Dense-Gated Network for Image Super-Resolution

Shumin Fan, Tianyu Song, Pengpeng Li, Jiyu Jin, Guiyue Jin, Zhongmin Zhu

Abstract: Deep-learning based methods have achieved great success in the field of Single Image Super-Resolution (SISR) by progressively exploring contextual and deep semantic features. However, existing methods do not make full use of scale-space features, resulting in the restoration of high-resolution image details being blurred. In order to address this issue, the Dense-Gated Network is proposed for SISR, called DGISR, which consists of two essential blocks: Pyramid Multi-scale Extraction Block (PMEB) and Gated Attention Distillation Block (GADB). The proposed PMEB allows capturing and integrating more complex features through pyramid pooling and multi-scale operation to enhance the modeling capability of the network. The proposed GADB can extract key regions of images and reduce data redundancy, while improving the convergence of the model. The proposed DGISR outperforms other methods in visual quality and quantitative metrics on the standard benchmark datasets, as demonstrated by experimental results. Our overall method significantly outperforms the state-of-the-art methods.