One-stage Retinex-based Framework (ORF)

Updated 10 July 2025

One-stage Retinex-based Framework (ORF) is a unified image enhancement technique that decouples illumination and reflectance in a single, interpretable process.
It effectively addresses challenges like non-uniform lighting, shadows, and noise by combining decomposition, enhancement, and artifact suppression in one stage.
Recent deep and transformer variants of ORF have improved contrast, detail preservation, and computational efficiency across diverse imaging applications.

A one-stage Retinex-based Framework (often abbreviated as ORF) refers to a class of image processing and enhancement methodologies that directly utilize Retinex theory to decouple illumination and reflectance in images within a single, unified computational process. These frameworks are employed to address diverse imaging challenges arising from non-uniform illumination, shadows, noise, and low-light conditions. Rather than relying on multi-stage or alternating procedures, one-stage frameworks integrate decomposition, enhancement, and artifact suppression into an interpretable, often end-to-end trainable, pipeline. Originally applied in palm vein recognition (1605.08154), the principles of ORF have since influenced various domains, including low-light enhancement, dehazing, and intrinsic image decomposition.

1. Theoretical Foundations of Retinex and One-Stage Formulation

The Retinex theory models an observed image as the product of two components: the intrinsic reflectance $R(x, y)$ and the spatially varying illumination $L(x, y)$ ,

$S(x, y) = R(x, y) \cdot L(x, y)$

or, in vectorized formulations for color images:

$S_{\text{RGB}}(x, y) = R_{\text{RGB}}(x, y) \cdot L(x, y)$

The core objective is to recover $R(x, y)$ from observed $S(x, y)$ , compensating for illumination effects (including uneven lighting or shadow), without introducing artifacts or amplifying noise.

A one-stage Retinex-based framework distinguishes itself by performing:

Decomposition into $R$ and $L$ ,
Enhancement (e.g., histogram equalization, illumination correction),
Artifact and noise suppression (e.g., via filtering or learned denoising), within a single, coherent optimization or network architecture, rather than through a sequence of separate or iterative modules (1605.08154, 1906.06690, 2210.05436).

2. Classical Algorithmic Workflow: Palm Vein Extraction Case Study

The archetypal one-stage approach, as presented for palm vein extraction, consists of the following key steps (1605.08154):

Preprocessing and Normalization: The input image $f(x, y)$ is normalized,

$I(x,y) = \frac{f(x, y) - \min f}{\max f - \min f}$

ensuring contrast consistency.

Single Scale Retinex (SSR):
- The image is log-transformed:
$R(x, y) = \log I(x, y) - \log(G(x, y) * I(x, y))$

where $G(x, y)$ is a Gaussian kernel estimating the locally smooth illumination. The parameter $\delta$ in the Gaussian controls the receptive field; values such as $\delta=25$ are empirically selected.
Postprocessing:
- Dynamic range stretching via histogram equalization,
- Denoising (e.g., median filtering),
- Structure extraction by thresholding,
- Morphological cleaning to remove small artifacts.

This workflow, though rooted in classical signal processing, forms the structural template for many later data-driven ORF variants (1605.08154).

3. Modern One-Stage Retinex Networks: Deep and Plug-and-Play Approaches

Recent developments embed Retinex principles directly into neural architectures that jointly learn decomposition and enhancement:

Deep Joint Decomposition & Enhancement: Architectures such as RetinexNet utilize subnetworks for decomposition (Decom-Net) and illumination adjustment (Enhance-Net), trained together with constraints ensuring consistency of reflectance and smoothness of illumination (1808.04560). Losses include:
- $L_{\mathrm{recon}}$ : reconstruction error between synthesized and target images,
- $L_{\mathrm{ir}}$ : penalizing differences in reflectance estimated from low-/normal-light pairs,
- Structure-aware smoothness loss for illumination, mitigating over-smoothing at edges.
Plug-and-Play with CNN Denoisers: Sequential one-stage frameworks estimate illumination (e.g., via meanRGB and ADMM-based TV refinement), then estimate reflectance through pixel-wise division and regularized denoising. Denoising is implemented by deep CNNs or explicit interpretable modules (e.g., unrolled wavelet shrinkage via Soft-AE), with gamma correction applied during recombination (2210.05436). This strategy allows modularity:
- Any state-of-the-art denoiser can be "plugged in" to serve as a reflectance prior.
- Gamma correction is employed for perceptual balancing of brightness and contrast.
Transformer-Based One-Stage Frameworks: State-of-the-art approaches such as Retinexformer further integrate illumination estimation and restoration of corruptions (e.g., noise, color artifacts) within a transformer architecture (2303.06705). Features include:
- Illumination-guided transformer blocks leveraging spatially-varying illumination to direct the modeling of non-local pixel dependencies,
- Efficient attention via techniques such as depthwise convolutions and illumination-prior fusion,
- End-to-end learning from low-light inputs to enhanced outputs.

