CoA: Towards Real Image Dehazing via Compression-and-Adaptation

TL;DR

This work tackles real image dehazing under limited compute and diverse real-world scenes by introducing Compression-and-Adaptation (CoA), a two-phase framework that first compresses a synthetic-domain dehazing model (MoC) to obtain a compact parameter space and then adapts it to real domains (BiA) through cross-domain bilevel optimization and EMA-inspired updates. The BiA phase combines a cross-domain bilevel formulation with a CLIP-guided loss to align real-world outputs while preserving synthetic-domain performance, all within a lightweight network built on a Res2Net encoder and a multi-scale dehazing decoder. Empirically, CoA demonstrates stability across multiple synthetic domains, flexibility across different dehazing models, and substantial parameter reductions (up to ~94% depending on the base model) with competitive or superior real-world restoration across daytime, nighttime, and underwater haze datasets. The approach offers practical impact by enabling real-time, adaptable dehazing on resource-constrained devices, with publicly available code to facilitate adoption and benchmarking.

Abstract

Learning-based image dehazing algorithms have shown remarkable success in synthetic domains. However, real image dehazing is still in suspense due to computational resource constraints and the diversity of real-world scenes. Therefore, there is an urgent need for an algorithm that excels in both efficiency and adaptability to address real image dehazing effectively. This work proposes a Compression-and-Adaptation (CoA) computational flow to tackle these challenges from a divide-and-conquer perspective. First, model compression is performed in the synthetic domain to develop a compact dehazing parameter space, satisfying efficiency demands. Then, a bilevel adaptation in the real domain is introduced to be fearless in unknown real environments by aggregating the synthetic dehazing capabilities during the learning process. Leveraging a succinct design free from additional constraints, our CoA exhibits domain-irrelevant stability and model-agnostic flexibility, effectively bridging the model chasm between synthetic and real domains to further improve its practical utility. Extensive evaluations and analyses underscore the approach's superiority and effectiveness. The code is publicly available at https://github.com/fyxnl/COA.

Figures (12)

• Figure 1: Performance evaluation. The proposed CoA incorporates model compression in synthetic domain for efficiency and bilevel adaptation in real domain for adaptability, as illustrated in the central sub-figure (b). The left sub-figure (a) presents efficiency across various metrics, clearly showing that our CoA outperforms others by a significant margin. The right sub-figure (c) shows adaptability across different scenes, where it is evident that our CoA consistently performs excellently.
• Figure 2: Qualitative comparisons of algorithmic properties. Arrows indicate specific regions that highlight visible differences.
• Figure 3: Quantitative comparison on three real-world datasets. Four no-reference image quality assessments were calculated.
• Figure 4: Qualitative comparisons on daytime haze and dusty scenes. All these observations come from RTTS and URHI datasets.
• Figure 5: Qualitative comparisons corresponding to Table \ref{['tab:MultiTask']}. All examples are sourced from RTTS and URHI datasets.