Learning noise-decoupled affine models for extreme low-light image enhancement

• Published in: Neurocomputing (Amsterdam) Vol. 448; pp. 21 - 29
• Main Authors: Liu, Maomei, Tang, Lei, Zhong, Sheng, Luo, Hangzai, Peng, Jinye
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 11.08.2021
• Subjects: Local affine color transformation; Low-light image enhancement; Noise suppression; Local affine color transformation; Low-light image enhancement; Noise suppression
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2021.03.107

How to handle the noise effectively is an important yet challenging problem for low-light image enhancement especially in real-world extreme low-light conditions. Furthermore, contrast enhancement and noise removal are coupled problems, it’s hard to trade off well between noise suppression and preservation of details. To this end, this paper proposes an end-to-end network for low-light image enhancement with a particular focus on handling this coupling relationship. The basic idea is to convert low-light image enhancement to local affine color transformations. Instead of image smooth denoising, a special noise processing mechanism is proposed to learn noise-decoupled affine models. Alternatively, to achieve efficient learning, the whole network is trained in bilateral space. Extensive experiments on several benchmark datasets have shown that the proposed method is very competitive to state-of-the-art methods. Especially when processing images captured in extreme low-light conditions, it has a significant advantage over other algorithms in reducing noise while retaining image details.