“Universal” transform image coding based on joint adaptation of filter banks, tree structures and quantizers

Summary form only given. Transform coding has become the de facto standard for image and video compression. The design of adaptive signal transforms for image coding usually follows one of the two approaches: adaptive tree/quantizer design with fixed subband filter banks and adaptive subband filter bank design with fixed quantizers and tree topology. The main objective of our work is to integrate these two paradigms in an image coder in which subband filter banks, tree structures and quantizers are all adapted. We design a codebook for the filters, tree and quantizers. The codebook design algorithm uses a training set made of images that are assumed to be representative of the broad class of images of interest. We first design the filters and then the quantizers. In the filter design phase, we visit nodes in a top-down fashion and design a filter codebook for each tree node. The optimal filter codebook for each node is designed so as to minimize the theoretical coding gain-based rate. The design of the quantizers and the weights for the splitting decisions is done jointly using a greedy iterative algorithm based on the single tree algorithm of Ramchandran et al. (1993). The actual coding algorithm finds, based on the codebook design, the optimized filter banks, tree structure, and quantizer choices for each node of the tree. In our experimental setup we used a training set of 20 images representative of four image classes