An improved method for single image super-resolution based on deep learning

AbstractThis paper strives for presenting an improved method for single image super-resolution based on deep learning, and therefore, a well-designed network structure is proposed by simultaneously considering the merits of convolutional sparse coding (CSC) and deep convolutional neural networks (CNN). In our model, contrary to most existing methods that directly operate on the raw input, we first perform a global decomposition on the input based on CSC for the purpose of extracting two specific components from it. Since the generated components are designed to have predefined physical meanings (i.e., residual or smooth), they can be discriminatively super-resolved according to their distinctive appearances. Specifically, a strong preference is given to the residual one as it is much more crucial to our task, while the other should just provide a quick reference. Based on this analysis, deep CNN and plain interpolation are selected to map them, respectively. In all, the proposed model integrates the above procedures into a completely end-to-end trainable deep network. Thorough experimental results demonstrate that our proposed network is able to gain considerable accuracy from this deep and delicate architecture, thereby outperforming many recently published baselines in terms of both objective evaluation and visual fidelity.