When Schrödinger Bridge Meets Real-World Image Dehazing with Unpaired Training

2025-07-13

AI Summary

Existing unpaired image dehazing methods—particularly GAN-based approaches—are limited by the generator’s inability to accurately model the complex transmission mapping between hazy and haze-free images, resulting in poor generalization to real-world scenes. To address this, we propose DehazeSB, the first dehazing framework leveraging Schrödinger Bridge (SB) theory, which establishes a few-step, differentiable, bidirectional distribution alignment grounded in optimal transport. We further introduce a detail-preserving regularization to maintain structural fidelity and a haze-aware prompt learning mechanism that integrates CLIP’s semantic priors to enhance discriminative capability. Extensive experiments on multiple real-world dehazing benchmarks demonstrate that DehazeSB significantly outperforms state-of-the-art unpaired methods: outputs exhibit richer texture details, more natural color rendition, and fewer artifacts—validating SB’s effectiveness and superiority in modeling the physical haze degradation process.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Learning to Search
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Recent advancements in unpaired dehazing, particularly those using GANs, show promising performance in processing real-world hazy images. However, these methods tend to face limitations due to the generator's limited transport mapping capability, which hinders the full exploitation of their effectiveness in unpaired training paradigms. To address these challenges, we propose DehazeSB, a novel unpaired dehazing framework based on the Schrödinger Bridge. By leveraging optimal transport (OT) theory, DehazeSB directly bridges the distributions between hazy and clear images. This enables optimal transport mappings from hazy to clear images in fewer steps, thereby generating high-quality results. To ensure the consistency of structural information and details in the restored images, we introduce detail-preserving regularization, which enforces pixel-level alignment between hazy inputs and dehazed outputs. Furthermore, we propose a novel prompt learning to leverage pre-trained CLIP models in distinguishing hazy images and clear ones, by learning a haze-aware vision-language alignment. Extensive experiments on multiple real-world datasets demonstrate our method's superiority. Code: https://github.com/ywxjm/DehazeSB.

Problem

• Overcoming GAN limitations in unpaired image dehazing
• Bridging hazy and clear image distributions optimally
• Preserving structural details in dehazed outputs

Innovation

• Uses Schrödinger Bridge for image dehazing
• Leverages optimal transport theory directly
• Introduces detail-preserving regularization alignment