The basic idea is to predict a pixel value in a non-distorted frame, and then to compare this value to that in the input (corrupted) image. Usually these two values are different and a decision has to be made upon the pixel value at the filter output. The output can be considered as a sum of the prediction and the prediction error processed in some way. For example, large values of prediction errors can be set to zero because they are classified as caused by impulse noise samples. Soft decisions on classification of prediction errors lead to good results for test images. Three-dimensional median filters of various types are used here as predictors. First of all, vector median filters are used. A median is calculated over windows in consecutive frames, e.g. from the last past frame, from the current frame and the next future frame. Windows in past and future frames are motion-compensated. Prediction error is processed either by a memoryless nonlinear element or even by a 2-D filter. Experimental results with videophone test sequences prove that the filters described are quite efficient.
"Motion-compensated nonlinear filters for video restoration", Proc. SPIE 3646, Nonlinear Image Processing X, (5 March 1999); doi: 10.1117/12.341088; https://doi.org/10.1117/12.341088