Near optimal non-uniform interpolation for image super- resolution from multiple images

Non-uniform interpolation is a common procedure in image processing. A linear interpolation filter is generally a weighted combination of the inputs. Optimal filter coefficients (in a least squares sense) can be derived if the interpolated image is known beforehand. The weights of a general linear interpolation filter are independent of content and only depend on the relative positions of the available samples. The optimal coefficients are shown to be relatively independent of the content experimentally in absence and presence of noise, allowing non-uniform images to be interpolated using coefficients that have been optimised on a synthetic image. This results in a linear interpolation with computational complexity of the same order as nearest neighbour or bilinear, but with a near optimal performance.