Image Recovery from Compressed Video Using Multichannel Regularization

In this chapter we propose a multichannel recovery approach to ameliorate coding artifacts in compressed video. The main shortcomings of previously proposed recovery algorithms for this problem was that only spatial smoothness was explicitly enforced. In this chapter we attempt to ameliorate the above problem. According to the proposed approach, prior knowledge is enforced by introducing regularization operators which complement the transmitted data. The term multichannel implies that both the spatial (within-channel) and temporal (across-channel) properties of the image sequences are used in the recovery process. More specifically, regularization operators are defined that in addition to spatial smoothness explicitly enforce smoothness along the motion trajectories. Since the compressed images due to quantization belong to known convex sets, iterative gradient projection algorithms are proposed to minimize the regularized functional and simultaneously guarantee membership to these sets. Numerical experiments are shown using H.261 and H.263 coded sequences. These experiments demonstrate that introduction of temporal regularization offers a significant improvement both visually and from a peak signal-to-noise ratio point of view.