Single image super-resolution using self-examples and texture synthesis

Abstract

An algorithm for single image super-resolution based on example-based super-resolution and example-based texture synthesis is proposed. While many other techniques for single image super-resolution are mainly effective on edges, the proposed algorithm enhances both edges and texture detail. The algorithm does not use an additional example database as it uses self-examples to synthesize new detail and texture, assuming that images contain a sufficient amount of self-similarity. The texture synthesis component of the algorithm enables the re-synthesis of texture at the output resolution to achieve super-resolution. The algorithm aims to create plausible, visually pleasing detail rather than reconstructing the true high-resolution image. Experimental results for natural images confirm the algorithms ability to create visually pleasing results, but also indicate that its performance is highly content dependent. Future efforts will be aimed at improving the robustness of the method.

Fast Single Image Super Resolution Reconstruction via Image Separation

“…The basic super resolution scheme for cartoon component is related to the framework proposed by Freedman et al in [13]. Unlike other previously used learning based strategies [9], [14] which directly upscale the image into the desired magnification size and then add the details, this framework performs multiple upscaling steps of small scaling factors to achieve the desired magnification size. And the detailed high-frequency informations are merged into the images gradually in each small upscaling steps.…”

“…Also, when the desired magnification factor is large, their performance will degrade rapidly. To overcome the limitations of the reconstruction-based algorithms, some machine learningbased techniques have been proposed [8]- [14]. This kind of approaches usually contain two steps: 1) learning the coherence from a training data set including both LR and HR image patches; and 2) predicting the details of the HR image through utilizing those learned coherence.…”