Low Complexity H.264/AVC Encoding And Transcoding

The latest H.264/AVC standard outperforms previous video coding standard in variousaspects such as bit rates, video quality due to the introduction of many new conceptsand tools. The appearance of small block encoding ensures the coding quality be keptat comparably a high level, but on the other hand, this also results in huge computationtime in finding the best mode for the encoded macroblock (MB). So many researchworks have been focused on reducing the computation complexity. The main contentof the thesis also contributes to implementing low complexity H.264/AVC encoding ortranscoding. Through experimental analysis, it is known that motion estimation is themost computation intensive part in the whole encoder. Thus, proposals of fast modedecision and fast transcoding can all efficiently reduce the complexity of the system.The content of the thesis has been mainly discussing the following issues and alsothe related research progress done:Firstly, the background knowledge of the research is summarized, including modedecision and video transcoding. With the discussion of the current progress in relatedtopics, the main issue as well as difficulties of the following two topics are outlined: fastmode decision for DIRECT mode in B frame and fast heterogeneous video transcoding.For fast spatial DIRECT mode decision in B frame, the limitations of the currentwork is discussed. And the method proposed in this thesis is explained in details. Thismethod is based on the analysis of relation of temporal coding information. Currently,major works have been done on intra mode, inter mode and SKIP mode to speedup the encoding process, but the work on DIRECT mode is rather limited. In fact,DIRECT mode has been widely used in coding of everyday life sequences. So, if DIRECTmode can be determined at an early stage, then the performance improvement by using the proposed method would be expected. In this thesis, a fast mode decisionmethod dedicated to spatial DIRECT mode is proposed. Through experimental analysis,the relation of the current block and the co-located block in mode decision andrate-distortion cost are analyzed. The analysis result of this method has been utilized togenerate a set of conditions for early DIRECT mode decision. To enable better codingperformance with various video sequences, the thesis further proposes a dynamic parameterupdate model, which ensures the overall algorithm to be robust. In addition, thealgorithm achieves satisfying result even in small QP cases, which also distinguishesit from other proposals.For fast transcoding from AVS to H.264/AVC with down-sampling, the widelyaccepted models and structures are discussed in the thesis. Then the proposed multistagemotion and mode mapping strategy applied in the proposed method are introduced.Previously, many works have been focused on transcoding under the same resolutionwhile the proposal in this thesis deals with transcoding with down-sampling.The proposal also has a very clear application scenario: to provide robust and efficientsolutions for transcoding standard AVS streams to CIF resolution H.264/AVC streamswhich can be played on mobile devices. Besides, the proposal can be widely appliedto meet other application needs as long as the down-sampling rate is 2:1. Due to thefact that the stream will be down-sampled, how to efficiently utilize the motion andmode information extracted from the decoder becomes a critical issue for the evaluationof the performance. Compared with the existing works, the proposed method inthis thesis is innovative in the following aspects: 1) provide solutions for motion andmode mapping under down-sampling process. 2) for intra predicted blocks in originalstreams which has no motion information available, a motion information derivationstrategy is proposed for these blocks. 3) multiple merging patterns are proposed fordifferent level motion information merging to assure the accuracy of re-utilizing theinformation.Finally, the pros and cons of the proposed methods are discussed in details andfuture directions to improve the performance of these methods are discussed.