Single Image Super-Resolution Using Frequency-Dependent Convolutional Neural Networks

In this paper, we propose a single image super-resolution (SR) method based on frequency-dependent training of convolutional neural networks. Several researchers have focused on the reconstruction of super-resolution images by training a single convolutional neural network. In the proposed method, we divided the input images into three sub-frequency groups and then trained a convolutional neural network for each sub-frequency group. Then, the final output images were reconstructed by combining the SR images from the multiple networks. Experimental results show that the proposed training method produces promising performance.