Vector set partitioning and successive refinement VQ for wavelet image and video compression

The ever increasing drive for efficient compaction of visual data, along with the need for attractive features such as successive transmissibility, bit stream scalability, and noise resilience, present an overwhelming urge to combine them into a common framework. The wavelet transform, in conjunction with zerotree quantization techniques has recently been shown to be very efficient for natural image compression. In this dissertation, a generic framework for embedded vector zerotree coding, called Vector Set Partitioning in Hierarchical trees (VSPIHT), is presented, where the individual coding units are vectors rather than scalars. Wavelet coefficients grouped as vectors, are classified by the set-partitioning methodology, and successively refined using trained or lattice vector quantization (VQ) techniques. Efficient compression schemes using the VSPIHT methodology are presented for monochrome and color images, as well as for color video. In all cases, an embedded bitstream is generated allowing functionalities such as SNR scalability, and progressive transmissibility. Since VSPIHT works especially well for images or parts of images with a great deal of detail, multimode VSPIHT techniques are introduced that can adapt the vectoring mode and other associated parameters based on image content. It is further shown that high-dimensional VSPIHT can effectively alter the balance of bits in the mixed VSPIHT bitstream so that much fewer critical bits carrying significance information are produced. This makes the use of unequal error protection for transmission of images over a noisy channel particularly convenient. Finally, two generic methodologies for successive refinement of vectors, applicable to all kinds of embedded coding are presented. The first leads to the development of Voronoi Lattice VQs for successive refinement of uniformly distributed vectors, and the second aims at developing new constrained storage multistage VQ structures allowing fine trade-offs between storage and efficiency.