Image Super-resolution Reconstruction Based On Sparse Representation

Image super-resolution is to reconstruct a high resolution image form its single or multi-low resolution images. Interpolation-based image super resolution is a classical SR method, its advantage is that the algorithm is simple and easy to implement. Yet the blur effect and lost of high frequency information will occur when the magnification factor is very large. Up to now, more and more researchers focus on the sparse representation based image super resolution. It is very effective in reconstruct the image details, yet due to lack of the nonlocal spatial redundancy and structure redundancy, some high frequency lost and reconstructed details blurred. Based on the problems, our thesis contains three contributions in image super resolution based on dictionary learning and structure redundancy. As follows:(1) We proposed an image super resolution algorithm based on dictionary learning and structural similarity, to compensate the problems occurred in the classical method. We utilize the SSIM in dictionary learning and sparse representation, which makes the low resolution sparse coefficient more approximate with the high resolution sparse coefficient, also dictionary becomes more reasonable in image super resolution. Experimental results show that the proposed method can achieve a state-of-the art performance with more clear and plentiful details.(2) The current dictionary-based method assume that the low and high resolution image have the same sparse coefficient over the corresponding dictionary. It ignores the structural differences in image patches. In the thesis, we first employ the structure clustering to train a dictionary based on different datasets, then reconstruct the high resolution image patches and obtain the finial HR image through weighted summation. Experimental results show that the proposed method has advantages not only in objective value but also in subjective value.(3) Nonlocal spatial redundancy is one of the classical redundancies. In our thesis, we design a new regular function by introducing the nonlocal redundancy and prior and proposed a new image super resolution method based on nonlocal total variation. In this method, we transform the image super resolution problem into an image restoration problem. We design a new regular function based on nonlocal total variation, to restore the image. Experimental results show that the proposed method not only can guarantee the consistency of the smooth region in reconstructed high resolution image, but also can retain the image details and the integrity of the edge profile.