Image super-resolution using progressive residual multi-dilated aggregation network

Recently, single image super-resolution based on convolutional neural network (CNN) has achieved considerable improvements against traditional methods. However, it is still challenging for most CNN-based methods to obtain satisfactory reconstruction quality for large-scale factors. To solve the issues, we propose a progressive residual multi-dilated aggregation network (PRMAN), which performs multi-level ×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document}2 upsampling to reconstruct images with large-scale factors. Specially, we design a residual multi-dilated aggregation block to simplify the model and supply enriched features with different receptive fields. Simultaneously, the channel attention mechanism is adopted to select informative features. Furthermore, to speed up the convergence and attain better performance, we train the model with two-stage training strategy. Extensive experimental results show that our proposed PRMAN exceeds the state-of-the-art methods in most cases.