Image coding based on multiple projections and multistage vector quantization

The vast majority of practical image coding systems used today are based on the transform coding paradigm, where image blocks are projected into a series of basis functions, and the expansion coefficients are subsequently quantized. In this paper we introduce a novel constrained vector quantizer (VQ), which we call seg-VQ. As an extension of the transform coding framework, in our approach codevectors are constrained to be located on a series of line segments in the multidimensional space. These segments are designed sequentially based on a training set. The advantages of seg-VQ are twofold: first, the encoding complexity is proportional to the number of segments rather than to the number of codevectors, and second, it can efficiently exploit the directional preferences (correlations) in sources such as images. For image sources, at low dimensions (e.g., 4 by 4 blocks), at the same encoding complexity with TSVQ, seg-VQ outperforms TSVQ by 0.5 dB at 0.4375 bpp achieving a performance close to unconstrained VQ obtained by pairwise nearest neighbor (PNN) initialized GLA. At higher dimensions (e.g., 8 by 8 blocks) we use multistage seg-VQ where the input block (as in transform coding) is projected into a series of segments in order to be quantized.