Low-Light Image Enhancement With Regularized Illumination Optimization and Deep Noise Suppression

Yu Guo, Yuxu Lu, Ryan Wen Liu, Meifang Yang, Kwok Tai ChuiIEEE Access 2020 · DOI

Introduction

Maritime safety and traffic supervision rely heavily on computer vision for object detection and navigation. However, at night or in low-light conditions, maritime images suffer from two major issues: extremely low contrast and significant random noise. Existing solutions often fall into two camps: traditional Retinex-based methods, which frequently over-amplify noise when boosting light, and end-to-end deep learning models, which can struggle with the specific geometric structures of maritime scenes or require massive, perfectly paired datasets.

This paper introduces a hybrid framework that bridges this gap. By combining classical variational optimization with deep learning, the proposed method tackles the "blind" denoising problem—removing noise without prior knowledge of noise levels—while preserving the sharp edges of ships and horizons essential for maritime applications.

Methodology

The proposed framework is based on Retinex Theory, which models an image ($I$) as the product of reflectance ($R$) and illumination ($L$): $I = R \cdot L$. The process is divided into two primary stages: regularized illumination optimization and deep noise suppression.

Overview of the proposed hybrid enhancement framework, showing the transition from a low-light input to the final denoised result.

1. Regularized Illumination Optimization

The process begins with an initial coarse illumination map generated via the Max-RGB method. To refine this map, the authors employ a hybrid variational model. This model is a key innovation, combining $L_0$-norm gradient sparsity to maintain sharp structural edges with Relative Total Variation (RTV) to smooth out unwanted textures. The refined map is then adjusted using adaptive gamma correction, which varies based on the local mean brightness of the image to ensure natural-looking results.

2. Deep Noise Suppression

Brightening a low-light image inevitably amplifies underlying noise. To address this, the authors implement a blind denoising network composed of two specialized sub-networks:

Architecture of the blind denoising network, featuring the E-Net for noise estimation and the D-Net for non-blind denoising.
  • E-Net (Estimation): A 5-layer CNN that predicts a noise level map from the brightened image.
  • D-Net (Denoising): A 16-layer modified U-Net that utilizes both the noisy image and the predicted noise map to produce a clean, sharp final output.

By decoupling the physical lighting model (handled by optimization) from the statistical noise model (handled by deep learning), the framework achieves a balance that pure end-to-end networks often miss.

Experimental Setup and Datasets

The study utilizes several data sources for training and evaluation:

  • Synthetic Maritime Dataset: Created by darkening the V-channel of high-quality maritime images in the HSV color space and adding additive Gaussian noise at various levels ($\sigma = 5, 15, 25$).
  • Realistic Maritime Images: A collection of real-world low-light images captured in maritime environments to test generalization.
  • Training Data: The denoising sub-networks were trained on 2,000 noise-free images with synthetic noise ranging from $\sigma \in [0, 50]$.

Results and Discussion

The proposed method was compared against ten state-of-the-art (SOTA) methods, including LIME, Retinex-Net, and MBLLEN.

Quantitative Performance: In synthetic tests, the method achieved significantly higher PSNR (often >21 dB compared to SOTA's 14-19 dB) and SSIM (>0.95), indicating superior structural and color fidelity.

Qualitative Analysis: Visually, the method avoids "black halo" artifacts and color distortions. It is particularly effective in "sky" regions, which typically become grainy in other frameworks. In real-world tests without ground truth, the method scored best in NIQE (Naturalness Image Quality Evaluator), proving it produces images that appear natural to human observers.

A key insight is that mathematical variational models allow for better control over structural preservation—such as the thin lines of a ship's mast—than black-box neural networks. The combination of $L_0$-norm and RTV successfully separates textural details from main structures.

Conclusion and Future Work

While the framework provides superior image quality, its primary limitation is computational speed. The iterative numerical optimization for illumination refinement takes roughly 40 seconds for a $720 \times 1280$ image on a modern PC. Future work will focus on GPU acceleration and further optimization of the variational model to achieve real-time performance, a prerequisite for live maritime navigation systems.

Summary updated

BibTeX
@article{guo2020low,
  author = {Yu Guo and Yuxu Lu and Ryan Wen Liu and Meifang Yang and Kwok Tai Chui},
  title = {{Low-Light Image Enhancement With Regularized Illumination Optimization and Deep Noise Suppression}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3015217},
}
Figure