Context-dependent tree-structured image classification using the QDA distortion measure and the hidden Markov model

Vector quantization based on the Gauss mixture model (GMM) and the quadratic discriminant analysis (QDA) distortion measure has been shown to perform well in statistical image classification problems. Previous work in this area has concentrated on designing a separate GMM-based vector quantizer using the QDA distortion measure for each class using full search. We design a single vector quantizer for all classes using a tree-structured algorithm based on the (generalized) BFOS algorithm. This reduces the search complexity, while it increases the correct classification rate. Further, the pruning stage of our algorithm takes into account the dependencies between the image blocks assuming a hidden Markov model (HMM). During the test stage, our algorithm aims to iteratively maximize the joint probability of occurrence of all image blocks based on the HMM. Our simulation results indicate that our algorithm performs better (both in terms of computational complexity and classification rate) when compared to the previously published algorithms based on the GMM.