Vector quantization with compressed codebooks

Abstract

Abstract Codebook storage is the key limitation in reducing distortion of vector quantization (VQ). We propose an effective lossless compression scheme with reduced computational requirements. Variable precision representation (VPR) for each vector y stores exp( y ), the number of leading bits which are zero in all elements, and avoids storing the zero bits. We show that storing the difference of centroid codevectors effectively removes the redundancy in the codebook. This difference is non-stationary. As the mean square error of the VQ encoder decreases, the added codevector differences become smaller and more amenable to compression. In mean-residual VQ, VPR can save more than 50% in storage, and achieve more than 75% bit reduction of entropy coding. Moreover, VPR reduces the cost of VQ. The VPR inner product module has 10% less area × time cost than the conventional inner product. The designs are verified and their speed estimated for 1.0-μ CMOS gate array technology. Using tree-structured VQ, they can encode in real-time television quality video (720 × 576 pixels × 30 frames/s).