The point spread function (PSF) maps fi,j to g or vice versa. To do this, blurring, decimation and motion need to be modeled at high resolution.
In case of superresolution with scale factor 2 a simple 2x2 box filter is used for blurring and downsampling. For motion field more complex interpolation is used. Nearest neighbor interpolation results in the 3x3 PSF, in other word each low-resolution image pixel is mapped to 3x3 area in high-resolution image.
Another useful mode of this algorithm is x1 upsampling, it may be used for quick de-noising and de-blocking of chrominance planes. In this case motion field is not interpolated, and blurring is not performed, thus it can be done with the 2x2 PSF.
Figure “Mapping Low-resolution frame with the PSF” shows the above described types of PSF.
