Single Image Super-Resolution Neural Network Using Frequency-Domain Information

Singe image super-resolution (SISR) has been widely researched because of its application value. Despite the remarkable performance of the super-resolution methods based on deep convolutional neural networks, the lack of high-frequency information in the recovered images remains a core problem. In this work, we proposed a new loss function based on the spectral information of the image and discussed the effectiveness of this loss function in guiding the recovery of image details. Our experiment results on extensive benchmark datasets reveal the superiority of our proposed loss function, and the visual effects of the networks trained with the proposed loss function were significantly improved compared with those without this loss function.