From Fog Chamber to Aircraft Window: Pixel-Registered Imaging and Synthetic Fine-Tuning Enable Cross-Domain Defogging

Abstract: A deep defogging pipeline pretrained on controlled laboratory fog and fine-tuned with domain-randomized synthetic fog applied to clear outdoor scenes generalizes across a graded sequence of out-of-distribution settings with no target-domain training, from chamber-free free-flowing fog to iPhone video recorded through an aircraft cabin window in flight, an entirely unseen sensor, scene, and optical path. This directly addresses an open transfer limitation reported for real-world binocular defogging. Two design choices support the transfer. First, a single-camera fog imager photographs a flat-panel display through an artificial-fog enclosure with a fixed 114~mm scattering path, producing 5{,}495 pixel-aligned foggy/clear pairs. Exact registration permits a paired Laplacian ratio that predicts per-image restoration quality far better than single-image proxies (Spearman $ρ= 0.632$ versus $0.399$) and supports pixel-exact $L_1$ reconstruction training that avoids adversarial hallucination. Second, the fog-chamber checkpoint is fine-tuned on Mapillary Vistas crops overlaid with on-the-fly randomized synthetic fog spanning a broad range of strengths, spatial variations, airlights, and noise conditions. On a 552-image held-out split, a uniform comparison of 30 restoration backbones places NAFNet at the top (24.33~dB~/~0.7912~SSIM), with a compact alternative within 1.29~dB at 3\% of the parameter count, and a ResNet-50 classifier confirms that the restoration preserves semantic content rather than only pixel-level structure. On unpaired aircraft-window video, NIQE decreases from a mean of 6.22 to 4.97 after fine-tuning, with temporally stable output across full-motion sequences. The same backbone, under paired supervision, also reaches 20.71~dB~/~0.683~SSIM on a non-overlapping O-HAZE/NH-HAZE split (a transferability check rather than a competitive ranking).

• The paper introduces a paired laboratory fog-chamber dataset that achieves exact pixel registration for defogging using a direct L1 pixel loss approach.
• The paper devises a fog-difficulty metric based on the mean-absolute Laplacian ratio and benchmarks 30 restoration backbones, with NAFNet attaining a top PSNR of 24.33 dB.
• The paper applies domain-randomized synthetic fine-tuning on outdoor scenes, demonstrating effective cross-domain generalization to aircraft-window video without further retraining.

Pixel-Registered Cross-Domain Defogging Through Controlled Imaging and Synthetic Fine-Tuning

Introduction

This work addresses a fundamental open problem in single-image defogging: the domain transfer gap between models trained on synthetic or laboratory fog and those required to operate robustly in challenging real-world environments, including variable outdoor fog and through unseen imaging pipelines such as consumer mobile devices. The study introduces a paired-laboratory dataset and a synthetic fine-tuning strategy that together enable a deep defogging network to generalize across a sequence of out-of-distribution conditions without target-domain retraining, specifically achieving perceptual improvements on unpaired aircraft-window video captured in flight.

Methodological Contributions

Exactly Registered Fog-Chamber Dataset

The authors present an optically controlled imaging apparatus: a single-camera setup photographs a flat-panel display through a 114 mm artificial-fog enclosure, generating 5,495 pixel-aligned foggy/clear image pairs. This design enables exact registration—unlike previous binocular datasets that suffer from inter-camera misalignment—and supports direct $L_1$ pixel loss during training, circumventing the adversarial hallucination artifacts often encountered in GAN-based or shift-robust learning regimes.

Paired Fog-Difficulty Metric

Leveraging exact pixel alignment, the paper introduces the paired mean-absolute Laplacian ratio $R_{\nabla^2}$ as a fog-difficulty metric, quantifying high-frequency content retention across paired images. This metric demonstrates a Spearman correlation $\rho = 0.632$ with restoration PSNR on held-out data, outperforming the best single-image proxies by a significant margin and providing a content-normalized measure of per-image restoration challenge.

