Super-resolution from learning the enhancement ratio and texture/residual dictionary

Image super resolution (SR) is the process of generating a high-resolution (HR) image by using one or more low-resolution (LR) inputs. In this study, a framework toward the single image SR process is investigated. A single image SR algorithm framework containing multiple steps is proposed. First, the enhancement ratio of each input image patch is estimated from the original image and the respective sub-sampling image patch. The enhancement ratio is estimated according to each pair of input LR patch and its down-sampled version. To construct an HR image patch, the input patch up-sample is then multiplied by the estimated enhancement ratio to generate an SR patch. Subsequently, the learned residual information on interest images is added to the initial super-resolved image, refining the result. The iterative back-projection procedure is then adapted to provide excellent visual quality. The experimental results indicate that the proposed framework provides high constraint and excellent visual quality, particularly regarding non-smooth texture areas.