Efficient Dual-domain Image Dehazing with Haze Prior Perception

The paper proposes DGFDNet, a dual-domain dehazing network that integrates a Dark Channel Guided Frequency-aware Dehazing Block and a Prior Correction Guidance Branch to effectively align spatial and frequency features for achieving state-of-the-art performance with improved robustness and efficiency.

The Gist

Imagine you are trying to take a beautiful photo of a city skyline, but a thick, gray fog has rolled in. Everything looks washed out, colors are dull, and details are hidden. This is what "hazy images" look like to a computer. For a long time, computers have struggled to "clean" these photos because the fog isn't just a blanket; it's a tricky, uneven layer that changes how light behaves.

This paper introduces a new AI system called DGFDNet (Dark Channel Guided Frequency-aware Dehazing Network) that acts like a super-smart photo editor. It doesn't just try to wipe the fog away; it understands how the fog distorts the image and fixes it from two different angles at once.

Here is how it works, explained with simple analogies:

1. The Two-Pronged Approach: Looking at the "Big Picture" and the "Fine Print"

Most old methods tried to fix the image by looking at it like a regular photograph (the Spatial Domain). They looked at pixels next to each other. But fog affects the whole image at once, so looking just at neighbors wasn't enough.

Other methods tried to look at the image as a sound wave or a radio signal (the Frequency Domain). Think of an image as a song. The "low notes" are the big shapes and colors, while the "high notes" are the sharp edges and tiny details. Fog mostly messes up the "low notes" (making the image dull) but leaves the "high notes" (the structure) mostly intact.

DGFDNet's Superpower: It looks at the photo in both ways simultaneously. It's like having a detective who listens to the whole song and reads the sheet music at the same time. By combining these two views, it knows exactly what to fix without accidentally blurring the sharp edges.

2. The "Haze Detector" (The Dark Channel Prior)

To know where the fog is, the system uses a trick called the "Dark Channel Prior."

• The Analogy: Imagine you are looking at a forest. Even in a foggy forest, if you look at a tiny patch of leaves, at least one color (red, green, or blue) is usually very dark. But if there is fog, that patch turns gray and bright.
• The Problem: This trick works great in forests but fails in the sky. The sky is naturally bright and blue, so the computer thinks the sky is "foggy" when it's actually just clear blue.
• The Fix: DGFDNet uses a special branch called PCGB (Prior Correction Guidance Branch). Think of this as a smart editor who double-checks the initial guess. If the computer thinks the sky is foggy, this editor says, "Wait, that's just the sky! Don't touch that." It constantly corrects the map of where the fog actually is, especially in tricky outdoor scenes.

3. The "Frequency Tuner" (HAFM)

Once the system knows where the fog is, it needs to clean it.

• The Analogy: Imagine the image is a radio signal full of static. The Haze-Aware Frequency Modulator (HAFM) is like a high-tech radio tuner. It doesn't just turn the volume down; it specifically targets the "static frequencies" caused by the fog and filters them out, while keeping the clear signal (the actual image) loud and crisp.
• Because it uses the "Haze Detector" map, it knows exactly how much to tune for each part of the image. It's not a one-size-fits-all fix; it's a custom tune for every pixel.

4. The "Detail Refiner" (MGAM)

Sometimes, after cleaning the fog, the image might look a bit too smooth or lose its texture (like the grain in a photo or the texture of a brick wall).

• The Analogy: This module, called MGAM, is like a master sculptor. After the fog is gone, the sculptor goes in with fine tools to carve out the tiny details that were lost. It uses a "gating" mechanism, which is like a bouncer at a club, deciding which details are important to keep and which noise to throw out. It ensures the final photo looks sharp and realistic, not just "clean."

Why is this a big deal?

• Speed vs. Quality: Usually, you have to choose between a fast method (which makes a mediocre photo) or a slow method (which makes a great photo but takes forever). DGFDNet is like a Formula 1 car that is also a family sedan: it is incredibly fast and efficient (lightweight) but produces the highest quality results.
• Real-World Ready: It works not just on computer-generated fog, but on real photos taken in cities, mountains, and rainy days. It handles the messy, uneven fog that confuses older AI models.

In a nutshell: DGFDNet is a smart, dual-view camera cleaner. It uses a "haze map" to find the fog, a "frequency tuner" to remove the grayness, and a "detail sculptor" to sharpen the edges, all while constantly double-checking its work to make sure it doesn't accidentally erase the sky or the sun. The result is a crystal-clear photo, generated quickly and efficiently.

Technical Summary

1. Problem Statement

Single-image dehazing aims to recover a clear scene from a hazy input, but it remains an ill-posed problem due to unknown and spatially varying haze distributions. Existing methods face three primary limitations:

• CNN-based methods: While effective at local feature extraction, they suffer from limited receptive fields, hindering the modeling of long-range dependencies essential for global structural consistency.
• Transformer-based methods: Although they excel at capturing global context via self-attention, they incur high computational costs (quadratic complexity), making real-time application difficult.
• Dual-domain (Spatial-Frequency) methods: Recent approaches attempt to combine spatial and frequency domains. However, they typically use loosely coupled dual-branch designs that lack explicit alignment of degradation cues between domains. Furthermore, they often rely on static physical priors (like the Dark Channel Prior) which fail in complex outdoor scenes (e.g., sky regions), leading to inaccurate haze localization.

