Pre-/post-filtering and context modeling for DCT based block coding systems

Block coding based on the Discrete Cosine Transform (DCT) is very popular in image and video compression, especially for resource-constrained and speed-critical applications. The two well-known problems with current DCT based block coding algorithms are: (i) low coding efficiency; (ii) annoying blocking artifacts at low bit rates. This dissertation proposes to solve the problems by the combination of pre-/post-filtering and context modeling of DCT coefficients. Pre-filtering is attached to the encoder and post-filtering is attached to the decoder without affecting the existing DCT based infrastructure. Pre-filtering removes correlation between blocks, increasing coding efficiency; post-filtering eliminates blocking artifacts. Context modeling optimized for block transforms takes full advantage of correlation among DCT coefficients, resulting in highly efficient entropy coding. We first study the basic theory of pre-/post-filtering for both critically-sampled case and non-critically-sampled case with emphasis on high performance and low complexity implementations. Then we design a family of pre-/post-filters to provide desired tradeoffs between coding efficiency and robustness to transmission errors for image/video communication over unreliable channels. After that, we proceed to present 2 DCT based block image coding algorithms: adaptive runlength coding and context based entropy coding of block transform coefficients. Benefiting from pre-/post-filtering and advanced context modeling, the algorithms demonstrate state-of-the-art coding performance despite the strict complexity constraint. Finally, a lapped transform based video coding algorithm is described. The algorithm offers excellent coding performance and superior visual quality. We will show how pre-/post-filtering plays a critical role in vastly reducing the prohibitive complexity of the algorithm.