Linear prediction with low-pass filter and interpolator gives blurred image. But it can deal with aliasing and noise well.

Nonlinear prediction based on median filter and interpolator must be carefully designed due to the aliasing effect and noise.

We combine these two schemes using a criteria based on the difference of adjacent pixel. This criterion recognizes edges and areas with large fluctuation and applies an appropriate filter. This method works fairly well with the test images.

Future work

The proposed scheme has a lot of complexity and implementation issues. A study can be made to find a fast algorithm.

The filter parameters and threshold value for the parameter "p" are based largely on experimental result. There may be a systematic way to choose these parameters.