A Self-Learning Single Image Super-Resolution by Considering Consistency in Adjacent Pixels

In this paper, we propose a self-learning single image super-resolution. In our proposed method, adjacent pixels information in smooth area is used. Low and high-resolution pyramids are built by applying up-sampling and down-sampling techniques on input image, as training data. In training phase, we apply support vector regression (SVR) to model the relationship between the pair of low and high-resolution images. For each patch in the low-resolution image, sparse representation is extracted as a feature vector. In this paper, in order to reduce the edge blurring effects, we first separate edge pixels from non-edge pixels. In the smooth area, because of the similar colors around the each pixel, the center pixel value is determined by considering the reconstructed adjacent pixels. Experimental results show that the proposed method is quantitatively and qualitatively outperform the competitive super-resolution approaches.