Multi-scale cascaded network with high-low frequency for low-light image enhancement

Low-light images affect human visual perception and computer vision downstream tasks because of low illumination, blurred details, and severe noise. Most existing methods optimize the illumination prior and reflectance of the image to accomplish low-light image enhancement. However, in these methods, the acquired illumination features cannot be effectively restored, and the spatial structure cannot be adequately rendered. To address the above issues, this paper proposes a high and low-frequency enhanced low-light image enhancement framework based on a cascaded UNet. To obtain high-quality illumination features, we design a UNet architecture to capture both local and global semantic priors, which are then used to illuminate low-light images. The second UNet module extracts local details and fine spatial structures to repair degraded image information using illumination-guided restoration with high and low-frequency enhancements. At the second UNet skip connections, we quote the channel reduction attention mechanism to enhance the interaction of feature channel information. Experiments on public datasets show that the proposed method achieves superior enhancement performance.