Abstract

Super-resolution is a fundamental problem in computer vision which aims to overcome the spatial limitation of camera sensors. While significant progress has been made for single image super-resolution, most existing algorithms only perform well on unrealistic synthetic data, which limits their applications in real scenarios. In this paper, we study the problem of real-scene single image super-resolution to bridge the gap between synthetic data and real captured images. Specifically, we focus on two important problems of existing super-resolution algorithms: first, lack of realistic training data; second, underutilization of the information recorded by the cameras. To address the first issue, we propose a new pipeline to generate more realistic training data by simulating the imaging process of digital cameras. For the second problem, we develop a two-branch convolutional neural network to exploit the originally-recored radiance information in raw images. In addition, we propose a dense channel-attention block for better image restoration as well as a learning-based guided filter network for more effective color correction. Our model is able to generalize to different cameras without deliberately training on images from specific camera types. Extensive experiments demonstrate that the proposed algorithm and the new data generation pipeline can help recover fine details and clear structures, and achieve superior results for single image super-resolution in real scenarios.

Code GitHub