Multi-resolution mean-shift algorithm for vector quantization

The generation of stratified codebooks, providing a subset of vectors at different scale levels, has become necessary with the emergence of embedded coder/decoder for scalable image and video formats. We propose an approach based on mean-shift, invoking the multi-resolution framework to generate codebook vectors. Applied to the entire image, mean-shift is slow because it requires each sample to converge to a mode of the distribution. The procedure can be sped up with three simple assumptions: kernel truncation, code attraction and trajectory attraction. Here we propose to apply the mean-shift algorithm to the four image subbands generated by a DWT, namely the LL, LH, HL and HH subbands. It can be concluded from experimental results that the proposed MR-MS achieves similar PSNR to the LBG algorithm but outperforms it in terms of computation time.