Backbone Benchmarking and Semantic Assessment

Thirty canonical and recent restoration backbones are benchmarked under identical training protocols on the fog-chamber dataset. NAFNet achieves the highest PSNR (24.33 dB, SSIM 0.7912), with SpecAT S2 providing a highly parameter-efficient alternative (within 1.29 dB, $\sim$3% of NAFNet's parameters). A ResNet-50 classifier is used to assess semantic content recoverability, showing that NAFNet-restored outputs restore a substantial portion of recognizability degraded by fog.

Domain-Randomized Synthetic Fine-Tuning

To drive cross-domain generalization, the fog-chamber-trained NAFNet is fine-tuned using spatially randomized synthetic fog overlays on high-resolution outdoor scenes from Mapillary Vistas. Synthetic fog parameters (strength, airlight, spatial variation, noise, etc.) are sampled from broad distributions, eschewing attempts to explicitly match target-domain statistics. This procedure pushes the model to adapt to varying scattering and illumination conditions without overfitting to a specific reference set.

Empirical Results

Generalization to Chamber-Free and Aircraft-Window Settings

Zero-shot deployment on unpaired outdoor frames captured through free-flowing fog and, crucially, on iPhone video acquired through in-flight aircraft windows, provides strong evidence of practical domain transfer. In the latter scenario, the synthetic fine-tuned model provides perceptually improved, temporally consistent sequences with a mean NIQE reduction from 6.22 to 4.97. This outcome represents a direct resolution to limitations highlighted in prior work, where real-fog-trained models (notably on binocularly-captured datasets) failed to generalize across camera hardware and scenes without retraining.

Adaptation on Public Real-Haze Benchmarks

When paired supervision is available (e.g., O-HAZE and NH-HAZE), the same backbone fine-tuned to these datasets achieves strong, though non-competitive, performance (mean 20.71 dB / 0.683 SSIM), validating the approach's adaptability to canonical dehazing challenges. However, the model fine-tuned solely on synthetic fog is shown to over-correct in these paired real-haze settings, indicating that synthetic and real-haze regimes remain complementary.

Benchmark Analysis

The comprehensive backbone comparison reveals that direct $L_1$ reconstruction enabled by exact registration decisively outperforms adversarial backbones such as pix2pix in this fully registered paired-imaging regime. The benchmark establishes a clear accuracy--efficiency Pareto frontier and demonstrates systematic improvements over classical priors, which fail due to violation of underlying assumptions (e.g., the back-illuminated display invalidates the dark channel prior).

Theoretical and Practical Implications

This study provides formal evidence that pixel-registered laboratory data, when coupled with domain-randomized synthetic augmentation, can overcome the historical lack of cross-hardware and cross-scene transferability in defogging. The paired Laplacian difficulty metric offers a robust, interpretable difficulty estimator for benchmarking and calibration. The release of the dataset, simulator, and codebase directly supports reproducibility and onward research.

Practically, the findings designate NAFNet and SpecAT S2 as the current most effective choices—balancing considerations of accuracy and model complexity—for application on fog-impaired imagery where paired training data is attainable. For out-of-distribution scenarios, synthetic fine-tuning demonstrates strong performance, but, as noted, the absence of ground-truth in real-world evaluation sets (such as aircraft-window video) constrains quantitative assessment to no-reference image quality metrics.

Limitations and Future Directions

The dataset is limited to two-dimensional display imagery (not real 3D outdoor scenes), and system calibration (radiometric or geometric) is incomplete. While the synthetic pipeline compensates for some realism limitations, acquiring aligned cross-hardware video benchmarks under real fog remains an open task. Future work should focus on dataset expansion to real dynamic scenes with synchronized cross-domain ground-truth, radiometric calibration, and further exploration of backbone architectures in this exactly registered paradigm.

Conclusion

This work demonstrates that the combined use of pixel-registered fog-chamber imaging and domain-randomized synthetic fine-tuning produces deep models that generalize effectively to egregiously out-of-distribution settings without target-domain data. The approach directly addresses domain transfer limitations identified in real-world binocular defogging, establishes new empirical and methodological benchmarks for the field, and provides a resource base for ongoing advancement of transferable defogging techniques.