Computationally Efficient Image Super Resolution from Totally Aliased Low Resolution Images

This paper considers the problem of super-resolution (SR) image reconstruction from a set of totally aliased low resolution (LR) images with different unknown sub-pixel offsets. By assuming the translational motion model, a linear compact representation between the LR image spectrums and SR image spectrum, based on multi-coset sampling is provided. Based on this model, we formulate the joint estimation of the unknown shifts and SR image spectrum as a dictionary learning problem and alternating minimization approach is employed to solve this joint estimation. Two different approaches for obtaining the SR image; one based on estimated shifts and another based on estimate SR spectrum are described. The significant advantage of the proposed approach is the smaller matrix sizes to be handled during the computation; typically on the order of number of images and enhancement factors, and is completely independent on the actual dimensions of LR and SR images, hence requiring significantly lesser resources than the current state of the art approaches. Brief simulation results are also provided to demonstrate the efficacy of this approach.