Image classification using tree-structured discriminant vector quantization

According to the principle of minimum description length, the best classifier is the one that minimizes the sum of the complexity of the model and the description length of the training data. As the complexity of any realizable model is finite, the emphasis should be on minimizing the description length of the training data. Discriminant vector quantization (DVQ) tries to achieve precisely this goal by minimizing the description length of the training data through a two-stage vector quantization. We propose a tree-structured version of DVQ based on the generalized BFOS algorithm. This reduces the search complexity, while increasing the correct classification rate. Further, we propose a split criterion based on the mismatch due to quantizing a source with a quantizer optimized for a probability distribution function different from that of the source.