Performing Vector Quantization Using Reduced Data Representation

We propose an improvement to VQ by applying data and dimensionality reduction techniques on the original dataset and using different resolutions generated by the reduction techniques as input to the GLA with splitting at different stages of codebook generation. The different resolutions are pre-computed (a one time cost), and are used instead of the original dataset in each iteration of the GLA. We propose three variants of the reduced data representation VQ (RDR- VQ) based on the three reduced data representation techniques we use: 1. Reduced precision VQ (RP-VQ), where we use a simple rounding off of precision of the original dataset; 2. Wavelet RDR-VQ, where we use a discrete wavelet transform of the original dataset 3. PCA RDR-VQ, where we use a linear transform of the original dataset using PCA The difference in our work with existing algorithms that use VQ with wavelets is while these algorithms use multiple quantizers on the wavelet subbands, our method uses the different subbands as input to different iterations of a single quantizer. Different resolutions of the dataset are used as input for the different Lloyd iterations. An important consideration is choosing an appropriate resolution level for each of these iterations. We call the different resolution levels used by the algorithm to compute the final codebook a resolution sequence.