Backbone-Agnostic Stochastic Perturbation Learning for End-to-End Real-World Image Dehazing

Bingcai Wei, Yuning Cui, Mingyu Liu, Jinni Geng, Ling Li, Benwang Chen, Ziwei Li, Alois KnollarXiv 2026 · arXiv

TL;DR

The paper introduces Backbone-Agnostic Stochastic Perturbation Learning (BSPL), a plug-and-play framework that enhances real-world image dehazing across various network architectures (CNNs, Transformers, and Mambas). By actively probing feature sensitivity using learnable stochastic perturbations and enforcing bidirectional consistency (forward dehazing and backward haze reconstruction), BSPL achieves substantial image quality improvements with virtually zero additional inference overhead.

What Problem Does This Paper Address?

Real-world image dehazing is notoriously difficult because haze is spatially non-uniform, highly dependent on illumination, and physically ambiguous. Most state-of-the-art models attempt to learn a direct, deterministic mapping from a hazy image to a clean target.

However, this approach suffers from two major drawbacks:

  1. Ignoring Feature Sensitivity: Standard networks treat small feature disturbances or noise as nuisances to suppress, missing out on the valuable information that feature sensitivity can reveal about degradation.
  2. Lack of Physical Constraints: Purely data-driven models rarely verify whether their restored "clean" image can logically reconstruct the original hazy input under physical laws, often leading to over-smoothed textures or severe color shifts.

Methods

BSPL addresses these gaps by introducing a modular framework that can be inserted into any existing encoder-decoder restoration backbone without altering its core topology.

Figure 1: Overall integration of BSPL into a hierarchi- cal encoder–decoder restoration network. The green Block modules denote the backbone network into which BSPL is embedded; LSPM units are interleaved with these original blocks without changing the backbone topology.

The framework consists of three main components:

  1. Learnable Stochastic Perturbation Modulator (LSPM): Interleaved between backbone blocks, LSPM samples channel-wise and spatial-wise random perturbations from an input-conditioned Gaussian distribution. By measuring the "response discrepancy" (how much the features change before and after the perturbation), it identifies which channels and spatial locations are most sensitive to haze and converts these discrepancies into adaptive attention weights.
Figure 2: Illustration of LSPM. It estimates channel-wise and spatial-wise perturbation distributions, measures feature-response discrepancies under sampled perturbations, and converts them into modulation weights for feature enhancement.
  1. Prior-informed Perturbation-guided Reconstruction Module (PPRM): Active only during training, PPRM enforces a physics-inspired constraint. It takes the restored clean image and attempts to reconstruct the original hazy observation using the classical atmospheric scattering model, guided by Dark Channel Prior (DCP) estimates and the same perturbations learned by LSPM.
Figure 3: Illustration of PPRM. It uses transmission and atmospheric-light priors together to reconstruct the hazy ob- servation through a physics-inspired formulation. AT denotes an affine transformation. I-LSPM denotes an inherited LSPM that receives Nc and Ns from the backbone bottleneck rather than learning independent perturbations.
  1. Dual-space Domain-diversified Distribution-aware Contrastive Loss ($D^3$CL): This loss regularizes both the clean restoration space and the hazy reconstruction space. It leverages both real-world and synthetic negative samples to prevent the network from collapsing into a simple paired reconstruction mapping.
Figure 4: Illustration of D3CL. Contrastive regularization is performed in both clean restoration and hazy reconstruction spaces, with real-world and synthetic negatives forming di- versified degradation domains.

Inference Efficiency: At inference time, PPRM and $D^3$CL are completely discarded. The random seed for LSPM is fixed to ensure deterministic evaluation. As a result, the model gains the benefits of physics-inspired training without slowing down deployment.

Datasets

The authors evaluated BSPL across five diverse, real-world paired image dehazing benchmarks:

  • Dense-Haze: Characterized by dense, homogeneous haze.
  • I-HAZE / O-HAZE: Focused on indoor and outdoor real-world haze, respectively.
  • NH-HAZE: Features highly non-homogeneous, spatially varying haze.
  • LMHaze: A large-scale dataset covering multiple haze intensities across large scenes.

