Classification and segmentation of images using hidden Markov Gauss mixture models

As interest in content-based image retrieval grows with the volume of image data on the Internet, automatic classification of images is important for seeking and finding portions of images that “look like” a given image. Because an important component of many images is texture, a variety of algorithms have been developed for classifying textures. Segmentation of images is used in multimedia services for extracting explicit information about content so that human observers can interpret images clearly by highlighting specific regions of interest. Vector quantization is a lossy compression technique, based on principles of statistical clustering, that can be used for classification purposes. This thesis introduces several new methods for classification and segmentation of images based on a hidden Markov model and a Gauss mixture vector quantizer (GMVQ) combining ideas from vector quantization with Gauss mixture modeling.; For texture classification, we design a codebook or Gauss mixture for each texture using separate GMVQs. Most importantly, superblocks consisting of multiple Gauss quantization vectors are used to capture the macro features of the texture with low-complexity implementation. Our multi-codebook GMVQ classifier, applied to the Brodatz texture database, has proven to outperform other texture classifiers including classifiers based on Gauss Markov random fields, tree-structured wavelet transforms, and Gabor wavelet classifiers.; Conventional block-based segmentation algorithms determine the class of a block by examining only its feature vector and ignoring contextual information. In order to improve segmentation by context, we have devised an algorithm that models images by combining GMVQ and the Ising model to produce a two-dimensional non-causal hidden Markov Gauss mixture model (HMGMM). The stochastic EM algorithm is applied to optimize the MAP hidden states of the HMGMM of the image. This approach is used to identify man-made regions in aerial images and to segment textures of interest. Application to such images shows that HMGMM attains better segmentation than several popular methods, including the causal hidden Markov model (HMM), multi-resolution HMM, a classification and regression tree, and learning vector quantization in terms of Bayes risk and the spatial homogeneity of the segmented objects, with a computational load similar to that of causal HMM.