Effective Use of Low-level Features for Image Super-Resolution

In this paper we present an improved method based on Super-Resolution Convolutional Neural Network(SRCNN) to explore the application of lightweight CNN structure in single-image super-resolution. We keep the three layer depth unchanged and add a skip connection from input to the output and add twice skip connections across the second convolutional layer. This method can better transfer low-level features to higher layers to improve the reconstruction. In addition, we also study the effect of dataset augmentation to results. The simulation shows that the proposed method is easy to converge during training and has better image reconstruction quality.