Variable rate vector quantization of images

Vector quantization is a lossy compression technique that has become popular in the last decade. Its performance for image compression applications can be significantly improved by using variable rate codes, which are able to code active regions of an image, such as the edges, at a higher resolution. At the same time, variable rate coding saves bits by coding the less active regions, such as the background or regions of constant intensity, at a lower resolution. In this thesis, we present several applications of a recently developed pruning technique for variable rate tree-structured vector quantizer design to images. Pruned tree-structured vector quantization (PTSVQ) is particularly suitable for progressive transmission of images, in which an increasingly higher quality image can be reconstructed by the decoder. A variation of PTSVQ incorporates a predictive preprocessor that improves the performance of the coders by close to 3 dB. Next, a technique is introduced for directly designing a variable rate tree-structured vector quantizer. Here the tree is grown one node at a time rather than the typical one layer at a time. This is less constrained than growing a balanced tree and the resulting unbalanced tree outperforms a balanced tree of the same average rate. When the tree is pruned, additional improvement is measured in the signal to noise ratio at high rates over standard PTSVQ. The tree growing algorithm can be interpreted as a constrained inverse operation of the pruning algorithm. Finally, the pruning algorithm is applied to a bit allocation problem in which differing numbers of bits are allocated in a classified vector quantizer application. The algorithm is conceptually very simple and has very low complexity under convexity constraints on the class quantizer functions.