Abstract
Low-light image enhancement (LLIE) is the process of improving the quality of images taken in low-light conditions while striking a balance between enhancing image illumination and maintaining their natural appearance. This involves reducing noise, enhancing details, and correcting colors, all while avoiding artifacts such as halo effects or color distortions. We propose LIVENet, a novel deep neural network that jointly performs noise reduction on lowlight images and enhances illumination and texture details. LIVENet has two stages: the image enhancement stage and the refinement stage. For the image enhancement stage, we propose a Latent Subspace Denoising Block (LSDB) that uses a low-rank representation of low-light features to suppress the noise and predict a noise-free grayscale image. We propose enhancing an RGB image by eliminating noise. This is done by converting it into YCbCr color space and replacing the noisy luminance (Y) channel with the predicted noise-free grayscale image. LIVENet also predicts the transmission map and atmospheric light in the image enhancement stage. LIVENet produces an enhanced image with rich color and illumination by feeding them to an atmospheric scattering model. In the refinement stage, the texture information from the grayscale image is incorporated into the improved image using a Spatial Feature Transform (SFT) layer. Experiments on different datasets demonstrate that LIVENet’s enhanced images consistently outperform previous techniques across various quality metrics.
Introduction
Capturing images in low-light environments is a fundamental challenge in computer vision. These images typically suffer from three major issues: extreme noise, low contrast, and color distortion. Existing Low-Light Image Enhancement (LLIE) techniques often fall into two traps: noise amplification and color/texture loss. For instance, aggressive illumination adjustments can cause unrealistic color shifts, such as a white wall turning orange.
LIVENet aims to decouple these problems, treating noise reduction and illumination enhancement as distinct but collaborative tasks. By predicting a clean grayscale version of the scene and swapping it into the luminance channel of an enhanced color map, the model achieves state-of-the-art results in restoring realistic colors and sharp textures without typical noise amplification.
Methods
LIVENet uses a coarse-to-fine architecture divided into two main stages:
Stage 1: Image Enhancement & Denoising
This stage consists of two parallel modules:
- Grayscale Image Prediction (GSIP): This module predicts a noise-free grayscale version of the image. It introduces a novel Latent Subspace Denoising Block (LSDB), which uses low-rank matrix representation to separate the "true" underlying structure of the image from high-rank noise.
- Coarse Map Image Generation (CMIG): This module uses an Atmospheric Scattering Model (ASM)—a technique usually reserved for dehazing—to estimate light transmission and atmospheric light. This produces a brightened but potentially noisy "coarse" color map.
The key innovation lies in converting the coarse color map into the YCbCr color space. Since the human eye is more sensitive to luminance (Y) than color (CbCr), and noise is most prevalent in the Y channel, the authors replace the noisy Y channel with the clean grayscale image predicted by the GSIP module.
Stage 2: Refinement
The denoised coarse map is then passed through a refinement stage. This stage uses Spatial Feature Transform (SFT) layers to "condition" the image. It takes the high-quality structural features from the grayscale prediction and injects them back into the color image to recover lost textures and fine-tune the final appearance.
Results and Discussion
The model was trained and evaluated on several benchmark datasets, including paired sets like LOL-v1 and LOL-v2-real, and unpaired sets for testing generalization (DICM, LIME, MEF, NPE, and VV).
LIVENet consistently outperformed previous state-of-the-art models like LLFlow, SNR, and DCCNet. On the LOL-v1 dataset, LIVENet achieved a PSNR of 24.68, significantly higher than LLFlow (23.84) and DCCNet (22.97). Qualitative results show that LIVENet avoids the "over-saturated" or "washed out" look common in other models. It handles reflections much better, creating fewer artifacts around bright light sources. On unpaired datasets, LIVENet achieved the best (lowest) NIQE scores in 4 out of 5 categories, proving it generalizes well to real-world photos outside of the training set.
Conclusion
LIVENet provides a robust framework for simultaneous denoising and enhancement of low-light images. By using a clean grayscale "skeleton" as a structural anchor and repurposing the Atmospheric Scattering Model, it achieves state-of-the-art performance. While the current model is designed for single images, future work will focus on extending LIVENet to handle low-light videos by incorporating temporal and spatial information, as well as optimizing computational complexity for real-time applications.
Summary updated