Recursive Wiener filter for motion parameter estimation in three-parameter motion model

摘要

Motion compensation is used to reduce the displaced frame difference (DFD) during video coding. To increase the accuracy of the point correspondence during the compensation, a three-parameter motion model is considered. The matching error can be significantly reduced as compared with that of the two-parameter block matching. To derive the parameters, a partial full search method is used. The full search is used when the zoom value is set at one. Otherwise, the gradient-based algorithm is used. Since the DFD is a nonlinear function of the image gradients and the motion parameters, a linearized model is considered. To eliminate the linearization error, Wiener filtering is used to smoothen the DFD to improve the convergence condition of the iterative gradient search. To make the gradient-based search more robust to the gradient variation, several gradient estimation methods are also compared.