HYBRID GLOBAL/LOCAL SEARCH STRATEGIES FOR VQ CODEBOOK GENERATION BASED ON OTSU AND LBG ALGORITHM

Abstract

Vector quantization algorithms have been extensively used for image compression, pattern recognition, image steganography, image retrieval, and anomaly intrusion detection. For large N(p) training vectors and N(c) clusters, vector quantization algorithms can hardly find the global optimal Classification without requiring a great deal of the squared Euclidean distance calculation. This paper proposes an efficient global division algorithm based on histogram threshold to improve computation time of the squared Euclidean distance from O(kN(p) log N(c)) to O(kN(p)N(c)). The experimental results and comparisons show that the global division algorithm can reduce computational complexity, find better codewords to improve the quality of the codebook and cooperate with the local search algorithm to tune it efficiently.