RTE-FM-Dehazer: Radiative Transfer Equation Inspired Flow Matching for Real-World Image Dehazing

2026-07-02 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

RTE-FM-Dehazer is a new single-image dehazing approach addressing limitations of current methods, which often suffer from residual haze and color drift due to reliance on the Atmospheric Scattering Model (ASM) and a scarcity of realistic training data. This novel technique replaces ASM with the Radiative Transfer Equation (RTE), which better models non-homogeneous, multiple-scattering hazy scenes by accounting for both scattering and absorption. It integrates a diffusion-absorption regularizer, derived from a reduced RTE, to guide the flow matching trajectory. Complementing this, the authors developed an automated data pipeline using vision-language models, releasing P-HAZE, a dataset of 50,000 realistic hazy/clear image pairs. Trained exclusively on P-HAZE, RTE-FM-Dehazer effectively eliminates artifacts, demonstrates strong cross-domain generalization, and achieves leading performance on five real-world dehazing benchmarks.

Key takeaway

For computer vision engineers developing robust image processing solutions, RTE-FM-Dehazer offers a significant advancement in real-world dehazing. You should consider integrating RTE-inspired flow matching models, as they overcome the limitations of traditional atmospheric scattering models by better handling complex haze. Furthermore, explore leveraging vision-language models to generate large, realistic synthetic datasets like P-HAZE, which can dramatically improve model generalization and reduce reliance on costly real-world data collection.

Key insights

RTE-FM-Dehazer combines the Radiative Transfer Equation with flow matching and a large synthetic dataset to achieve superior real-world image dehazing.

Principles

• Radiative Transfer Equation models complex haze better than ASM.
• Flow matching can be steered by physical priors.
• Vision-language models enable scalable synthetic data generation.

Method

RTE-FM-Dehazer integrates a diffusion-absorption regularizer, derived from a reduced Radiative Transfer Equation, into a flow matching framework. It is trained on P-HAZE, a 50,000-pair dataset generated via an automated vision-language model pipeline.

In practice

• Use RTE-FM-Dehazer for robust real-world image dehazing.
• Generate large-scale synthetic data with vision-language models.
• Apply RTE-inspired regularizers in diffusion models.

Topics

• Image Dehazing
• Radiative Transfer Equation
• Flow Matching
• Vision-Language Models
• Synthetic Data Generation
• P-HAZE Dataset

Best for: Research Scientist, AI Scientist, Computer Vision Engineer

Related on AIssential

• Towards UAV Image Dehazing: A UAV Atmospheric Scattering Model, Benchmark, and Geometry-Aware Deep Unfolding Network — A new UAV atmospheric scattering model and deep unfolding network effectively dehaze UAV images by accounting for non-uniform haze geometry.
• Single Image Defogging Using a Fourth-Order Telegraph PDE Guided by Physical Haze Modeling — This paper introduces a novel hybrid defogging model that integrates a fourth-order nonlinear Telegraph Partial Differential Equation (PDE) with a physical haze formation model for single image defogging.
• RelativeFlow: Taming Medical Image Denoising Learning with Noisy Reference — RelativeFlow improves medical image denoising by learning relative noise mappings from heterogeneous noisy references.
• Synthetic Image Detection with CLIP: Understanding and Assessing Predictive Cues — CLIP-based synthetic image detectors rely on high-level photographic attributes, not just artifacts, but struggle with cross-generator generalization.
• DDTNet: Degradation Disentanglement and Transfer Network for Test-Time All-in-One De-weathering Adaptation — DDTNet enhances all-in-one de-weathering by disentangling and transferring degradation patterns to generate domain-adaptive training data.
• Judging to Improve: A De-biased VLM-as-3D-Judge Protocol for Single-Image 3D Generation — A hardened VLM-as-3D-judge protocol can specialize 3D generators, but exceeding strong baselines requires more than lightweight PEFT.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.