DehazeSB: Segmentation & SB Dehazing

• DehazeSB is a collection of advanced image dehazing methods leveraging segmentation strategies and Schrödinger Bridge theory to optimize haze removal across varied conditions.
• One variant employs a two-stage process with explicit sky/non-sky segmentation and region-adaptive dehazing, significantly enhancing edge visibility and color fidelity.
• Another technique integrates large-scale segmentation priors and stochastic optimal transport to achieve superior quantitative performance under dense or unpaired haze scenarios.

DehazeSB refers to several distinct, technically advanced image dehazing methods and frameworks whose shared designation arises from segmentation- or Schrödinger Bridge-based approaches. These systems are characterized by either explicit sky-region segmentation, integration of large-scale segmentation priors, or optimal transport via Schrödinger Bridge theory, and have each contributed significant advances in real-world image dehazing under challenging, unpaired, or low-data regimes (Lan et al., 13 Jul 2025, Li et al., 2020, Jin et al., 2023). This entry systematically details all primary DehazeSB variants, their theoretical formulations, algorithmic structures, and quantitative benchmarks.

1. Sky/Non-Sky Segmentation Hybrid (DehazeSB, 2020)

One foundational approach labeled DehazeSB is the two-stage hybrid system combining region segmentation and region-adaptive dehazing (Li et al., 2020). The method centers on decomposing the input image into sky and non-sky regions for specialized treatment.

Algorithmic Pipeline

• Stage I: Sky/Non-Sky Segmentation
  • The input image $I(x)$ is first segmented using edge detection with confidence weighting and MeanShift clustering in the spatial–color domain.
  • Morphological post-processing yields a binary sky mask $M_{\text{sky}}(x)$.
• Stage II: Region-Adaptive Dehazing
  • Non-sky regions: Dehazed by an improved Dark Channel Prior (DCP), which substitutes the local minimum with a local average to mitigate haze overestimation:

  $$\tilde{D}(x) = \frac{1}{|\Omega(x)|} \sum_{y \in \Omega(x)} \min_{c \in \{r,g,b\}} I^c(y)$$

  Raw transmission is:

  $\hat{t}(x) = 1 - \omega \frac{\tilde{D}(x)}{A}$

  with $\omega=0.95$, further refined by guided filtering. Scene radiance is recovered via:

  $$J_{\text{non-sky}}(x) = \frac{I(x)-A}{\max(\hat{t}(x), t_0)} + A$$

  and adjusted for mean brightness. - Sky regions: Processed by an enhanced DehazeNet convolutional neural network comprising multi-scale feature blocks and PReLU activations. - Fusion: Output image is

  $$J(x) = M_{\text{sky}}(x) \cdot J_{\text{sky}}(x) + (1 - M_{\text{sky}}(x)) \cdot J_{\text{non-sky}}(x)$$

Quantitative Performance

The DehazeSB segmentation hybrid achieves higher entropy, improved edge visibility, and reduced saturation artefacts versus He et al. (DCP) and Cai et al. (DehazeNet), with runtime $2-3\,\mathrm{s}$ for $473 \times 283$ images (Li et al., 2020).

Method	Entropy↑	VisRatio↑	AvgGrad↑	Sat%↓	Time (s)
He et al.	5.7130	0.1146	3.6574	32.29	3.15
Cai et al.	—	—	—	—	—
DehazeSB	5.7706	0.2492	4.0763	12.96	2.20

This variant addresses colour distortion in sky regions and improves gradient-based perceptual quality.

2. Segmentation-Boosted Dehazing with Large-Model Priors

A second major line designates DehazeSB as "Dehazing with Segmentation Boost," which exploits large-scale segmentation (SAM) priors for fine-grained, region-aware restoration (Jin et al., 2023).

System Architecture

• Segmentation-Aided Module: The pipeline uses an off-the-shelf, frozen Segment Anything Model (SAM), producing a multi-instance object mask $M_{\text{raw}}$ for each scene.

• Grayscale Coding: A coding mechanism compresses k-way masks into an 8-bit grayscale channel without index loss.

• Channel-Expansion Block: Input $M_{\text{sky}}(x)$0 and coded mask $M_{\text{sky}}(x)$1 are concatenated and projected to 64 feature channels through $M_{\text{sky}}(x)$2 convolution, batch normalization, and ReLU.

• Dehaze Network: A U-Net backbone with multi-scale segmentation/feature fusion:

$M_{\text{sky}}(x)$3

• Optional Pre-Dehaze: Running a preliminary dehazer before segmentation further improves performance under dense haze.

