An efficient Group Skip-Connecting Network for image super-resolution

摘要

This paper proposes an efficient Group Skip-Connecting Network (GSCN) for image super-resolution to increase the reconstruction performance and reduce the running time. Different from ResNet and DenseNet, which are two typical representative networks based on the skip connection, the proposed network designs a group skip connection layer to fuse the features between different separated groups. Hence GSCN can enjoy the merits of the lightweight from the group convolution and the high-efficiency from the skip connection. Considering the difficulty of restoring the high-frequency details, we further propose an Enhanced-GSCN (E-GSCN), which uses an enhanced attention module (EAM) to adaptively enhance the high-frequency details of the intermediate layer features. In addition, to better exploit the hierarchical information, we also design a hierarchical feature fusion framework to adaptively learn the hierarchical information at both the local and global levels. Experiments on the benchmark test sets show that the proposed models are more efficient than most of the state-of-the-art methods.