Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling

arXiv:2606.25732 · cs.CV · Submitted 2026-06-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling".

Jane: Here's the gist of the paper. The gist The proposed PGL-Net is a lightweight framework that incorporates physical inductive biases via operator-level emulation to achieve state-of-the-art real-world image dehazing with significantly reduced computational cost.

Tom: How it works PGL-Net decouples the dehazing process into two main components: global distribution rectification and local structural refinement, which is achieved through a U-shaped architecture.

Jane: The framework embeds physical inductive biases by preserving the inverse Atmospheric Scattering Model operator form, implementing it as feature-space affine rectification rather than explicit parameter estimation.

Tom: The global distribution rectification is handled by the Physics-Inspired Affine Fusion (PAF) module.

Title and authors: Tom: So we're looking at the paper titled "Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling." I mean, it sounds a bit technical, but the big idea is that they've created this lightweight framework for dehazing that uses physical rules to guide the learning process.

Jane: It does sound heavy on the physics part, Tom. For someone just driving or walking, what does that actually mean? Is this going to be some massive model we need a supercomputer for, or something practical?

Lu: What they're aiming for is exactly the opposite of that. They are proposing PGL-Net as a lightweight framework that incorporates physical inductive biases through operator-level emulation. This means instead of having to explicitly estimate parameters like the transmission or atmospheric light, they are preserving the form of the inverse Atmospheric Scattering Model and implementing it as feature-space affine rectification >

Meng: Operator-level emulation sounds promising if we can keep the complexity low. I mean, existing approaches either use fragile physical inversions under simple assumptions or they adopt heavy blind architectures that just don't work for edge deployment >

Lalam: The core concept here is decoupling the dehazing process into global distribution rectification and local structural refinement. They do this using a U-shaped architecture where a Physics-Inspired Affine Fusion module handles the global part, and a Degradation-Aware Modulation block handles the local part >

The paper's summary: Tom: Exactly. So, they are breaking it down into two main components. First, there’s the global distribution rectification handled by something called the Physics-Inspired Affine Fusion module. This module interprets dehazing as an affine rectification where a scale factor gamma compensates for attenuation and a shift factor beta accounts for airlight-induced bias >

Jane: That sounds like they are using those physical concepts to fix the big, overall haze problem first, before we worry about the fine details. How do they figure out those scale and shift factors? Is it just guessing, or is there some learning involved?

Lu: It involves extracting global descriptors from both the decoder and encoder features projecting them into a joint representation z, and then splitting that representation to yield the channel-wise affine parameters gamma and beta >

Meng: So they are using the network's own features to figure out how much to stretch or shift the image globally. That sounds like it could be efficient if those feature extractions aren't too costly >

Lalam: And then for the local part, they use a Degradation-Aware Modulation block. This block performs residual feature recalibration using stable normalization and lightweight modulation to handle things that are spatially and spectrally variant after the global step >

The paper's improvements: Tom: Looking at how they improved this, the training objective is pretty interesting too. They use a composite loss function. One term just enforces pixel-wise reconstruction of the clean image, and the second term constrains spectral consistency by penalizing both the real and imaginary components of delta F >

Jane: Penalizing both real and imaginary components sounds smart for keeping things stable in terms of color and tone. It stops the model from messing up the spectral information when it tries to reconstruct the scene >

Lu: The improvements really focus on making this decoupled approach work well. They show that combining both modules yields the best performance with minimal overhead, which is key since they are trying to avoid explicit parameter estimation >

Meng: If it keeps the overhead low, that's what I care about for deployment. I need to know if these physical biases actually help them learn better than just a standard blind model on real data >

Lalam: The experimental results show that this works well. For instance, the Tiny variant, PGL-Net-T, improves PSNR by up to two point six dB and consistently enhances downstream object detection accuracy >

Conclusion: Tom: So, wrapping it up on "Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling," the main point is that they achieve state-of-the-art restoration quality while significantly reducing computational cost. They manage to integrate physical priors implicitly >