Training and Performance

The method is trained using an L1 reconstruction loss. Training is accelerated (30–40% fewer epochs) with mask priors. Quantitative gains shown below (PSNR/SSIM on NTIRE2020):

Method	PSNR (dB)	SSIM
DehazeNet	18.4	0.62
AOD-Net	18.9	0.64
Dehazeformer	19.8	0.65
C2PNet	20.0	0.66
DehazeSB	20.69	0.659

The performance benefits are most salient under medium or severe haze benchmarks (Jin et al., 2023).

3. Schrödinger Bridge-Based Unpaired Dehazing

The 2025 DehazeSB system is based on Schrödinger Bridge (SB) theory, addressing the fundamental bottleneck of unpaired dehazing via stochastic optimal transport (Lan et al., 13 Jul 2025).

Theoretical Foundation

• Schrödinger Bridge Problem: Given source and target distributions $M_{\text{sky}}(x)$4, $M_{\text{sky}}(x)$5 (hazy and clear), seeks probability law $M_{\text{sky}}(x)$6 minimizing KL divergence to a reference Wiener measure, subject to endpoint marginals:

$M_{\text{sky}}(x)$7

Leading to an entropy-regularized OT problem.

• Neural Transport: Discretization into $M_{\text{sky}}(x)$8 steps; generator $M_{\text{sky}}(x)$9 predicts final clear image given intermediate state, and intermediate samples are drawn conditionally using the SB tractable Gaussian.

Loss Terms

• Adversarial Loss: Patch-level Markovian and CLIP-based discriminators for local and global distribution matching.

• Detail-Preserving Regularization:

  • PatchNCE (contrastive, deep feature) loss.
  • Physical prior (ASM-consistent) reconstruction loss.
  • High-frequency preservations: DFT, SSIM, and Sobel gradient matching.
• Prompt Learning with CLIP: Learns a haze-reflecting embedding $$\tilde{D}(x) = \frac{1}{|\Omega(x)|} \sum_{y \in \Omega(x)} \min_{c \in \{r,g,b\}} I^c(y)$$0, enforcing distance of dehazed outputs from "hazy" semantics.

Training and Results

SB-based DehazeSB achieves state-of-the-art FID, NIQE, and perceptual scores on real-world unpaired datasets; e.g., on OHAZE: PSNR=18.83 dB, SSIM=0.838, VSI=0.961. Ablations confirm each regularization is necessary for optimal balance of detail, contrast, and artifact suppression (Lan et al., 13 Jul 2025).

4. Comparative Evaluations

Cross-publication benchmarks indicate that DehazeSB approaches yield improved or competitive results compared to prior state-of-the-art, particularly under dense or non-uniform haze. Direct comparisons (PSNR/SSIM, FID, NIQE, etc.) are presented in each respective paper, with the methods generally excelling on perceptual structure, entropy, and convergence speed.

DehazeSB Variant	Key Mechanism	Principal Benchmark Leader
Sky segmentation hybrid (Li et al., 2020)	Edge + MeanShift + CNN	Gradient/entropy/fusion
SAM segmentation-boost (Jin et al., 2023)	Large-segmentation prior	PSNR/SSIM/hard fog
Schrödinger Bridge (Lan et al., 13 Jul 2025)	Stochastic optimal transport	FID/MUSIQ/no-ref metrics

5. Current Limitations and Design Insights

Analyses across all three DehazeSB families point to several limitations and open questions:

• Overfitting to Haze Distribution: Learned prompt embeddings (SB variant) may be over-specialized if haze types in training are insufficiently varied (Lan et al., 13 Jul 2025).
• Extreme Haze Cases: All methods exhibit some degradation under extremely dense haze, particularly in fine texture recovery (Lan et al., 13 Jul 2025).
• Architectural Choices: In the segmentation-boosted pipeline, pre-dehaze modules and multi-scale fusion most benefit scenarios with heavy and non-uniform fog (Jin et al., 2023).
• Absence of Fine-Tuning: Off-the-shelf segmentors/SAM can transfer well without additional optimization, contributing to generalization and faster training (Jin et al., 2023).

6. Significance in Dehazing Research

DehazeSB methodologies represent the convergence of domain-aware segmentation, neural optimal transport, and the use of large vision-language priors within efficient or lightweight architectures. Their adoption in both paired and unpaired scenarios, across synthetic and real datasets, illustrates the field's shift toward hybrid, data- and prior-efficient models robust to data gaps and challenging haze conditions (Lan et al., 13 Jul 2025, Jin et al., 2023, Li et al., 2020). Each approach provides modular advances—segmentation fusion, region-adaptive CNNs, stochastic-OT regularization—that inform subsequent design in image restoration and generative modeling.