IATN: illumination-aware two-stage network for low-light image enhancement

Shuying Huang, Huiying Dong, Yong Yang, Yingzhi Wei, Mingyang Ren, Shuzhao Wang

Abstract: Images captured in low brightness environment have issues with low contrast and noise due to uneven lighting, which can seriously affect the accuracy of high-level computer vision tasks. Currently, most enhancement methods still suffer from color distortion and noise amplification. To overcome these issues, this paper proposes an illumination-aware two-stage network (IATN) for low-light image enhancement. In the first stage, a tiny illumination estimation network based on Retinex theory is constructed to generate a coarse enhanced image. In the second stage, an illumination-aware correction network (IACN) is designed by building an illumination map to guide the reconstruction of features, which can reduce color distortion and suppress noise in the results obtained in the first stage, thereby obtaining refined enhancement results. In IACN, considering the exposure difference in different regions of the image caused by uneven lighting, multiple illumination-aware modules are constructed to correct features at different scales by utilizing the long-range dependence of features. Numerous experiments conducted on public benchmark datasets have shown that our IATN generates enhanced images that are more natural, colorful, and superior to some state-of-the-art methods. The source code of this work will be available on GitHub.