Fast VQ algorithm using scalar prequantization

Although the average distortion of full search vector quantization is minimized, its utilization is severely limited by the computational complexity of the encoding process. The authors propose a new method such that, to reduce the number of codewords for which the distortion must be calculated, each sample of the input vector is quantized before the vector quantization. The number of codewords to which any vector existing in some hypercube will be quantized is very much smaller than the number of the original codewords. The method is applicable to various distortion measures based on the norm. The simulation result shows that the proposed method is 62 times faster than the conventional full search algorithm for the IID Gaussian source when the dimensionality is 4 and the rate is 10[bits/vector]