Hierarchical flow learning for low-light image enhancement

Abstract

Low-light images often have defects such as low visibility, low contrast, high noise, and high color distortion compared with well-exposed images. If the low-light region of an image is enhanced directly, the noise will inevitably blur the whole image. Besides, according to the retina-and-cortex (retinex) theory of color vision, the reflectivity of different image regions may differ, limiting the enhancement performance of applying uniform operations to the entire image. Therefore, we design a Hierarchical Flow Learning (HFL) framework, which consists of a Hierarchical Image Network (HIN) and a normalized invertible Flow Learning Network (FLN). HIN can extract hierarchical structural features from low-light images, while FLN maps the distribution of normally exposed images to a Gaussian distribution using the learned hierarchical features of low-light images. In subsequent testing, the reversibility of FLN allows inferring and obtaining enhanced low-light images. Specifically, the HIN extracts as much image information as possible from three scales, local, regional, and global, using a Triple-branch Hierarchical Fusion Module (THFM) and a Dual-Dconv Cross Fusion Module (DCFM). The THFM aggregates regional and global features to enhance the overall brightness and quality of low-light images by perceiving and extracting more structure information, whereas the DCFM uses the properties of the activation function and local features to enhance images at the pixel-level to reduce noise and improve contrast. In addition, in this paper, the model was trained using a negative log-likelihood loss function. Qualitative and quantitative experimental results demonstrate that our HFL can better handle many quality degradation types in low-light images compared with state-of-the-art solutions. The HFL model enhances low-light images with better visibility, less noise, and improved contrast, suitable for practical scenarios such as autonomous driving, medical imaging, and nighttime surveillance. Outperforming them by PSNR = 27.26 dB, SSIM = 0.93, and LPIPS = 0.10 on benchmark dataset LOL-v1. The source code of HFL is available at https://github.com/Smile-QT/HFL-for-LIE.