Abstract
Images captured in poorly lit conditions are often corrupted by acquisition noise. Leveraging recent advances in graph-based regularization, we propose a fast Retinex-based restoration scheme that denoises and contrast-enhances an image. Specifically, by Retinex theory we first assume that each image pixel is a multiplication of its reflectance and illumination components. We next assume that the reflectance and illumination components are piecewise constant (PWC) and continuous piecewise planar (PWP) signals, which can be recovered via graph Laplacian regularizer (GLR) and gradient graph Laplacian regularizer (GGLR) respectively. We formulate quadratic objectives regularized by GLR and GGLR, which are minimized alternately until convergence by solving linear systems -- with improved condition numbers via proposed preconditioners -- via conjugate gradient (CG) efficiently. Experimental results show that our algorithm achieves competitive visual image quality while reducing computation complexity noticeably.
1. Introduction
Capturing images in low-light environments presents a significant challenge: the resulting images are typically both dark (low contrast) and grainy due to high acquisition noise. Traditional Retinex-based methods decompose an image into reflectance (intrinsic details) and illumination (lighting) components. However, a common pitfall is that brightening the illumination often amplifies noise in the reflectance.
This paper introduces a fast, graph-based approach to simultaneously denoise and brighten low-light images. By combining Retinex theory with Graph Signal Processing (GSP), we utilize a novel "Gradient Graph Laplacian Regularizer" (GGLR) to ensure smooth lighting transitions while preserving sharp object details. This approach achieves state-of-the-art visual quality at a fraction of the computational cost of previous methods.
2. Methodology
The proposed framework treats image pixels as nodes on a graph, employing specific regularizers to penalize unlikely signal configurations.
2.1 Reflectance and Illumination Modeling
- Reflectance Modeling (GLR): Reflectance is assumed to be piecewise constant. We use a standard Graph Laplacian Regularizer (GLR) to encourage similarity between neighboring pixels with similar colors.
- Illumination Modeling (GGLR): Illumination is typically smooth and gradient-like. Our key innovation is the Gradient Graph Laplacian Regularizer (GGLR). Instead of regularizing pixel values directly, we regularize the gradients between pixels, promoting continuous piecewise planar surfaces.
2.2 Optimization and Preconditioning
Both regularizers are formulated as quadratic objectives, allowing the problem to be solved via linear systems using the Conjugate Gradient (CG) method. This is significantly faster than the non-differentiable L1-norm optimizations used in traditional Total Variation methods. To further accelerate convergence, we designed a custom "Jacobi-like" diagonal preconditioner to address numerical instabilities in very dark or bright patches.
3. Experimental Results
The method was evaluated on 12 standard test images from established low-light enhancement benchmarks, with synthetic Gaussian noise added to test joint denoising and enhancement.
3.1 Visual Quality and Denoising
Unlike simpler methods like LIME that amplify noise during brightening, our approach successfully suppresses noise while enhancing contrast. This produces cleaner edges and smoother surfaces, as shown in the enhanced results.
3.2 Computational Efficiency
The proposed method demonstrates a significant speedup over competing models. On average, it processed images in 34.74 seconds, compared to 312.39 seconds for the LR3M model—a roughly 9x improvement. Furthermore, it outperformed others in Lightness-Order-Error (LOE), indicating more realistic brightness preservation.
4. Discussion and Future Work
The transition to quadratic forms for regularizers is a key factor in achieving high speed. By applying the graph Laplacian to gradients rather than values, we avoid "blocky" artifacts common in standard graph tools.
Future work includes addressing parameter sensitivity through deep learning integration, potentially "unrolling" the iterative algorithm into neural network layers for automatic parameter learning. Additionally, further optimization is needed for real-time mobile deployment, as 34 seconds remains high for instant camera applications.
Summary updated