Singular vector decomposition based adaptive transform for motion compensation residuals

Abstract

Video coding standards commonly use discrete cosine transform (DCT) to transform the motion compensation (M-C) residuals. However, the MC residuals have much weaker correlation than image pixels, and DCT is not the optimized transform for them. In this paper, we propose an adaptive transform structure for MC residuals. Unlike traditional approaches which use a predefined transform core, we apply singular value decomposition (SVD) on the prediction block and use the eigenvector matrices as the transform core. Experiments show that this adaptive transform is more efficient compared with the traditional approach. An average 2.0% bit rate reduction is achieved when implemented on H.265/HEVC.