Adaptive block-based decomposition with pre/post-filtering

Efficient storage, archiving, processing, retrieval, and communication of natural visual information sources such as still images and video sequences continue to pose challenging problems to engineers and scientists. Generally speaking, modelling, analysis, estimation, and coding of non-stationary signals such as visual image/video information are very difficult problems, requiring highly adaptive systems. Block coding has been the popular method of choice in digital image/video compression, especially for resource-constrained and speed-critical applications. Unfortunately, the main disadvantage of previous block-based coding approaches is their ungraceful degradation at low bit rates. Coding efficiency suffers since inter-block correlation has not been taken full advantage of. Another disadvantage of current block-based coding systems is their lack of adaptivity to input signal characteristics.; To improve visual reconstruction quality, coding efficiency, but still retain most attractive features of block-based techniques, this dissertation studies the theory, structure, design, and implementation of adaptive block-based signal decomposition systems with pre-/post-filtering. This framework not only retains traditional advantages of block decomposition such as maintaining fast, efficient and VLSI-friendly implementations but it also adds on top a high level of signal adaptivity, allowing cascading block operators be chosen on the fly. Trade-off between complexity and coding performance can be obtained through different combinations of adaptivity in the pre-/post-filtering support, in the adaptive transform block size, in the degree of regularity as well as in the number of decomposition stages. The novel adaptive block-based signal decomposition system solves both the blocking artifact and ringing artifact problems in traditional block-transform-based image coders. Extensive image coding examples demonstrate that the new adaptive signal decomposition, when appropriately designed and utilized, offer significant improvements in both objective and subjective coding performances over all transforms reported in previous works.