Global Edge Retinex (GER) Model

• Global Edge Retinex (GER) is a decomposition theory that reformulates low-light image formation by integrating explicit edge, reflectance, illumination, noise, and artifact components.
• It improves detail preservation by treating object edges as a first-class element, decoupling their structure from illumination to reduce blurring and feature aliasing.
• Empirical validations show that substituting standard Retinex with GER in DRWKV yields significant gains in PSNR, SSIM, and NIQE, underscoring its impact on edge fidelity.

Global Edge Retinex (GER) is a Retinex-inspired decomposition theory for low-light image enhancement introduced within DRWKV, “DRWKV: Focusing on Object Edges for Low-Light Image Enhancement” (Bai et al., 24 Jul 2025). GER reformulates low-light image formation by making edge structure an explicit component alongside reflectance, illumination, spatially heterogeneous noise, and artifact residuals. In that formulation, the central objective is not only illumination correction but also preservation of object edge continuity and fine structural details under extreme illumination degradation.

1. Conceptual basis and stated motivation

GER is presented as a response to a limitation attributed to standard Retinex formulations: they do not explicitly model hierarchical edge structure and they couple illumination and edges too tightly. The stated consequence of that coupling is degraded edge fidelity, including edge blurring, feature aliasing, misclassification of noise versus edges, and distortion in extremely low-light conditions (Bai et al., 24 Jul 2025).

Within this account, low-light enhancement is framed as a problem in which brightness restoration alone is insufficient. The paper argues that an illumination-centered strategy can amplify background noise and distort edges, while an approach focused only on contours can produce abrupt transitions and underexposure. Under controllable global exposure, the “essence” of low-light enhancement is therefore described as refining object edges. GER is the theoretical mechanism proposed to support that claim by separately representing stable reflectance or content, edge detail, illumination, noise, and artifacts.

The paper’s Hypothesis (1) states that hierarchical detail enhancement depends on a gradient-guided mechanism with illumination invariance. GER is introduced as the decomposition theory that substantiates that hypothesis. A plausible implication is that the framework treats edge preservation not as a secondary regularization target but as part of the image formation model itself.

2. Mathematical formulation

The baseline Retinex model used for contrast is

$I = R \cdot L + N$

where the observed image is modeled through reflectance, illumination, and noise. GER modifies that formulation to

$$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$

with $I$ denoting the observed low-light image, $R$ the reflectance component, $E$ the edge feature term, $L$ the illumination component, $N$ spatially heterogeneous noise, $S$ the artifact residual or structured artifact component, and $\alpha, \beta, \gamma$ learned or set weights balancing the components; $\odot$ denotes element-wise multiplication (Bai et al., 24 Jul 2025).

This decomposition differs from the baseline in four explicit ways. First, it adds the edge term $$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$0. Second, it adds the artifact term $$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$1. Third, it redefines noise as spatially heterogeneous noise. Fourth, it treats edge detail as part of the enhanced reflectance representation rather than leaving it implicit. For that reason, GER is characterized as more than “Retinex plus edge detection”; it is a reformulated decomposition in which edge structure becomes a first-class component.

The paper gives the reflectance estimate as

$$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$2

and describes it as the “clean benchmark” for detail enhancement. In that interpretation, $$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$3 removes spatially heterogeneous noise, division by $$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$4 neutralizes illumination interference, and the result serves as the stable base for subsequent edge-guided enhancement. The enhanced reflectance is then defined as

$$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$5

which is the key edge-preserving step in GER.

The illumination component is estimated in Light Preprocessing using the gray-world assumption:

$$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$6

Artifact suppression is introduced as

$$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$7

and the paper gives the final enhanced-image expression as

$$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$8

while noting, in the extracted account, that the original LaTeX is somewhat malformed. The intended meaning is a reconstruction that combines edge-enhanced reflectance with illumination and then adds controlled noise residual and suppressed artifact residual. This makes GER a multi-component reconstruction framework rather than a simple two-term Retinex model.

3. Placement in the DRWKV pipeline

Within DRWKV, GER is one of two principal theoretical pillars, alongside the Evolving Scanning mechanism. The pipeline is described as having two stages: Light Preprocessing and Deep Detail Mining. GER is especially central to Deep Detail Mining, but it also informs preprocessing (Bai et al., 24 Jul 2025).

In Light Preprocessing, GER is explicitly connected to the estimation of illumination, noise, and artifacts. The preprocessing decomposition is given as

$$I = (R + \alpha \cdot E) \odot L + \beta \cdot N + \gamma \cdot S$$9

followed by reconstruction of a brightness-restored baseline image through

$I$0

This stage is framed as preparing a baseline image before deeper GER-guided detail mining. The paper further states that by combining the gray-world illumination estimate $I$1 with the artifact component $I$2, the method can “explicitly define the interference boundaries to be avoided,” allowing brightness enhancement without amplifying artifacts.

