10.1109/97.551685
中图分类号
TM [电工技术]; TN [电子技术、通信技术];
学科分类号
0808 ; 0809 ;
摘要
The K-means algorithm is widely used in vector quantizer (VQ) design and clustering analysis, In VQ context, this algorithm iteratively updates an initial codebook and converges to a locally optimal codebook in certain conditions. It iteratively satisfies each of the two necessary conditions for an optimal quantizer; the nearest neighbor condition for the partition and centroid condition for the codevectors. In this letter, we propose a new algorithm for both vector quantizer design and clustering analysis as an alternative to the conventional K--means algorithm, The algorithm is almost the same as the K-means algorithm except for a modification at codebook updating step, It does not satisfy the centroid condition iteratively, but asymptotically satisfies it as the number of iteration increases, Experimental results show that the algorithm converges to a better locally optimal codebook with an accelerated convergence speed.
引用
收藏
页码:2 / 4
页数:3
相关论文
共 7 条
[1]  
ANDERBERG M, 1973, CLUSTER ANAL APPLICA
[2]  
BOW ST, 1992, PATTERN RECOGNITION
[3]  
Gersho A., 1992, VECTOR QUANTIZATION
[4]   A New Initialization Technique for Generalized Lloyd Iteration [J].
Katsavounidis, Ioannis ;
Kuo, C. -C. Jay ;
Zhang, Zhen .
IEEE SIGNAL PROCESSING LETTERS, 1994, 1 (10) :144-146
[5]   ALGORITHM FOR VECTOR QUANTIZER DESIGN [J].
LINDE, Y ;
BUZO, A ;
GRAY, RM .
IEEE TRANSACTIONS ON COMMUNICATIONS, 1980, 28 (01) :84-95
[6]   K-MEANS-TYPE ALGORITHMS - A GENERALIZED CONVERGENCE THEOREM AND CHARACTERIZATION OF LOCAL OPTIMALITY [J].
SELIM, SZ ;
ISMAIL, MA .
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 1984, 6 (01) :81-87
[7]   GLOBALLY OPTIMAL VECTOR QUANTIZER DESIGN BY STOCHASTIC RELAXATION [J].
ZEGER, K ;
VAISEY, J ;
GERSHO, A .
IEEE TRANSACTIONS ON SIGNAL PROCESSING, 1992, 40 (02) :310-322