Retinex-based Image Denoising / Contrast Enhancement Using Gradient Graph Laplacian Regularizer

Yeganeh Gharedaghi, Gene Cheung, Xianming Liu

TL;DR

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), the authors use a novel "Gradient Graph Laplacian Regularizer" to ensure smooth lighting transitions while preserving sharp object details, achieving state-of-the-art visual quality at a fraction of the computational cost of previous methods.

What Problem Does This Paper Address?

Capturing images in low light is a double-edged sword. Because sensors collect fewer photons, the resulting images are both dark (low contrast) and grainy (high acquisition noise).

The standard "Retinex" approach to fixing this involves splitting an image into two components: reflectance (the actual colors and details of objects) and illumination (how light hits those objects). However, traditional methods often fall into a trap: brightening the illumination component inadvertently amplifies the noise hidden in the reflectance. While complex optimization techniques exist to solve this jointly, they are often too slow or memory-intensive for mobile devices. This paper seeks a "sweet spot"—an algorithm that is mathematically robust enough to denoise and enhance simultaneously, yet fast enough for practical use.

Methods

The authors frame the problem using Graph Signal Processing (GSP). They treat the image pixels as nodes on a graph and use "regularizers" to penalize unlikely configurations.

  1. Reflectance Modeling (GLR): They assume reflectance is piecewise constant (flat areas of color). They use a standard Graph Laplacian Regularizer (GLR), which essentially tells the algorithm: "If two neighboring pixels have similar colors, keep them that way."
  2. Illumination Modeling (GGLR): This is the paper's key innovation. Illumination isn't usually flat; it's smooth and gradient-like. To capture this, they use a Gradient Graph Laplacian Regularizer (GGLR). Instead of regularizing the pixel values themselves, they regularize the differences (gradients) between pixels. This promotes continuous piecewise planar surfaces—think of a smooth spotlight hitting a wall.
  3. Optimization: Both regularizers are formulated as quadratic objectives. This is a big deal because quadratic problems can be solved using linear systems (specifically the Conjugate Gradient method), which is much faster than the non-differentiable $\ell_1$-norm optimizations used in older "Total Variation" methods.
  4. Preconditioning: To speed up the math even further, they designed a custom "Jacobi-like" preconditioner. This fixes numerical instabilities that occur when the algorithm encounters very dark or very bright patches, reducing the number of iterations needed to find a solution.

Datasets

The authors evaluated their method on 12 standard test images (including "Moonlight," "Cars," and "Man") sourced from established low-light enhancement benchmarks (specifically datasets provided by Fu et al. and Guo et al.). They added synthetic Gaussian noise to these images to test the joint denoising/enhancement capabilities.

Key Insights and Ideas

  • Quadratic is King for Speed: By choosing regularizers that can be expressed in quadratic form, the authors moved the problem from the realm of "slow iterative optimization" to "fast linear algebra."
  • Gradients Matter for Light: Standard graph tools often produce "blocky" artifacts. By applying the graph Laplacian to the gradient of the illumination rather than the illumination itself, the authors achieved much more natural-looking light transitions.
  • The Preconditioning "Aha": The authors realized that small values in the illumination or reflectance components make the underlying math "stiff" (ill-conditioned). Their diagonal preconditioner acts like a shock absorber, smoothing out the optimization process and preventing the solver from getting stuck.

Key Results

  • Speedup: The proposed method is significantly faster than the competing LR3M model. On average, it processed images in 34.74 seconds compared to LR3M’s 312.39 seconds—roughly a 9x speedup.
  • Visual Quality: The method outperformed others in terms of Lightness-Order-Error (LOE), a metric that measures how well the natural "order" of brightness is preserved. A lower LOE means the enhanced image looks more realistic and less "processed."
  • Denoising: Unlike simpler methods (like LIME) that amplify noise during brightening, this approach successfully suppressed noise while enhancing contrast, producing cleaner edges and smoother surfaces.

Limitations and Future Work

  • Parameter Sensitivity: The model relies on four empirical parameters ($\mu_r, \sigma_r, \mu_l, \sigma_l$). While the authors provide suggested values, these might need tuning for vastly different lighting conditions.
  • Deep Learning Integration: The authors suggest "unrolling" this iterative algorithm into neural network layers. This would allow the parameters to be learned automatically from data (end-to-end training) rather than being set manually, potentially improving performance on diverse datasets.
  • Mobile Deployment: While faster than SOTA, 34 seconds for a high-res image is still a bit slow for a "real-time" camera app; further optimization or hardware acceleration would be needed for instant results.

Summary updated

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.
BibTeX
@inproceedings{gharedaghi2023retinex,
  author = {Yeganeh Gharedaghi and Gene Cheung and Xianming Liu},
  title = {{Retinex-based Image Denoising / Contrast Enhancement Using Gradient Graph Laplacian Regularizer}},
  booktitle = {ICIP},
  year = {2023},
  doi = {10.1109/icip49359.2023.10222331},
}

Loading related content...

Figure