Classification vector quantization of image data using competitive learning

We present an implementation of vector quantization for use in the coding of image data. The implementation is based on the frequency sensitive competitive learning (FSCL) variant of the competitive learning neural network algorithm. Previous work has shown that this neural network provides a large computational advantage over existing methods such as the Linde, Buzo, and Gray (LBG) algorithm for small codebook sizes. However, for large codebooks which are necessary for good performance, the neural network implementation requires an excessive amount of training. We work to correct this deficiency by applying the classification vector quantization technique to the neural network implementation of the vector quantizer. In this manner we can separate the problem into the generation of codebooks for each of the two categories into which the data is divided. By appropriately choosing the categories we can vastly simplify the training process and thus substantially reduce the computational costs while improving the subjective performance