Dual Features Extraction Network for Image Super-Resolution

With the development of deep convolutional neural network, recent research on single image super-resolution (SISR) has achieved great achievements. In particular, the networks, which fully utilize features, achieve a better performance. In this paper, we propose an image super-resolution dual features extraction network (SRDFN). Our method uses the dual features extraction blocks (DFBs) to extract and combine low-resolution features, with less noise but less detail, and high-resolution features, with more detail but more noise. The output of DFB contains the advantages of low- and high-resolution features, with more detail and less noise. Moreover, due to that the number of DFB and channels can be set by weighting accuracy against size of model, SRDFN can be designed according to actual situation. The experimental results demonstrate that the proposed SRDFN performs well in comparison with the state-of-the-art methods.