2. Methodology: DGFDNet

The authors propose DGFDNet (Dark Channel Guided Frequency-aware Dehazing Network), a dual-domain framework that explicitly aligns haze degradation cues across spatial and frequency domains. The architecture follows a U-shaped encoder-decoder structure with a Prior Correction Guidance Branch (PCGB).

Core Components

A. DGFDBlock (The Fundamental Unit)
Each stage of the network contains multiple DGFDBlocks, which integrate two key modules:

1. Haze-Aware Frequency Modulator (HAFM):

   • Function: Performs global degradation-aware spectral filtering.
   • Mechanism: It uses the Dark Channel Prior to generate a haze confidence map (spatial attention). This map guides the modulation of frequency components.
   • Process:
     • Spatial Modulation: The input is processed via depth-wise convolutions and modulated by the haze confidence map to highlight degraded regions.
     • Frequency Modulation: The spatially modulated features undergo Fast Fourier Transform (FFT). The real and imaginary components are processed via 1×1 convolutions to adaptively filter haze-related frequencies, then converted back via Inverse FFT (IFFT).
     • Synergy: This alternating process ensures that frequency modulation is explicitly guided by spatial haze localization, avoiding the "loose coupling" of previous methods.

2. Multi-level Gating Aggregation Module (MGAM):

   • Function: Refines local details and recovers high-frequency structures.
   • Mechanism: It employs a hybrid gating mechanism with multi-scale convolutions (using $3\times3$ and $5\times5$ kernels with dilation).
   • Process: Low-level features rich in edges/textures generate gating signals to regulate the flow of high-level semantic information. This allows for adaptive feature fusion, preserving fine-grained details while mitigating the loss of local structures common in global modeling.

B. Prior Correction Guidance Branch (PCGB)

• Problem Addressed: The Dark Channel Prior (DCP) is often inaccurate in outdoor scenes (e.g., sky regions have naturally low intensity, causing false haze estimation).
• Solution: PCGB treats the DCP not as a static input but as a dynamic cue.
• Mechanism:
  • It extracts features from the initial dark channel.
  • It receives feedback from the MGAM in the main dehazing branch.
  • Using SKFusion (a shared fusion block), it iteratively refines the dark channel features.
  • Closed-Loop: This creates a feedback loop where the network corrects the prior's errors based on the refined spatial features, significantly improving haze localization accuracy in complex environments.

C. Loss Function
The model is trained using a Dual-domain L1 Loss, supervising both the spatial domain (pixel-wise difference) and the frequency domain (difference in Fourier magnitude) to ensure both structural consistency and texture recovery.

3. Key Contributions

1. Dual-Domain Alignment: Proposed DGFDNet, the first framework to explicitly align spatial and frequency degradation cues using a dark channel-guided attention mechanism, moving beyond loosely coupled dual-branch designs.
2. HAFM & MGAM Modules: Designed the Haze-Aware Frequency Modulator for global spectral filtering and the Multi-level Gating Aggregation Module for fine-grained detail recovery, balancing global context and local restoration.
3. Iterative Prior Correction: Introduced the PCGB, a closed-loop feedback mechanism that dynamically corrects the Dark Channel Prior during inference. This solves the issue of prior failure in complex outdoor scenes (e.g., sky regions) without requiring additional heavy supervision.
4. Efficiency: Achieved State-of-the-Art (SOTA) performance with a compact design, offering real-time efficiency compared to heavy Transformer-based models.

4. Experimental Results

The authors evaluated DGFDNet on four benchmark datasets: SOTS-Indoor, SOTS-Outdoor (synthetic), and Dense-Haze, NH-HAZE (real-world).

• Quantitative Performance:
  • SOTS-Indoor: Achieved 42.18 PSNR and 0.997 SSIM, outperforming the previous best (DCMPNet) and lightweight models (PGH2Net).
  • SOTS-Outdoor: Achieved 38.51 PSNR and 0.995 SSIM, surpassing DCMPNet by 1.95 dB.
  • Real-world: Set new records on Dense-Haze and NH-HAZE, demonstrating superior generalization to non-synthetic conditions.
• Efficiency:
  • Parameters: Only 2.08M (significantly lower than DCMPNet's 17.36M and Dehamer's 132.5M).
  • FLOPs: 13.65G, which is less than half of many competing CNN/Transformer models (e.g., FFA-Net at 287.8G).
• Ablation Studies:
  • Removing PCGB caused a significant drop in outdoor performance (up to 0.71 dB), validating the necessity of prior correction.
  • Removing either HAFM or MGAM resulted in substantial performance degradation, confirming the synergy between global frequency modeling and local detail aggregation.

5. Significance

This paper addresses a critical bottleneck in image dehazing: the trade-off between global modeling capability (usually requiring heavy Transformers) and computational efficiency. By explicitly aligning spatial and frequency domains and introducing a mechanism to iteratively correct physical priors, DGFDNet achieves:

• Robustness: It handles complex, non-uniform haze conditions (including sky regions) where traditional priors fail.
• Efficiency: It delivers SOTA performance with a lightweight architecture suitable for real-time deployment.
• Generalization: The iterative correction strategy allows the model to adapt to unseen real-world degradation patterns, making it highly practical for autonomous driving, surveillance, and remote sensing applications.