Single image superresolution by multiple geometrical regressors

In this paper, to improve the quality and enhance the edge sharpness of the reconstructed image, a novel example-based single image superresolution approach is proposed, where the mappings between a low-resolution (LR) image and the corresponding high-resolution (HR) image are established based on multiple regressors. At first, multiple pairs of LR and HR geometrical dictionaries are learned from the pre-classified example patches, respectively. Then, for each atom in the geometrical dictionary, the local regressor is built up by accumulating a certain number of the most similar patches in both LR and HR spaces. In the reconstruction process, for each input LR patch, the most similar atom in each dictionary is searched and the corresponding regressor is chosen. Thus, these multiple geometrical regressors are used to get the regression coefficients in the LR space, and its HR patch can be estimated by applying the same coefficients to the corresponding multiple HR regressors. Experimental results on benchmark dataset demonstrate that our proposed method could achieve competitive results both numerically and visually compared with some state-of-the-art methods.