Illumination-guided semi-supervised network for low-light image enhancement jointly with denoising

The biggest challenges faced by low-light image enhancement(LLIE) include under/overexposure, color distortion, and the amplified noises after enhancement. Most of the existing methods are to uniformly enhance the global image, which leads to the neglect of some essential semantic information and the inherent factors that decrease the image quality are also amplified at the same time. To address these issues, we propose a semi-supervised network guided by luminance information. The structure includes three parts: decomposition network(DecomNet), reconstruction network(ReconNet) and denoising network(DenoiseNet). DecomNet is constructed on the basis of a Retinex-based model we design, which can decompose low-light images into illumination layers without noise interference. And we propose a region-based illumination adjustment method to make the illumination layer extracted from DecomNet adjust adaptively. The illumination map after our adaptive adjustment is the key which is as embedded guidance for ReconNet to enhance the image luminance and color contrast balancedly on the whole. The final DenoiseNet combines the characteristics of the blind spot network to remove noise from the results of ReconNet. We conduct extensive experiments on multiple datasets, including paired data sets and no-reference data sets. The results demonstrate that our method is superior to other low-light image enhancement methods on multiple datasets.