Jane: It seems like this paper provides a practical trade-off for real world dehazing. It successfully integrates those physical priors without needing heavy, explicit parameter estimation, which is what makes it suitable for edge deployment >

Lu: The mechanism analysis of the PAF module shows that the learned affine factors actually exhibit meaningful degradation-aware patterns. Specifically, the bias term beta is predominantly negative across stages and becomes progressively more negative with depth >

Meng: From a deployment standpoint, this is good because latency evaluations show PGL-Net-T runs at only fourteen point one seven milliseconds on representative computing devices. That's a substantial reduction compared to recent models like SGDN, showing over a ten times reduction in inference latency >

Lalam: And they also found that the framework is robust to spatially varying haze distributions and preserves structures in high-resolution images, which is great for real-world scenes >

Yifei Qu, Ru Li, Junjie Chen, Jinyuan Wu

Harbin Institute of Technology

cs.CV

Submitted: 2026-06-24

Updated: 2026-10-21

Comments: Accepted by WACV 2027

Code: https://github.com/sc-30-bit/PGL-Net

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

The gist: The gist The proposed PGL-Net is a lightweight framework that incorporates physical inductive biases via operator-level emulation to achieve state-of-the-art real-world image dehazing with

Key concepts

Physics-Inspired Affine Fusion (PAF)
This module interprets image dehazing as an affine rectification process. It calculates global scale ($\gamma$) and shift ($\beta$) factors by extracting descriptors from the network features. This physically grounded transformation compensates for haze attenuation and airlight bias, aligning the global distribution of the image effectively.
Degradation-Aware Modulation (DAM)
This block handles local structural refinement using residual feature recalibration. It decomposes features into a value branch for structural cues and a modulation map prediction pathway. This allows the network to adaptively modulate local features, effectively correcting spatially variant degradation after the global step.
Joint Fidelity and Spectral Consistency Training
The model uses a composite loss function that combines pixel-wise reconstruction error with constraints on spectral consistency. By penalizing both real and imaginary components of the change in features ($\Delta F$), it ensures high spatial fidelity while preventing phase wrapping ambiguities, leading to more realistic color restoration.
Edge Deployment Efficiency
PGL-Net is designed for practical deployment by achieving over a 10x reduction in inference latency compared to existing models. Its lightweight architecture allows it to run at real-time performance on resource-constrained devices, such as running at only 14.17 ms, making high-quality dehazing feasible on edge platforms.

Terminology

Summary

The gist The proposed PGL-Net is a lightweight framework that incorporates physical inductive biases via operator-level emulation to achieve state-of-the-art real-world image dehazing with significantly reduced computational cost.

How it works

PGL-Net decouples the dehazing process into two main components: global distribution rectification and local structural refinement, which is achieved through a U-shaped architecture. The framework embeds physical inductive biases by preserving the inverse Atmospheric Scattering Model operator form, implementing it as feature-space affine rectification rather than explicit parameter estimation.

The global distribution rectification is handled by the Physics-Inspired Affine Fusion (PAF) module. This module interprets dehazing as an affine rectification where the scale factor γ compensates for attenuation and the shift factor β accounts for airlight-induced bias. PAF calculates these parameters by extracting global descriptors from decoder and encoder features, projecting them into a joint representation z, and then splitting them to yield the channelwise affine parameters γ, β

How it works

The local structural refinement is managed by the Degradation-Aware Modulation (DAM) block. The DAM block operates to perform residual feature recalibration using stable normalization and lightweight modulation. It decomposes the normalized features into two branches: a value branch extracts local structural cues using a depth-wise convolution, and a compact bottleneck pathway predicts a degradation-aware modulation map M. Finally, the local features are adaptively modulated and aggregated in a residual manner: F out = DSConv(V ⊙ M) + F in.

