Single image super-resolution based on a modified U-net with mixed gradient loss

Abstract

Single image super-resolution (SISR) is the task of inferring a high-resolution image from a single low-resolution image. Recent research on super-resolution has achieved great progress because of the development of deep convolutional neural networks in the field of computer vision. Existing super-resolution methods have high performances in mean square error (MSE), but most methods fail to reconstruct an image with shape edges. To solve this problem, the mixed gradient error, which is composed by MSE and mean gradient error, is proposed and applied to a modified U-net network. The modified U-net removes all batch normalization layers and one of the convolution layers in each block. The operation reduces the parameter number and therefore accelerates the model. Compared with the existing SISR algorithms, the proposed method has better performance and time consumption. The experiments demonstrate that modified U-net with mixed gradient loss yields high-level results on three widely used datasets, SET14, BSD300 and Manga109, and outperforms other state-of-the-art methods on the text dataset, ICDAR2003. Code is available online