Spatio-temporal super-resolution from compressed video employing global and local motion

In this paper, we propose a spatio-temporal approach for reconstructing high-resolution frames from compressed video. Utilizing motion information and inter-frame prediction residual, we form three estimates of the high-resolution frames: one based on spatial interpolation, one based on local motion, and one based on global motion. These three estimate are then linearly combined on a block-by-block basis using the Linear Minimum Mean Square Error (LMMSE) approach. Experimental results demonstrate the effectiveness of the proposed framework.