Gradient-Domain Weighted Guided Filter (GDWGIF)

• GDWGIF is an advanced image processing operator that employs spatially adaptive regularization and gradient-based constraints to enhance edge preservation, detail fidelity, and noise suppression.
• It refines the classical guided filter by integrating an edge-aware weighting mechanism that replaces scalar regularization with gradient-domain statistics for robust illumination correction.
• Integration with Retinex-based pipelines demonstrates GDWGIF’s practical effectiveness, achieving superior PSNR, SSIM, and BIQI metrics compared to traditional filters.

The Gradient-Domain Weighted Guided Filter (GDWGIF) is an image processing operator designed to address limitations of classical guided filters—specifically, edge blurring and noise amplification under complex illumination conditions. GDWGIF introduces spatially adaptive regularization and gradient-based constraints for enhanced edge preservation, detail fidelity, and effective noise suppression, while maintaining the linear computational complexity of original guided filtering. Its integration with Retinex-based enhancement pipelines enables simultaneous illumination correction and denoising in practical computer vision applications, as demonstrated in recent frameworks (Tao et al., 9 Dec 2025, &&&1&&&).

1. Mathematical Formulation and Theoretical Basis

GDWGIF generalizes the classical Guided Filter (GIF) by adapting its regularization term and local linear model based on pixel-wise gradient statistics. For standard GIF, given input $q$ and guidance $I$ in a window $\Omega_k$ centered at pixel $k$, the model is

$z_i = a_k I_i + b_k, \quad \forall i \in \Omega_k$

and $(a_k, b_k)$ are obtained by minimizing

$$E_{\rm GIF}(a_k, b_k) = \sum_{i\in\Omega_k} (a_k I_i + b_k - q_i)^2 + \lambda a_k^2$$

with closed-form solution

$$a_k = \frac{\mathrm{Cov}_{\Omega_k}(I, q)}{\mathrm{Var}_{\Omega_k}(I) + \lambda}, \quad b_k = \bar{q}_k - a_k \bar{I}_k$$

GDWGIF introduces two primary changes:

• Edge-aware regularization: Replace scalar $\lambda$ by $\lambda/\hat T_I(k)$, where $\hat T_I(k)$ is large for flat regions and small for edges, computed via gradient-domain statistics.
• Adaptive bias term: Add an edge-driven steering factor $(a_k - \psi_k)^2$ that softly enforces $a_k \to 1$ on strong edges and $a_k \to 0$ on flats.

The cost function becomes

$$E_{\rm GDW}(a_k, b_k) = \sum_{i\in\Omega_k} (a_k I_i + b_k - q_i)^2 + \frac{\lambda}{\hat T_I(k)} (a_k - \psi_k)^2$$

which yields the solution

$$a_k = \frac{\mathrm{Cov}_{\Omega_k}(I, q) + \tfrac{\lambda}{\hat T_I(k)}\psi_k}{\mathrm{Var}_{\Omega_k}(I) + \tfrac{\lambda}{\hat T_I(k)}}, \quad b_k = \bar{q}_k - a_k \bar{I}_k$$

The aggregation over windows produces the output

$$z_i = \frac{1}{|\{k : i \in \Omega_k\}|} \sum_{k : i \in \Omega_k} (a_k I_i + b_k)$$

An analogous formulation employing explicit edge-detection and data-dependent weights is presented in (Wang et al., 2022), confirming robustness and edge fidelity.

2. Edge-Aware Gradient Extraction and Regularization

Edge localization and regularization scaling are central innovations in GDWGIF. Gradient computation proceeds via finite differences or Sobel filtering; local gradient variance $\sigma_{g,\xi}(k)$ is compared to its mean, and pixels are split into weak and strong sets using a threshold $T$ (typically $0.2$ or $1.7 \times$ global gradient mean, depending on implementation).

Wavelet or thresholding operations refine gradients in weak/strong subsets, yielding a composite map $g'(k)$. The edge-aware weight $\chi(k)$ is then assembled from local coefficients of variation: $$\chi(k) = \varphi_{I,3}(k)\ \varphi_{I,\xi}(k)\ g'(k)$$ where $\varphi_{I,r}(k)$ measures gradient variation in windows of radius $r$. The regularization denominator is constructed as

$$\hat T_I(k) = \frac{1}{|\Omega_k|}\sum_{i\in\Omega_k} \frac{\chi(i) + \varepsilon}{\chi(k) + \varepsilon}$$

with a small constant $\varepsilon$ to avoid degeneracy.

The bias term $\psi_k$ involves a logistic transform on $\chi(k)$, guiding $a_k$ adaptively: $$\psi_k = 1 - \frac{1}{1 + \exp[\eta(\chi(k) - \mu_{\chi,\infty})]}, \quad \eta = \frac{4}{\mu_{\chi,\infty} - \min \chi(\cdot)}$$

This structure ensures (1) edge retention near boundaries, (2) strong smoothing in uniform regions, and (3) suppression of halo artifacts at sharp transitions.

