Image Dehazing as In-Context Tabular Inference

Le-Anh Tran

cs.CVOct 18, 2026

Image Dehazing as In-Context Tabular Inference

Authors: Le-Anh Tran

Abstract

Image dehazing is traditionally addressed through handcrafted physical priors or task-specific deep learning models, often requiring carefully designed architectures or extensive model training. Recent advances in tabular foundation models (TFMs), however, offer an alternative learning paradigm in which continuous relationships can be inferred directly from contextual examples without task-specific model training. Motivated by this capability, this paper investigates whether single-image dehazing can be effectively reformulated as an in-context tabular regression problem and introduces DehazeTab, a framework that estimates atmospheric transmission from structured image-patch representations. Specifically, overlapping patches extracted from hazy images are transformed into tabular samples using handcrafted haze-related descriptors, pretrained visual embeddings, or their hybrid combinations, while corresponding transmission patches are summarized as scalar regression targets. A pretrained TFM then performs in-context regression using labeled reference patches to predict transmission values for unseen images without any further fine-tuning. These predictions are spatially aggregated into a transmission map, refined through guided filtering, and incorporated into the atmospheric scattering model to recover haze-free images. Experiments on benchmark datasets demonstrate competitive restoration performance and reveal that physically motivated handcrafted features can outperform pretrained visual embeddings, while hybrid representations offer complementary benefits. Further analyses examine the effects of feature representations, patch selection, and contextual reference configurations on restoration quality. Beyond image dehazing, the findings highlight the potential of TFMs as an alternative paradigm for addressing low-level vision tasks through structured and in-context inference.

Figures & tables

Type	Patch	Features					Target
Type	Patch	$\boldsymbol{f_{1}}$	$\boldsymbol{f_{2}}$	$\boldsymbol{f_{3}}$	$\boldsymbol{\cdots}$	$\boldsymbol{f_{n}}$	Target
Context	$P_{1}$	0.61	0.18	0.43	$\cdots$	0.52	0.78
	$P_{2}$	0.73	0.11	0.57	$\cdots$	0.66	0.61
	$P_{3}$	0.45	0.27	0.31	$\cdots$	0.48	0.84
	$\vdots$	$\vdots$	$\vdots$	$\vdots$	$\ddots$	$\vdots$	$\vdots$
Unseen	$P_{u,1}$	0.68	0.15	0.49	$\cdots$	0.58	$\times$

TABLE I: Illustration of the patch-to-tabular formulation. Labeled patches constitute the context, while unseen patches are provided without transmission targets.

Fig. 1: Offline preparation of spatially aligned patch features and transmission targets for contextual table construction.

Type	Description	Research Question
(i) Handcrafted features	DCP, statistical, and texture features	Are physically motivated features effective for this reformulated task?
(ii) Local embedded features	Features from a pretrained CNN (e.g., ResNet)	Do local and hierarchical visual representations provide better transmission cues?
(iii) Global embedded features	Features from a pretrained ViT (e.g., ViT-Tiny)	Does global contextual information improve transmission estimation?
(iv) Hybrid features	Handcrafted + embedded features	Do handcrafted and learned features provide complementary information?

TABLE II: Regression input configurations for transmission estimation using tabular foundation models.

Fig. 2: Reference retrieval and in-context inference process of DehazeTab.

Method	Publication	Type	SOTS-Outdoor		HSTS
Method	Publication	Type	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
DCP [ 1 ]	TPAMI’11	Prior	18.38	0.819	17.01	0.803
BCCR [ 22 ]	ICCV’13	Prior	15.71	0.769	15.21	0.747
JEEF [ 23 ]	JVCIR’23	Prior	17.51	0.857	17.61	0.861
RSVT [ 18 ]	PCS’24	Prior	22.31	0.903	21.53	0.887
ZID [ 11 ]	TIP’20	Zero-shot	20.27	0.878	22.65	0.901

TABLE III: Performance comparisons on the SOTS-Outdoor and HSTS datasets ( bold : best, underline : second-best).

