Rate distortion optimized adaptive multiscale vector quantization

We present a new rate-distortion optimized algorithm for adaptive vector quantization. It is based on approximate matching of recurrent patterns. In our approach, the input vector is segmented in variable-sized blocks. The blocks are encoded using a set of codebooks, one for each block size. The codebooks are updated while the data is encoded, with no need for any side information. Also, no prior training is required. We use dynamic programming. techniques to optimize the segmentation tree. It performs well for a wide class of sources, with very good results for highly nonstationary sources, like compound documents