4. Evaluation Metrics and Performance Tracking

Performance of one-stage Retinex-based frameworks is quantitatively measured using:

Metric	Formula/Description	Significance
Contrast	$C = \frac{1}{MN} \sum_{i,j} \|x(i,j)-\mu\|$ (1605.08154)	Discrimination between structure and background
Entropy	$E = -\sum p_i\log p_i$ (1605.08154)	Information content (higher = more detail)
Definition	Gradient-based sharpness, e.g., mean local gradient	Measure of structural clarity, edge preservation
PSNR	Signal-to-noise ratio ( $10 \log_{10} \frac{MAX^2}{MSE}$ )	Fidelity to ground-truth in enhancement tasks
SSIM	Structural similarity index	Structural integrity of image enhancement
NIQE, BTMQI	No-reference metrics for perceptual quality	Used where reference is unavailable

For example, the SSR-based palm vein method reported 18.4% contrast improvement, 1.07% entropy increase, and 18.8% definition improvement over classical methods (1605.08154). Transformer-based approaches have shown gains of up to 6 dB PSNR on challenging low-light benchmarks (2303.06705).

5. Adaptation to Diverse Application Domains

One-stage Retinex-based frameworks have been adapted to address diverse imaging problems:

Biometric Identification: Enhanced palm vein pattern extraction under shadows and illumination artifacts (1605.08154).
Low-Light Image Enhancement: Deep ORFs jointly learn reflectance/illumination separation and enhancement, enabling robust object detection under poor lighting (1808.04560, 2303.06705).
Image Dehazing: A straightforward inversion strategy links Retinex to haze removal:

$\text{Dehazing}(I) = 1 - \text{Retinex}(1 - I)$

yielding performance matching or exceeding conventional methods without explicit estimation of atmospheric parameters (1712.02754).

Color Image Enhancement: Weighted guided filtering with Retinex preserves color fidelity while avoiding halo and hue distortion issues (1812.09930).
Underwater Image Enhancement: Integration of multi-scale Retinex defogging with deep SR networks achieves clarity and color restoration in highly degraded environments, as quantitatively measured by PSNR/SSIM (2410.14285).

6. Practical Implementation Considerations

Computational Efficiency: Plug-and-play designs allow modular insertion of denoisers for flexible trade-offs between speed and quality (2210.05436). Lightweight CNN-based ORFs (2406.09656) and state-space model augmentations (2405.03349) further reduce computational cost, allowing deployment on embedded and mobile devices.
Interpretability: Some one-stage ORFs enhance transparency by unrolling classical optimization (e.g., algorithm unrolling (2202.05972)) or adopting interpretable denoisers. Fused-attention mechanisms (2405.03349) improve interpretability of the enhancement process over traditional self-attention modules.
Automation and Adaptability: Self-supervised fine-tuning strategies adapt the enhancement pipeline per-image without paired data (2202.05972), making ORF solutions attractive for large-scale practical applications.
Limitations: Careful loss design is critical to avoid over-smoothing or artifact amplification, and the selection of priors (statistical, structural, or learned) determines the generalizability of the enhancement.

7. Impact, Comparative Outcomes, and Continuing Developments

Empirical benchmarking demonstrates that one-stage Retinex-based frameworks, particularly those coupling classic Retinex modeling with advanced deep learning or state space models, outperform traditional multi-module methods across standard testbeds for low-light, biometric, and underwater imagery (1605.08154, 2303.06705, 2410.14285). RSEND, for example, achieves higher PSNR and SSIM than transformer-based competitors with a fraction of the parameter count (2406.09656), while RetinexMamba introduces efficient interpretability and speed gains over Retinexformer on LOL datasets (2405.03349).

Ongoing research emphasizes further improvements in efficiency, robustness to extreme lighting artifacts, integration of spatial consistency (channel-, semantic-, and texture-level guidance (2305.08053)), and broader application to domains like medical imaging and surveillance.

In summary, the one-stage Retinex-based framework reconstructs the underlying reflectance of an image by modeling and correcting for illumination in a single, coordinated process, with methodological extensions spanning hand-crafted algorithms, deep neural architectures, and plug-and-play modularity. Its contemporary implementations underpin state-of-the-art performance for a wide variety of imaging applications challenged by complex lighting conditions.