3. Algorithmic Pipeline and Pseudocode

The practical algorithm proceeds in the following steps:

1. Gradient Extraction: Compute raw gradient map, segment into weak/strong using variance ratio, apply wavelet thresholding, merge results.
2. Edge Weights Calculation: For each pixel, compute $\chi(k)$, $\psi_k$, and $\hat T_I(k)$ using box-filtered coefficients of variation and local statistics.
3. Local Linear Model Solution: Use box filters to compute local means, variances, and covariances of $I$ and $q$; solve for $a_k, b_k$ as above.
4. Aggregation: Average local linear predictions at each pixel (optionally weighted for edge/flatness if using the data-dependent aggregation (Wang et al., 2022)).

Pseudocode summary:

function z = GDWGIF(q, I, λ, ξ, T=0.2, r=5) g_raw = ∇I σg = localVariance(g_raw, radius=ξ) μσg = mean(σg) classify weak/strong (ρ=|σg/μσg−1|<T) apply wavelet-thresholding,... for each pixel k: χ(k)=φ3*φξ*g′(k) ψ(k)=1−1/(1+exp[η(χ(k)−μχ∞)]) T̂(k)=mean_{i∈Ω_k}[(χ(i)+ε)/(χ(k)+ε)] compute means/covariances via box filters for each k: ak = (cov + (λ/T̂(k))*ψ(k)) / (var + (λ/T̂(k))) bk = mean_q(k)−ak*mean_I(k) z = average_{k: i∈Ω_k}( ak*I(i) + bk ) end

The entire procedure operates in $\mathcal{O}(N)$ time due to efficient box filtering, summed-area tables, and constant-per-pixel overhead.

4. Integration in Retinex-Based Enhancement and Practical Applications

In recent simultaneous enhancement and denoising frameworks (Tao et al., 9 Dec 2025), GDWGIF is embedded in a Retinex pipeline. The principal usage is twofold:

• Illumination Estimation: Initial illumination is set as the channel-wise maximum of the RGB input. Multi-scale GDWGIF is applied (at three window radii), and the per-scale results are fused to extract smooth illumination maps for both the original and inverted image, permitting correction of both under- and overexposed regions.
• Reflection Denoising: Reflectance is computed via $R_k = L_k / (\hat L_k + \tau)$, and GDWGIF is employed again, using the refined illumination as guidance to denoise and sharpen reflectance $R'$. Exposure fusion and linear stretching optimize the final dynamic range.

This inclusion allows for adaptive correction under complex illumination states, with empirical demonstration of enhanced contrast and reduced noise relative to earlier models (Tao et al., 9 Dec 2025).

5. Comparative Performance and Experimental Outcomes

Extensive evaluation (Wang et al., 2022) of GDWGIF against GIF, WGIF, GDGIF, and related filters highlights its superior edge preservation and halo suppression:

Method	PSNR (dB)	SSIM
GIF	25.42	0.9794
WGIF	28.78	0.9899
GDGIF	35.00	0.9976
GDWGIF	37.93	0.9982

Qualitative analysis shows preservation of fine edges and uniformity in flat areas, with no visible halo artifacts. For detail enhancement and denoising, GDWGIF also achieves high BIQI and SSIM scores, with PSNR performance matched to or exceeding previous filters. This suggests GDWGIF is optimal for joint edge preservation and smoothness across diverse imaging tasks.

6. Implementation Guidance and Parameter Choices

Recommended parameter values (Tao et al., 9 Dec 2025, Wang et al., 2022):

• Window radius $\xi = 5$–$16$ (11×11 for enhancement, 9×9 for denoising)
• Regularization $\lambda = 0.2$–$1$
• Gradient threshold $T = 0.2$ (or $1.7×$ global gradient mean)
• Small constant $\varepsilon = (0.001 \mathcal{L})^2$ for stability
• Adaptive window coefficient $r=5$ (optional, for anisotropic windows)
• Aggregation weights $w_{\mathrm{flat}}=0.1$, $w_{\mathrm{edge}}=1.0$
• Gamma correction $\alpha=2$ (if required postprocessing)

Box-filter acceleration, summed-area tables, and straightforward neighbor padding suffice for robust, numerically stable implementation. The algorithm remains single-pass and linear complexity, immediately applicable to real-time and high-resolution imaging workflows.

7. Extensions, Limitations, and Directions

GDWGIF retains the simplicity and speed of classic guided filtering but mitigates its principal artifacts. Limitations include potential sensitivity to gradient-domain noise at extremely low SNR and possible necessity for multi-scale refinement in images with extreme dynamic-range edges. Extensions under active investigation include:

• Video enhancement with temporal-gradient constraints for flicker suppression
• HDR tone-mapping via base/detail layer decomposition
• Joint upsampling/fusion using external high-resolution signals (e.g., IR, depth)

A plausible implication is that GDWGIF offers a general-purpose, computationally tractable solution for edge-aware, noise-resilient image enhancement in diverse computer vision and image processing domains (Tao et al., 9 Dec 2025, Wang et al., 2022).