Super Resolution Image Reconstruction Using Linear Regression Regularized Sparse Representation

Abstract

This thesis addresses the generation and reconstruction of the high resolution (HR) image by using the single low resolution (LR) image and the linear coalition of sparse coefficients from a suitably chosen over-complete dictionary.The study of compressive sensing shows that under vague conditions the sparse representation of a signal can be effectively recovered from the downsampled version of the original signal. By training both LR and HR image patches simultaneously by coupled dictionary learning, we are enforcing the similarity between the sparse representation(SR) of LR and HR image patch pairs with respective to their LR and HR dictionaries. Literature survey suggests that different extracted features are used to compute the coefficients to boost the prediction accuracy of the HR image patch reconstruction. A set of Gabor filters has been employed to extract useful features from the LR dictionary. As the super resolution is an ill posed problem, in this thesis we have considered it as an optimization problem for getting the sparsest representation of image patches using linear regression regularized with L1 norm, known as a LASSO in statistics.Our method is found to be outperforming the other previous state of art methods in both quantitative and qualitative analysis. The results reveal that proposed method shows promising results in reconstructing the image textures and edges.