Explore unsupervised exposure correction via illumination component divided guidance

Abstract

The capturing of images with poor exposures remains a major source of errors in camera-based photography. Under- or over-exposure problems greatly reduce contrast and naturalness, and the disappearance of detailed content affects the understanding of an image. While recent works have built a great deal of learning frameworks to address this problem, they mainly focus on low-light or general image enhancement. Towards this end, we explore an unsupervised exposure correction framework. Our model disentangles the exposure correction into two easier sub-tasks. Different branches focus on different missions, making it easy for the network to solve coupling correction problems. Additionally, a novel attention-based pixel-wise brightness estimation makes it possible for coarse-to-fine mapping adjustment and correction. Experiments on various widely used benchmark datasets show that, by relying on illumination perception and component divided guidance, the proposed network can outperform previous state-of-the-art methods, especially when recovering fine detailed contents.