Fig. 3: Typical visual results on SOTS-Outdoor.

Method	Type	Natural Images
Method	Type	NIQE $\downarrow$	BRISQUE $\downarrow$
DCP [ 1 ]	Prior	3.21	15.64
BCCR [ 22 ]	Prior	3.85	21.74
RSVT [ 18 ]	Prior	3.08	14.66
USID-Net [ 24 ]	Unsupervised	3.32	24.95
SGDRL [ 27 ]	Unsupervised	3.21	15.41

TABLE IV: Performance comparisons on natural scenes ( bold : best, underline : second-best).

Fig. 5: Typical visual results on natural hazy scenes.

Method	Sampling	SOTS-Outdoor		HSTS
Method	Sampling	PSNR	SSIM	PSNR	SSIM
DehazeTab H	Informative	24.71	0.940	23.45	0.918
DehazeTab CNN	Informative	23.74	0.898	23.27	0.893
DehazeTab ViT	Informative	23.06	0.884	21.18	0.858
DehazeTab Hybrid	Informative	24.87	0.942	23.36	0.916
DehazeTab H	Clustering	24.85	0.940	23.29	0.908

TABLE V: Analysis of different patch sampling methods.

Method	#References	SOTS-Outdoor		HSTS
Method	#References	PSNR	SSIM	PSNR	SSIM
DehazeTab H	5	24.71	0.940	23.45	0.918
DehazeTab CNN	5	23.74	0.898	23.27	0.893
DehazeTab ViT	5	23.06	0.884	21.18	0.858
DehazeTab Hybrid	5	24.87	0.942	23.36	0.916
DehazeTab H	7	24.79	0.942	23.38	0.919

TABLE VI: Analysis of different numbers of reference images.

Method	#Patches	SOTS-Outdoor		HSTS
Method	#Patches	PSNR	SSIM	PSNR	SSIM
DehazeTab H	50	24.33	0.934	22.82	0.900
DehazeTab CNN	50	23.49	0.890	23.36	0.883
DehazeTab ViT	50	22.96	0.885	20.89	0.851
DehazeTab Hybrid	50	24.36	0.934	23.10	0.902
DehazeTab H	100	24.71	0.940	23.45	0.918

TABLE VII: Analysis of different numbers of patches per image.

#Patches ( $P$ )	50	100	100	100	150
#References ( $R$ )	5	5	7	10	5
#Context Rows	250	500	700	1000	750
DehazeTab H,44	2.718	3.151	3.286	3.352	3.248
DehazeTab CNN,128	1.930	2.683	3.978	5.854	4.238
DehazeTab ViT,192	2.124	3.186	4.621	6.714	4.896
DehazeTab Hybrid,172	3.155	4.338	5.601	7.879	5.677

TABLE VIII: Analysis of inference time with different configurations (in seconds).

Fig. 6: Inference time comparison with different configurations.

Image Dehazing

Image dehazing restores clear scene details and colors from images obscured by haze, fog, or smoke.

See all 35 papers

Jun 15, 2026

Dehaze-GaussianImage: Zero-Shot Dehazing via Efficient 2D Gaussian Splatting Representation

Existing single image dehazing methods are often constrained by computational redundancy in pixel-level optimization and the lack of physical interpretability in implicit neural networks. These limitations hinder the balance between representation efficiency and reconstruction fidelity. To address these issues, we propose Dehaze-GaussianImage, the first zero-shot framework that introduces 2D Gaussian Splatting (2DGS) into the image dehazing domain to break the traditional pixel-grid processing paradigm. Distinct from static convolutional neural networks (CNNs) or Transformers, our approach models hazy images as continuous and dynamically evolvable anisotropic Gaussian fields. Specifically, we propose a novel reconstruction-decoupling zero-shot learning strategy that embeds the atmospheric scattering model into the Gaussian parameter space. This strategy drives Gaussian primitives to adaptively split, clone, and prune during optimization, achieving geometric-level decoupling of the transmission medium and clear textures. Furthermore, explicit structure-preserving constraints are introduced to suppress artifacts commonly caused by traditional physical priors. Experimental results demonstrate that the proposed method achieves state-of-the-art (SOTA) performance in a fully unsupervised manner with minimal parameters, highlighting the potential of explicit Gaussian representation for low-level vision tasks.

