Deep Residual Network for Single Image Super-Resolution

Abstract

This paper proposes a Deep Residual Network for Single Image Super-Resolution (DRSR). We build a deep model using residual units that remove unnecessary modules. We can build deeper network at the same computing resources with the modified residual units. Experiments shows that deepening the network structure can fully utilize the image contextual information to improve the image reconstruction quality. The network learns both global residuals and local residuals, making the network easier to train. Our network directly extracts features from Low-Resolution (LR) images to reconstruct High-Resolution (HR) images. Computational complexity of the network is dramatically reduced in this way. Experiments shows that our network not only performs well in subjective visual effect but also achieves a high level in objective evaluation index.