A Convolutional Neural Network with Two-Channel Input for Image Super-Resolution
Purbaditya Bhattacharya, Udo Zölzer · 2019
In this work a convolutional neural network with two input channels is proposed for image super-resolution. Initially, the original image is downsampled with bicubic and nearest neighbour interpolation methods and the low-resolution image pair is used as the input to our network. Additionally, the input channels are randomly swapped as an augmentation method during training. The proposed network is a combination of a feedforward architecture and a residual network architecture, to provide a smooth image estimate. We show that the additional input image obtained with nearest neighbour interpolation improves image super-resolution performance with our network, since it acts as an image specific prior information and constrains the solution space for the model to learn. We train the network end-to-end for multiple upscale factors, evaluate the method on standard test datasets, and show that the proposed approach can produce state-of-the-art results.