A kernel-based statistical analysis of the residual error in video coding

Video compression techniques exploit the statistical redundancy present in video signals to efficiently reduce the amount of information sent to the decoder. We contribute with a kernel-based analysis of the residual error blocks. In particular, we borrow dimension reduction techniques from machine learning, namely Principal Component Analysis (PCA) and nonlinear Kernel Principal Component Analysis (KPCA), to assess the spatial structure of block residuals. Interestingly, a nonlinear structure is observed that correlates to the rate-distortion costs of the blocks. Simulations by using a test set of videos with cropped Ultra High Definition (UHD) resolution show interesting results.