The network-level deployment is described through two Retinex-related blocks. Block1 is constrained by the need to receive the edge feature $I$3. Block2 is the location where GER is deployed for better optimization. This arrangement indicates that GER is not merely an auxiliary prior; it is integrated into the layered optimization strategy of DRWKV and is tied directly to how edge-aware features are routed and refined.

4. Relationship to edge continuity, spectral alignment, and training objectives

GER is presented as addressing the content or decomposition side of edge preservation, while Evolving WKV Attention addresses the spatial and topological side. The paper states that GER specifies what should be decomposed and preserved by making edge information explicit through $I$4, whereas Evolving Scanning and ES-RWKV determine how spatial continuity is modeled after decomposition is established. In the paper’s formulation, Evolving Scanning “transforms the geometric continuity of edge manifolds into temporal continuity capturable by the VRWKV propagation mechanism” (Bai et al., 24 Jul 2025).

GER also interacts with Bi-SAB and MS$I$5-Loss. The conceptual division is explicit: GER provides edge-aware decomposition and reflectance enhancement; Bi-SAB performs spectral alignment between luminance and chrominance; MS$I$6-Loss encourages structural, edge, illumination, and artifact consistency. Bi-SAB is described as including Scharr edge enhancement and cross-attention to fuse low-level and high-level features. This suggests that GER supplies the edge-aware representational substrate, while Bi-SAB constrains the brightened output to remain spectrally natural.

MS$I$7-Loss contains the following terms:

$I$8

$I$9

$R$0

$R$1

$R$2

with total loss

$R$3

The role of this objective is not to define GER directly but to regularize the very components GER introduces: $R$4, $R$5, $R$6, and the balancing coefficients $R$7. In the paper’s interpretation, edge sparsity encourages sharp but sparse true edges, illumination smoothness preserves natural lighting variation, artifact suppression prevents hallucinated structure, and coefficient regularization stabilizes the decomposition.

5. Reported empirical evidence

The paper’s most direct empirical evidence for GER is a block-comparison ablation on LOLv2 Real. In the baseline configuration, Block1 uses Robust Retinex and Block2 also uses Robust Retinex, yielding PSNR 22.57, SSIM 0.741, and NIQE 4.124. In the compared configuration, Block1 remains Robust Retinex while Block2 is replaced by GER, yielding PSNR 24.12, SSIM 0.832, and NIQE 3.926 (Bai et al., 24 Jul 2025).

The reported improvement rates for inserting GER into Block2 are 6.9% PSNR, 12.3% SSIM, and 4.8% NIQE. The paper uses these values as evidence that GER’s gradient-guided global-edge coupling is practically effective for low-light edge restoration. The qualitative effects repeatedly attributed to the resulting system include continuous object edges, high-fidelity details, better visual naturalness, and reduced edge distortion and blurring.

The extraction further states that the empirical validation is strongest in this block ablation. That characterization is significant because it locates the most direct evidence for GER specifically at the point where standard Retinex is replaced by the new decomposition, rather than in broader end-to-end performance alone.

6. Scope, significance, and common simplifications

The paper identifies GER as the first-time integration of its new Global Edge Retinex theory with the VRWKV model. Its stated conceptual contribution is to treat object edges as a global structural factor in Retinex decomposition, decouple illumination from edge structure, preserve continuity of object boundaries, retain fine details under extremely low light, and reduce noise or artifact amplification during brightness restoration (Bai et al., 24 Jul 2025).

A common simplification would be to describe GER as standard Retinex with an added edge detector. The formulation given in the paper is broader than that. GER modifies the decomposition itself, introduces explicit artifact modeling, recasts noise as spatially heterogeneous, and uses edge-enhanced reflectance as a reconstruction primitive. In that sense, the theory is positioned as an image formation and reconstruction model rather than a post hoc edge-refinement module.

Another important clarification concerns its role in the overall system. GER is not the sole edge-preserving mechanism in DRWKV. The paper explicitly distributes responsibilities across components: GER handles edge-aware decomposition and reflectance enhancement; Evolving WKV Attention models continuity efficiently; Bi-SAB aligns brightness and color features while also handling edge details; and MS$R$8-Loss stabilizes the decomposition during training. This suggests that GER is foundational but not exhaustive within the architecture.

In concise form, GER can be characterized as a Retinex extension that reformulates low-light image formation as an illumination-modulated sum of reflectance plus explicit edge detail, noise, and artifact terms, so that DRWKV can enhance brightness while preserving edge continuity and fine structure.