Improvement on learning-based super-resolution by adopting residual information and patch reliability

Learning-based super-resolution algorithms synthesize high-resolution details by using training data. However, since an input image does not belong to a training image set, there is a limitation in recovering its high-frequency details. In our approach, we build and utilize residual training data to complement missing details. We first estimate a pair of mid- and high-frequency images of each training image by using ordinary training data. We then build residual training data by obtaining the residual mid-and high-frequency images that denote the difference between the estimation and original. Thereby, we can synthesize high-resolution details better by using both ordinary and residual training data sets. In addition, in order to use training data more efficiently, we adaptively select low-resolution patches in an input image. Experimental results demonstrate that the proposed method can synthesize higher-resolution images compared to the existing algorithms.