Fusion diversion network for fast, accurate and lightweight single image super-resolution

In recent years, deep convolution neural network has been widely used in image super-resolution and achieved great performance. As the network becomes deeper and deeper, the accuracy of reconstruction is higher and higher. However, it also brings a large increase in the number of parameters and computational complexity, which makes the practical application more and more difficult. In this paper, we propose an efficient image super-resolution method based on fusion diversion network (FDN), where diversion and fusion block serves as the basic build module. By using the fusion and diversion mechanism, the information can be fully interactive and transferred in the network, and the expression ability of the model can be effectively improved. Extensive experimental results show that even with much fewer layers, the proposed FDN achieves the competitive results in both accuracy and speed.