Vector quantization of video with two codebooks

In this paper, we present a VQ-based two-codebook design which uses separate codebooks for predicted residuals and full pixel values. We show that our approach captures abrupt scene changes while exploiting inter-frame dependencies. We use a simple universal code consisting of two codebooks, an intra-codebook containing codewords that are used as reproduction of image blocks together with a residual-codebook. Codebooks are selected to minimize distortion for the block being coded. If there is relatively little motion in a frame, most blocks use the residual-codebook. On the other hand, if a frame is very different from the previous one, most blocks will be coded using the intra-codebook. When compared to other VQ schemes mentioned above, the resulting quantizer not only follows scene changes closely with satisfactory fidelity but also is robust against mismatch between the training and test sequence. We compare the PSNR of three coding schemes, intra-coding-only, inter-coding only, and the proposed method