Robust learning-based super-resolution

Learning-based super-resolution algorithms synthesize a high-resolution image based on learning patch pairs of low- and high-resolution images. However, since a low-resolution patch is usually mapped to multiple high-resolution patches, unwanted artifacts or blurring can appear in super-resolved images. In this paper, we propose a novel approach to generate a high quality, high-resolution image without introducing noticeable artifacts. Introducing robust statistics to a learning-based super-resolution, we efficiently reject outliers which cause artifacts. Global and local constraints are also applied to produce a more reliable high-resolution image. Experimental results demonstrate that the proposed algorithm can synthesize higher quality, higher-resolution images compared to the existing algorithms.