Restoration and modeling for multimedia compression

Compression is of crucial importance for multimedia signals as they carry huge amount of information and the raw data representation would consume unaffordable storage space or transmission bandwidth. In this thesis, two important issues about multimedia compression are investigated. The first issue is signal restoration for compression, and the emphases of our research works are video denoising and image deblocking. For video denoising, we first propose denoising based on multihypothesis motion compensation to fully utilize video temporal correlation for noise suppression. In our proposed methods, multiple temporal predictions (multiple hypotheses) are combined by weighted averaging to achieve temporal denoising. The temporal denoising is adaptively combined with spatial denoising to further improve the denoising capability. To accelerate the video denoising process, we develop a noise-robust fast motion estimation strategy which can efficiently find accurate temporal predictions under noisy environment. In addition, we propose a novel scheme for the combination of denoising and encoding where the denoising is reduced to very simple operations on residue coefficients. For image deblocking, we revisit the classical POCS-based image deblocking approach from the perspective of convex optimization theory, and propose a new convex optimization-based restoration approach. The second issue is the modeling of compression systems. It is expected that by analytically modeling compression systems, the compression performance can be improved from a theoretical point of view. We have proposed two analytic models for video compression systems. The first one is a model of motion compensated residue. With the analysis of video compression process, the proposed residue model captures both the spatial and the temporal property of video signal. Based on this model, a strategy has been developed to reduce the temporal search range in multiple reference frame motion estimation and alleviate the computational burden of the encoder. The second model is an analytic quantization-distortion (Q-D) of video encoder. To the best of our knowledge, this is the first analytic model that takes both quantization distortion and its effect on temporal prediction efficiency into consideration. The proposed model can accurately predict the Q-D behavior of video encoders