Multi-resolution Mean-Shift Algorithm for Vector Quantization

Abstract

This paper presents a new multi-resolution mean-shift algorithm for vector quantization of high-resolution images and generation of a stratified codebook. The algorithm employs the discrete wavelet transform (DWT) to circumvent the problem of bandwidth selection with the mean-shift algorithm. Here, the mean-shift algorithm is applied to a reduced set of samples in the color space. The detection of salient edges is performed on the DWT subbands for each level of decomposition to identify the pixels escaping the basin of attraction. The quantized image is reconstructed by upward interpolation of the salient pixels in the feature space. We also propose a Modified-Weighted mean-shift algorithm to speed up the image reconstruction stage. Experiments show that the proposed multi-resolution mean-shift provides significant speed-up compared to the Linde-Buzo-Gray (LBG) algorithm.