A comparison of Bayes risk weighted vector quantization with posterior estimation with other VQ-based classifiers

We compare the compression and classification performance of various vector-quantizer based classifiers on real images. These quantizers include a Bayes risk weighted vector quantizer, Kohonen´s “learning vector quantizer” (LVQ), and an independent design of quantizer and classifier. Both full search and tree-structured codes are considered. The quantizers are applied to aerial photographs and medical images where the goal is to both compress the images and classify particular features within the images. We demonstrate that for the examples considered, Bayes risk weighted vector quantization with posterior estimation obtains similar or superior classification and compression performance to that obtained with the other systems