Face Image Super-Resolution Based on Topology ICA and Sparse Representation

In this paper, a new learning-based face super-resolution (SR) algorithm is proposed considering the similarity of topology structure between low-resolution (LR) image and high-resolution (HR) image and the sparseness of cells’ response in visual cortex. Firstly, we obtain coupling dictionary which are corresponding to LR and HR image patch pairs by applying topology ICA. Then, the sparse coefficients of input LR image according to LR dictionary can be got based on sparse representation theory. Furthermore, primary HR face image is reconstructed using HR dictionary. Finally, finer HR face image can be got by back-projection step. Experiments demonstrate the proposed approach can get good SR results in subjective perception and objective evaluation.