Single Image Super-Resolution via Squeeze and Excitation Network

Single Image Super-Resolution (SISR) has obtained unprecedented breakthrough with the development of Convolutional Neural Networks (CNN). A majority of these methods try to increase the depth of the network to obtain a larger receptive field. However, we found that blindly stacking feature maps and the simple cascading structure can not achieve a high rate of utilization in super-resolution reconstruction. In this paper, we propose a refined fully convolutional network for Single Image Super Resolution. Based on the assumption that features maps from different depths or the same depth but different channels have different contributions during image reconstruct, we introduce the Squeeze and Excitation (SE) network to evaluate the importance of different feature maps while building the network. Besides, densely connection operation is also conducted in the framework for a better use of the contextual information and feature maps. Extensive experiments demonstrates that the proposed method can enhance the restoration performance and achieve the state-of-the-art results in super-resolution task.