The training set aggregated 4,065 paired images from these sets, and the synthetic Haze4K dataset was utilized specifically to provide synthetic negative samples for the contrastive loss.

Key Insights and Ideas

  • Noise as a Diagnostic Tool: Instead of treating noise as an enemy, BSPL uses learnable stochastic perturbations as a probe. Features that fluctuate wildly under perturbation are flagged as highly ambiguous, haze-affected regions that require adaptive restoration attention.
  • Bidirectional Consistency: True image restoration shouldn't just be a one-way street. Forcing the network to successfully reverse-engineer the haze degradation path ensures that the restored image preserves structural and physical validity.
  • Dual-Space Contrastive Boundaries: Applying contrastive learning simultaneously to both the clean and hazy domains—while mixing real and synthetic negatives—enables a much more robust boundary definition for what constitutes a "clean" vs. "degraded" feature representation.

Key Results

The authors integrated BSPL into five distinct baseline architectures: MIMOUNet (CNN), IRNeXt (CNN), $C^2$SSM (Mamba), HOGformer-S (Transformer), and VIVNet (Hybrid).

BSPL consistently delivered performance gains across all models and benchmarks:

  • Mamba-based $C^2$SSM saw the highest average improvement, leaping by 2.51 dB PSNR and 0.115 SSIM.
  • HOGformer-S improved by 1.49 dB, while MIMOUNet and VIVNet improved by 1.13 dB and 1.10 dB, respectively.
  • Qualitative Improvements: Absolute error maps demonstrate that BSPL significantly reduces color distortion and residual haze, particularly around complex haze boundaries and fine textures, outperforming the standalone baselines.

Limitations and Future Work

While the paper demonstrates robust results, a few limitations and future directions are apparent:

  • Dependence on Hand-Crafted Priors: PPRM relies on Dark Channel Prior (DCP) estimates to guide its reconstruction branches. In scenes where DCP naturally fails (e.g., highly reflective surfaces or large white objects), the reconstruction guidance might be suboptimal.
  • Extension to Other Tasks: The authors note that probing degradation-sensitive features via stochastic perturbations is a general concept. Future work could explore extending BSPL to other low-level vision challenges, such as low-light image enhancement, motion deblurring, or raindrops removal.

Summary updated

Abstract

Real-world paired image dehazing remains challenging because haze degradation is spatially non-uniform, illumination-dependent, and physically ambiguous even when haze-free references are available. Existing end-to-end restoration networks usually learn a deterministic mapping from a hazy observation to a clean target, while degradation-sensitive feature responses, reverse haze-formation consistency, and cross-domain negative structure remain insufficiently exploited. In this paper, we propose Backbone-Agnostic Stochastic Perturbation Learning (BSPL), a plug-and-play framework for end-to-end real-world image dehazing. BSPL first introduces a Learnable Stochastic Perturbation Modulator (LSPM), which learns input-conditioned channel-wise and spatial-wise perturbation distributions and converts the resulting feature-response discrepancies into adaptive modulation weights. It then develops a Prior-informed Perturbation-guided Reconstruction Module (PPRM), which reuses the learned bottleneck perturbations together with transmission and atmospheric-light priors to reconstruct the hazy observation from the restored result and enforce degradation consistency. Furthermore, we propose a Dual-space Domain-diversified Distribution-aware Contrastive Loss ($D^3$CL) to regularize both clean restoration and hazy reconstruction spaces with real-world and synthetic negatives. Experiments on five real-world paired benchmarks show that BSPL consistently improves multiple representative backbones with only marginal additional inference overhead.
BibTeX
@misc{wei2026backbone,
  author = {Bingcai Wei and Yuning Cui and Mingyu Liu and Jinni Geng and Ling Li and Benwang Chen and Ziwei Li and Alois Knoll},
  title = {{Backbone-Agnostic Stochastic Perturbation Learning for End-to-End Real-World Image Dehazing}},
  year = {2026},
  eprint = {2607.11623},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.11623},
}
Figure