Yuhan Chen, Wenxuan Yu, Guofa Li +5

College of Mechanical and Vehicle Engineering, Chongqing University, Chongqing 400044, China · Department of Electrical and Computer Engineering, Carnegie Mellon University, Moffett Field, CA 94035, USA · National Innovation Center of Intelligent and Connected Vehicles, Beijing 100089, China +1

Image Restoration Image Dehazing

Ask PDF

Aug 11, 2026

HNDiff: Haze-Noise Diffusion for Image Dehazing

Existing diffusion-based methods have recently made significant progress in image dehazing. However, they typically neglect the physics of haze formation and reconstruct clean images from pure Gaussian noise, thereby limiting their restoration potential. To address this issue, we propose Haze-Noise Diffusion (HNDiff), a novel diffusion framework that embeds the atmospheric scattering model as an inductive bias. By grounding diffusion in physical principles, HNDiff ensures that the restoration aligns more closely with underlying mechanisms of haze formation. In its forward process, we introduce joint haze-noise diffusion with a haze-aware noise scheduler, which progressively adds both haze and noise to an image. Essentially, the scheduler adapts noise levels according to haze density, meaning that regions with heavier haze receive stronger noise injection to encourage content generation, while clearer regions receive lighter noise to better preserve details, which directly links the forward degradation process with the physics of haze. In the reverse process, we then derive a physically consistent dehazing-denoising process that simultaneously removes haze and noise to restore a clean image in a manner aligned with the forward degradation process. To further enhance practicality, we propose Latent HNDiff, which compiles clean latent priors that can be seamlessly integrated into existing dehazing networks to boost performance. Extensive experiments show that our work significantly improves leading dehazing backbones and achieves state-of-the-art results on benchmark datasets. The project page is available at https://jin-ting-he.github.io/HNDiff .

Jin-Ting He, Fu-Jen Tsai, Yan-Tsung Peng +3

Department of Computer Science, National Yang Ming Chiao Tung University, Taiwan · National Tsing Hua University, Taiwan · National Chengchi University, Taiwan +1

Physics-Informed Diffusion Models Diffusion Models for Image Restoration Physics-Informed ML Image Dehazing

Ask PDF

Apr 17, 2026

IncepDeHazeGAN: Novel Satellite Image Dehazing

Dehazing is a technique in computer vision for enhancing the visual quality of images captured in cloudy or foggy conditions. Dehazing helps to recover clear, high-quality images from haze-affected remote sensing data. In this study, we introduce IncepDeHazeGAN, a novel Generative Adversarial Network (GAN) involving Inception block and multi-layer feature fusion for the task of single-image dehazing. Utilizing the Inception block allows for multi-scale feature extraction. On the other hand, the multi-layer feature fusion design achieves efficient reuse of features as the features extracted at different convolution layers are fused several times. Grad-CAM XAI technique has been applied to our network, highlighting the regions focused on by the network for dehazing and its adaptation to different haze conditions. Experiments demonstrate that our network achieves state-of-the-art results in several datasets.

Tejeswar Pokuri, Shivarth Rai

No Institute Given

Image Restoration Image Dehazing Generative Adversarial Networks

Ask PDF

Abstract

Image Dehazing

Explore similar work

Dehaze-GaussianImage: Zero-Shot Dehazing via Efficient 2D Gaussian Splatting Representation

HNDiff: Haze-Noise Diffusion for Image Dehazing

IncepDeHazeGAN: Novel Satellite Image Dehazing