A Deep Hybrid Few Shot Divide and Glow Method for Ill-Light Image Enhancement

Abstract

Images captured in low and ill-lighting conditions with non-uniform illumination distribution contain over-exposed and under-exposed regions simultaneously. The existing methods use handcrafted parameters for the ill-posed image decomposition and rely on image pairs or priors. The capacity of these models is limited to specific lighting conditions. The existing deep learning-based models are also not competitive when there is a lack of large-scale dataset, and some models even require paired training data. But in practice, it is challenging to capture low-light and normal-light images of the same scene. In this paper, we propose a few shot divide and glow (FSDG) method to enhance low ill-light images without exactly relying on the large scale and paired training data. We divide the input images into reflection ${R}$ and illumination transmission $T$ components by using MID-Net and amplify the illumination map in a Glow-Net. A contrast enhancement strategy is proposed to upgrade image division, which maintains regularization consistency for a few- shot divide and glow network (FSDG-Net). The FSDG-Net is end-to-end trained to learns from the correlation consistency of the input image decomposition itself. Experiments are organized under both very low exposure and ill-light conditions, where a new dataset is also proposed with challenging test images. Results show that our method consistently shows the superior performance when comparing to other state-of-the-art approaches.