FSFN: feature separation and fusion network for single image super-resolution

In recent years, image super-resolution (SR) based on deep learning technology has made significant progress. However, most methods are difficult to apply in real life because of their large parameters and heavy computation. Recently, residual learning has been widely applied to the problem of super-resolution. It can make the shallow features extracted from the input image act on each middle layer through long and short connection. Therefore, residual learning can be focused on processing high-frequency feature information, which significantly improves the SR performance of the network. However, with the improvement of network depth, the features that can be effectively utilized are still the shallow ones extracted from the input image. In this paper, we propose the feature separation and fusion network(FSFN). We further enrich the high-frequency feature information by separating and fusing the extracted and unextracted features in the internal shallow layer of each feature separation and fusion module. As the depth of the network increases, the shallow features extracted from the input image can be updated in a direction closer to those extracted from the real high-resolution image. A large number of experimental results show that this method has a strong performance compared with the existing SR algorithm with similar parameters and computation.