Abstract: In recent years, deep convolutional neural networks have performed well in solving single image super-resolution problems. For improving the disadvantage that, the deeper the layer number of convolutional neural network is, the greater the amount of calculation is, the slower the real-time reconstruction speed is, combined with the existing convolutional network model, a lightweight network structure is proposed. This paper reduces the number of network layers in the neural network layer and uses channel split to build a multi-scale enhanced structure of local features. Then it combines the residual network for model construction. Experiment results show that, compared with LapSRN method, VDSR method, and traditional interpolation method, this method is faster in real-time reconstruction and is not weaker than others in peak signal to noise ratio and structural similarity.

Key words: convolutional neural networks, super-resolution, lightweight networks, channel split, residual network