Image Super-Resolution Based on the Down-Sampling Iterative Module and Deep CNN

Abstract

Most deep learning-based image SR algorithms do not apply the down-sampling to the reconstructed process. Given this fact and inspired by the iteration idea, we propose a novel image SR method based on the down-sampling iterative module and deep CNN, which explores a new basic iterative module combining up- and down-sampling processes. Each iteration of the iterative module generates the intermediate LR prediction and the HR image. The final reconstructed result is obtained by the weighted summation of the intermediate predicted images generated by multiple iterations. During the training, we adopt the adaptive loss function to achieve fast convergence and accurate reconstruction. Detailed experimental comparisons and analyses show that our method is superior to some state-of-the-art methods in objective performance evaluation and visual effects.