4KDehazeFlow: Ultra-High-Definition Image Dehazing via Flow Matching

TL;DR

4KDehazeFlow addresses UHD image dehazing by reframing restoration as Flow Matching over a continuous vector field. It integrates a haze purifier and a learnable Haze-LUT within a time-dependent vector field and solves the resulting ODE with a fourth-order Runge-Kutta integrator, enabling stable, efficient 4K dehazing. Key contributions include a learnable 3D LUT for adaptive color correction, a haze-aware vector field design, and a data-driven, architecture-agnostic framework that outperforms seven SOTA methods while reducing computation. The approach demonstrates strong quantitative gains (e.g., ~2 dB PSNR improvements) and favorable runtime on UHD datasets, with robust qualitative restoration and color fidelity in real-world hazy scenes.

Abstract

Ultra-High-Definition (UHD) image dehazing faces challenges such as limited scene adaptability in prior-based methods and high computational complexity with color distortion in deep learning approaches. To address these issues, we propose 4KDehazeFlow, a novel method based on Flow Matching and the Haze-Aware vector field. This method models the dehazing process as a progressive optimization of continuous vector field flow, providing efficient data-driven adaptive nonlinear color transformation for high-quality dehazing. Specifically, our method has the following advantages: 1) 4KDehazeFlow is a general method compatible with various deep learning networks, without relying on any specific network architecture. 2) We propose a learnable 3D lookup table (LUT) that encodes haze transformation parameters into a compact 3D mapping matrix, enabling efficient inference through precomputed mappings. 3) We utilize a fourth-order Runge-Kutta (RK4) ordinary differential equation (ODE) solver to stably solve the dehazing flow field through an accurate step-by-step iterative method, effectively suppressing artifacts. Extensive experiments show that 4KDehazeFlow exceeds seven state-of-the-art methods. It delivers a 2dB PSNR increase and better performance in dense haze and color fidelity.

Figures (8)

• Figure 1: Comparison of different dehazing paradigms. (a) Progressive dehazing performs restoration step-by-step, which may lead to suboptimal convergence at the local optima. (b) Diffusion model-based dehazing achieves global optimization through iterative sampling, but has limitations including high computational cost and high noise sensitivity. (c) Our Flow Matching-based dehazing method models the dehazing process as an ordinary differential equation of continuous vector fields, enabling efficient and stable dehazing with adaptive color correction.
• Figure 2: The overall architecture of 4KDehazeFlow. Our method employs a continuous stepwise inference process with adjustable amount of inference steps. The core innovation lies in the vector field, which integrates two key components: (1) an atmospheric scattering purifier implemented through the CNN feature extraction with potential dehazing function, and (2) a data-driven learned 3D LUT for degraded color reconstruction. Notably, the role of ODE is to optimize the entire vector field.
• Figure 3: Visual comparison on 4KID Zheng_2021_CVPR, I-HAZE I-HAZE, and O-HAZE O-HAZE datasets. The results on UHD (4K) scenes demonstrate that the proposed 4KDehazeFlow consistently outperforms state-of-the-art methods across varying haze densities and diverse scene complexities.
• Figure 4: Visual comparison on the SOTS SOTS dataset. The results highlight that the proposed 4KDehazeFlow achieves superior image restoration performance in low-resolution scenes.
• Figure 5: Visualization results of ablation experiments. As evidently demonstrated, the removal of key components leads to significant degradation, including a higher blue peak and greater red deviation from the ground truth values.