API: Empowering Generalizable Real-World Image Dehazing via Adaptive Patch Importance Learning

TL;DR

Real-world image dehazing suffers from limited paired data and highly variable haze distributions. The authors introduce API, combining Automatic Haze Generation (AHG) for diverse training data and Density-aware Haze Removal (DHR) with Adaptive Patch Enhancement to handle spatially varying haze, guided by Multi-Negative Contrastive Dehazing (MNCD) losses. The approach achieves state-of-the-art performance on multiple real-world paired and unpaired benchmarks, demonstrating strong generalization and perceptual quality improvements. By integrating patch-wise density awareness with contrastive learning in both spatial and frequency domains, the method offers a practical, scalable solution for robust real-world image restoration under complex haze conditions.

Abstract

Real-world image dehazing is a fundamental yet challenging task in low-level vision. Existing learning-based methods often suffer from significant performance degradation when applied to complex real-world hazy scenes, primarily due to limited training data and the intrinsic complexity of haze density distributions.To address these challenges, we introduce a novel Adaptive Patch Importance-aware (API) framework for generalizable real-world image dehazing. Specifically, our framework consists of an Automatic Haze Generation (AHG) module and a Density-aware Haze Removal (DHR) module. AHG provides a hybrid data augmentation strategy by generating realistic and diverse hazy images as additional high-quality training data. DHR considers hazy regions with varying haze density distributions for generalizable real-world image dehazing in an adaptive patch importance-aware manner. To alleviate the ambiguity of the dehazed image details, we further introduce a new Multi-Negative Contrastive Dehazing (MNCD) loss, which fully utilizes information from multiple negative samples across both spatial and frequency domains. Extensive experiments demonstrate that our framework achieves state-of-the-art performance across multiple real-world benchmarks, delivering strong results in both quantitative metrics and qualitative visual quality, and exhibiting robust generalization across diverse haze distributions.

Figures (10)

• Figure 1: Visual comparison of dehazing results on real-world images. Unlike synthetic data, haze in real-world scenes typically exhibits diverse distributions, including dense haze (first row), non-uniform haze (second row), and light haze (third row). Our method produces more realistic and comprehensive dehazing results than previous state-of-the-art approaches.
• Figure 2: Overview of the proposed API framework. AHG adopts a hybrid data augmentation strategy to generate realistic and diverse hazy images, serving as the high-quality training data. DHR further considers hazy regions with varying haze density distributions in an adaptive patch importance-aware manner for generalizable real-world image dehazing.
• Figure 3: Architecture of the Adaptive Patch Enhancement (APE). APE handles spatial and frequency patch-wise features separately, with adaptive residual connections applied to each patch.
• Figure 4: A qualitative comparison between our method and state-of-the-art methods on real-world unpaired dataset Fattal fattal, RTTS SOTS and URHI SOTS.
• Figure 5: Visualization of hazy images generated by AHG. The first six examples illustrate the effect of varying haze intensity controlled by Equation \ref{['equ3']}, while the last three demonstrate the modulation of spatial haze distribution via Equation \ref{['equ5']}.