Multi-reference neighborhood search for vector quantization by neural network prediction and self-organized feature map

Reference neighbor search (RNS) is a new technique for fast searching of vector quantization (VQ). However, the optimal performance is not guaranteed and the performance is greatly affected by the selection of reference point. In this research, the authors employed the Kohonen self organized feature map to generate a codebook of high degree of neighborhood and multi-layer perceptron (MLP) neural network to adaptively predict the reference. The predicted reference is closer to the input, thus the search distance will be reduced together with multiple queues and a look up table, the number of searches is significantly reduced while maintaining optimal performance