Image compression and classification using vector quantization
K.L. Oehler · 1993
Vector quantization is a lossy compression technique based on principles from statistical clustering. In this thesis, we present several vector quantization methods for compression and combined compression and classification of images. After a review of the basic ideas of image compression and vector quantization, we describe an image compression technique involving product vector quantization that uses three tree-structured codebooks to encode the mean, gain and shape features of vectors. Joint pruning of the codebooks is shown to provide optimal bit allocation among the features. The method provides low encoding complexity and reduced memory requirements while producing good coding performance. Next, the application of vector quantization to image classification is discussed. Vector quantization can perform block-based classification and compression simultaneously by associating a class label with each codeword. We present two methods of modifying traditional vector quantizer design algorithms to improve the classification ability of the quantizer. One approach modifies the tree-growing algorithm used for constructing tree-structured codebooks by incorporating classification information into the splitting criterion. The second approach explicitly incorporates a Bayes risk component into the distortion measure used for quantizer design, permitting a flexible trade-off between compression and classification priorities. These approaches are used to analyze simulated data, identify tumors in computerized tomography (CT) lung images, and identify man-made regions in aerial images.