Single Image Super-Resolution with Gradient Guidance

Recovering high-frequency image details such as edges and textures is a challenge of image super-resolution. To improve the reconstruction accuracy, image gradient maps are widely introduced as an additional input or a regularized term directly to existing methods. We argue that the best way to exploit gradient information is to learn from the training data. We propose a convolutional neural network for image super-resolution which is guided by image gradient maps. The gradient guidance provides a selective condition during super-resolution, leading to a more faithful super-resolved image. Our method is a flexible framework for image super-resolution, which can be easily incorporated into existing methods. Extensive benchmark evaluation shows that the proposed method achieves highly competitive performance, outperforming state-of-the-art performance in single image super-resolution.