Image super-resolution using multi-level high-frequency feature fusion

Recently, single image super-resolution based on deep convolutional neural networks has made significant progress, but there are still problems such as lacking high-frequency texture details and poor visual quality. As the network depth grows, the loss of high-frequency information becomes more and more serious. In this paper, we propose a method to enhance the high-frequency details by fusing multi-level high-frequency features through skip connection and high-pass filters. In order to enhance the representation of global features, our network structure combines both Transformer and channel attention as the base module. Finally, to further improve the visual perceptual quality, we design a contrast loss using gaussian blurred images as negative samples. Comprehensive experiments demonstrate the effectiveness of our method.