How it works

The training objective enforces both spatial fidelity and spectral consistency through a composite loss function. The first term enforces pixel-wise reconstruction, while the second term constrains spectral consistency by penalizing the real and imaginary components of ΔF.

How it works

The performance evaluation across multiple benchmarks demonstrates that PGL-Net achieves state-of-the-art restoration quality with significantly reduced complexity. Specifically, the Tiny variant, PGL-Net-T, improves PSNR by up to 2.6 dB and consistently enhances downstream object detection accuracy.

How it works

The model is designed for practical deployment by achieving a substantial reduction in inference latency. For instance, PGL-Net-T runs at only 14.17 ms on representative computing devices. This efficiency is validated by the fact that PGL-Net achieves over a 10× reduction in inference latency compared to the recent SOTA SGDN.

How it works

The mechanism analysis of PAF shows that the learned affine factors exhibit meaningful degradation-aware patterns. Specifically, the bias term β is predominantly negative (> 90% across stages) and becomes progressively more negative with depth 90% across stages) and becomes progressively more negative with depth (−0.30 → −0.57 → −1.31)>

How it works

The ablation study confirms the complementary nature of the proposed modules. Combining both modules yields the best performance with minimal overhead.

How it works

PGL-Net is well-suited for practical edge deployment because its latency evaluations show it runs at real-time performance across all devices

How it works

Despite its success, limitations exist when the input image contains extremely dense haze or nighttime scenes with colored local light sources. These conditions may cause the feature-space affine modulation to fail in fully recovering fine structures or introduce color bias.

How it works

In conclusion, PGL-Net provides a more practical quality–efficiency trade-off for unpaired real-world dehazing. The framework successfully integrates physical priors implicitly while maintaining high efficiency for real-time edge deployment>

How it works

The proposed method achieves superior restoration performance across five paired real-world benchmarks while maintaining substantially lower computational complexity. For instance, on the RUDB benchmark, PGL-Net-T improves PSNR over SGDN by 2.57 dB. The results validate the practical utility of PGL-Net in real-world scenarios>

How it works

The framework is robust to spatially varying haze distributions and preserves structures in high-resolution images. This capability is further supported by the qualitative evaluation showing that PGL-Net successfully penetrates dense haze and restores faithful colors and sharp structures closest to the ground truth <ref:2606.

Improvements for AI systems

  1. Decoupled Dehazing for Edge Deployment: The PGL-Net framework can perform state-of-the-art image dehazing with over a 10 times reduction in inference latency compared to recent SOTA SGDN models, making it suitable for real-time deployment on resource constrained edge platforms.

  2. Physics Inspired Affine Fusion (PAF) for Feature Alignment: The PAF module performs globally conditioned channel wise affine rectification across skip connections, mitigating the haze induced distribution mismatch between encoder and decoder features by using a physically grounded transformation derived from the inverse atmospheric scattering model.

  3. Degradation Aware Modulation (DAM) for Local Refinement: The DAM block adaptively restores spatially and spectrally variant details through dynamic feature modulation, allowing PGL-Net to handle local haze variations effectively after global rectification.

  4. Joint Fidelity and Spectral Consistency Training: Training with the composite loss L = J hat - J 1 + lambda (Re(Delta F) 1 + Im(Delta F) 1) jointly enforces pixel wise reconstruction and constrains spectral consistency via frequency domain discrepancy, avoiding phase wrapping ambiguity.

  5. Enhanced Downstream Object Detection: The PGL-Net models consistently improve downstream object detection accuracy, with the Tiny variant (PGL-Net-T) improving PSNR by up to 2.6 dB and achieving over a 10 times reduction in inference latency compared to SGDN.

  6. Robust No Reference Perception: For unpaired real world benchmarks, PGL-Net-T achieves the best NIQE on RTTS and URHI, indicating favorable perceptual naturalness under real haze, while also maintaining low inference latency of 14.17 ms.

Sources

Related papers