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

2026-06-14

AI Summary

Existing single-image dehazing methods are often hindered by computational redundancy and a lack of physical interpretability, making it difficult to simultaneously achieve efficiency and fidelity. This work proposes the first zero-shot dehazing framework by introducing 2D Gaussian splatting into this task, modeling hazy images as continuous, dynamically evolving anisotropic Gaussian fields. The atmospheric scattering model is explicitly embedded into the Gaussian parameter space, enabling geometric-level disentanglement of the transmission map and the haze-free image. Through adaptive Gaussian splitting, cloning, and pruning mechanisms, along with structure-preserving constraints, the method effectively suppresses artifacts and achieves state-of-the-art performance under fully unsupervised conditions with remarkably low parameter counts, thereby demonstrating the potential of explicit Gaussian representations in low-level vision tasks.

Abstract

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.

Problem

• image dehazing
• computational redundancy
• physical interpretability
• representation efficiency
• reconstruction fidelity

Innovation

• 2D Gaussian Splatting
• zero-shot dehazing
• atmospheric scattering model
• geometric decoupling
• explicit representation