Non-Reference Enhanced Low-Light Image Estimation Using Zero-DCE with Filtering Refinement.

Abstract This paper introduces Modified Zero-Reference Deep Curve Estimation (Zero-DCE), a novel technique that treats light enhancement as image-specific curve estimation task using a deep curve network. Our network, powered by DCE-Net, trains a light-weight deep network to estimate pixel-wise and higher order curves, facilitating dynamic range adjustment in an images. The design of our curve estimation network prioritizes considerations such as pixel value range, differentiability, and monotonicity. An intriguing aspect of modified Zero-DCE lies in its departure from strict requirements on the reference images. It doesn't rely on unpaired or paired data during training. Instead, it leverages a set of meticulously crafted the non-reference loss functions. These loss functions implicitly evaluate image enhancement quality and guide the network's learning process. Despite its simplicity, our method demonstrates robust generalization across the diverse lighting conditions. To further refine enhanced images generated through this network, a post-processing step involves filtering to eliminate noise that emerges after transforming the low-light images into high-light ones. This noise reduction step significantly enhances images, resulting in remarkable improvements in SSIM (Structural Similarity Index) and PSNR (Peak Signal-to-Noise Ratio) metrics when comparing the de-noised enhanced images to their high-light images.