Scientific Image Dehazing Benchmarks

• Scientific image dehazing benchmarks are standardized protocols that use unpaired datasets and physics-informed models to assess image restoration in biomedical and microscopic applications.
• They evaluate unsupervised dehazing methods, such as the Equivariant Image Dehazing (EID) framework, by integrating haze consistency and equivariance constraints for robust performance.
• The benchmarks employ quantitative metrics like NIQE, BRISQUE, FID, PSNR, and SSIM to ensure that dehazed images maintain naturalness and structural integrity despite real-world imaging challenges.

Scientific image dehazing benchmarks establish standardized protocols, datasets, and evaluation criteria for removing haze from images in scientific contexts, such as cellular microscopy and medical endoscopy. These benchmarks enable rigorous assessment of dehazing algorithms where acquisition of ground-truth haze-free images is often unfeasible. The recent introduction of the Equivariant Image Dehazing (EID) framework exemplifies fully unsupervised, physics-informed approaches that advance state-of-the-art performance without reliance on paired datasets (Wen et al., 20 Jan 2026).

1. Formulation of the Scientific Image Dehazing Problem

Image dehazing (ID) is the process of reconstructing a clear image $x$ from a hazy observation $y$, with the haze model denoted as $\mathcal{H}$. The general formulation is $y = \mathcal{H}(x)$, where $\mathcal{H}$ may encapsulate complex, unknown physical processes distinct from standard atmospheric scattering models. In scientific imaging modalities—such as fluorescence microscopy and endoscopy—the underlying parameters (e.g., attenuation $\beta$, depth $d(x)$, atmospheric light $A$) are unknown or intractable to estimate. This precludes direct application of classical prior-based or supervised dehazing protocols, necessitating physics-informed, data-driven solutions. A benchmarked method must thus support restoration via maximization of $p(x|y)$, or equivalently:

$$\min_x \left\{ \ell(\mathcal{H}(x), y) + \mathcal{R}(x) \right\}$$

where $\ell$ enforces model consistency and $\mathcal{R}$ embeds desirable priors.

2. Benchmark Datasets and Domain-Specific Challenges

Recent evaluations in scientific dehazing utilize curated benchmarks that reflect domain constraints:

Benchmark Name	Imaging Modality	Key Characteristics
Cholec80-Haze	Medical endoscopy	Real, heavily hazed, unpaired
Cell97	Cellular fluorescence microscopy	High noise, fine structures
RESIDE-OTS/HSTS	Natural scenes (for comparison)	Standard outdoor haze

In these benchmarks, paired haze/clean samples are absent. Instead, large sets of unpaired images of both types are provided to enable unsupervised training and quantitative evaluation. This design reflects the challenge of collecting haze-free ground truth in biomedical and microscope environments. A plausible implication is that advances must center on leveraging indirect supervision and domain-invariant constraints.

3. Unsupervised, Physics-Informed Frameworks

The EID paradigm represents an overview of self-supervision, data-driven pseudo-physics, and equivariant learning:

• Pseudo-haze operator ($G_h$): A differentiable surrogate for the unknown haze process, trained adversarially and with cycle-consistency objectives to map unpaired clear images to synthetic hazy images. Training does not require paired data; CycleGAN-style losses are used:
  • Adversarial loss: Encourages $G_h(x)$ to be indistinguishable from real hazy images by discriminator $D_h$.
  • Cycle-consistency loss: Enforces round-trip structural preservation.
• Dehazing network ($f_\theta$): Typically a U-Net trained under two complementary constraints:
  • Haze consistency: Enforces $\mathcal{H}(f_\theta(y)) \approx y$ via $\mathcal{L}_{hc}$.
  • Equivariance: Ensures $$f_\theta(\mathcal{H}(T_g f_\theta(y))) \approx T_g f_\theta(y)$$ for transformations $T_g$ (e.g., rotations), via $\mathcal{L}_{ec}$.

Training proceeds by first freezing the learned pseudo-haze module, then optimizing $f_\theta$ on raw hazy images using a total loss:

$$\mathcal{L}_{tot}(\theta) = \mathbb{E}_{y,g} \left[ \| \mathcal{H}(f_\theta(y)) - y \|^2 + \lambda \| f_\theta(\mathcal{H}(T_g f_\theta(y))) - T_g f_\theta(y) \|^2 \right]$$

with hyperparameter $\lambda$ set to $0.1$.

4. Evaluation Metrics and Benchmarking Protocols

Benchmarks employ quantitative image quality metrics, both reference-free and reference-based, tailored to the absence of ground truth in scientific imaging:

Metric	Interpretation	Preferred Direction
NIQE	Naturalness Image Quality Evaluator	Lower is better
BRISQUE	Blind/Referenceless Image Spatial Quality	Lower is better
FID	Fréchet Inception Distance	Lower is better
PSNR	Peak Signal-to-Noise Ratio	Higher is better
SSIM	Structural Similarity Index	Higher is better

EID reported state-of-the-art performance across these benchmarks, e.g. on Cholec80-Haze, NIQE=$3.10$, BRISQUE=$4.06$, FID$\approx57.69$; on Cell97, NIQE=$10.68$, BRISQUE=$44.12$, FID$\approx387.74$; on RESIDE-OTS/HSTS, PSNR=$25.18$/$24.15$ dB, SSIM=$0.919$/$0.921$ (Wen et al., 20 Jan 2026).

Qualitative evaluation confirms that fine structural details (e.g., cellular nuclei, mucosal surfaces) and natural color/contrast are preserved in the absence of paired data.

5. Loss Ablation and Transformation Analysis

Ablation studies on benchmark datasets reveal the necessity of combining both haze consistency and equivariance:

• Using $\mathcal{L}_{hc}$ only: NIQE$\approx5.54$ (Cholec80).
• Using $\mathcal{L}_{ec}$ only: NIQE$\approx6.08$.
• Combined: Best result, NIQE$=3.10$.

Regarding group transformations, rotation alone furnished the strongest equivariance supervision; alternative or compound transformations (shift, scale, affine, pan-tilt-rotate) delivered marginally inferior results.

6. Significance and Implications for Benchmark Design

These scientific dehazing benchmarks highlight three core implications:

• Unsupervised frameworks—specifically, those merging physics modelling and group-equivariant constraints—enable effective restoration where ground truth is inaccessible.
• Benchmark curation must reflect the real data limitations of scientific imaging, emphasizing unpaired sets and challenge-matched evaluation metrics.
• The demonstrated performance of EID on both biomedical and natural scene benchmarks suggests such strategies generalize across modalities, enabling cross-domain assessment. A plausible implication is accelerated development of post-processing pipelines in scientific visualization and diagnostics.

By measuring progress through rigorous benchmarks constructed around unpaired, real-world data, scientific image dehazing establishes a foundation for reproducible, domain-agnostic, and state-of-the-art algorithmic advancement (Wen et al., 20 Jan 2026).