Abstract: Vector quantization (VQ) has emerged as a viable and practical bandwidth compression technique due to its promising performance and a simple decompression architecture that requires a single look-up table. The improvement in the performance of VQ is realized only when large dimension input vectors could be utilized. This is hampered by the exponential growth in the complexity and the storage requirements of VQ for large dimension vectors. This presentation summarizes the results of a study that considers the hardware completely of VQ based on both Linde-Buzo-Gray (LBG) classification and neutral networks (NNs). The result of the study shows that a single chip implementation of large dimension VQ at video rates using either LBG or NN approach is not feasible if a full search algorithm is utilized. Modified forms of LBG VQ, with suboptimal performance, can be implemented using a single chip at moderate vector dimensions and bit rates. The most efficient implementation of neural network vector quantization (NNVQ) is the one that uses a combination of an analog and a digital chip.!14
展开▼
