Side information extrapolation with temporal and spatial consistency

Xianming Liu,Deming Zhai,Debin Zhao,Ruiqin Xiong,Siwei Ma,Wen Gao

DOI: https://doi.org/10.1109/ISCAS.2011.5938242

2011-01-01

Abstract:In this paper, we present an efficient side information extrapolation scheme with temporal and spatial consistency for low delay Wyner-Ziv video coding. Our method is based on the regularized local linear regression (RLLR) model, in which each pixel in SI is approximated as a linear weighted combination of samples within a local temporal neighborhood. The optimal model parameters are estimated by projecting the transformation function onto the temporal training samples to exploit motion-related dependency. During this procedure, moving weights are incorporated into the objective function to express the relative importance of training samples in estimating parameters of the model. Furthermore, spatial correlation is explored by imposing an additional local smoothness penalty, which does good to estimate the occluded regions and complex motion regions. The learned function is smooth and locally linear, and can be obtained with a closed-form solution by solving a convex optimization problem. Experimental results demonstrate that the RLLR method achieves very competitive SI extrapolation performance compared with the state-of-the-art methods.