SCDE-Dehaze: Semantic-Consistent Detail-Enhancement Learning for Real-World Image Dehazing

Zhiyuan Song, Hannan Lu, Haiqian Han, Bo Li, Pengxu Wei, Xiangyang Ji, Liang Lin

Keywords: real-world dehazing, unsupervised learning, detail-enhancement, semantic-consistent dehazing

TL;DR: The first self-supervised haze-agnostic detail-enhancement learning task with semantic-consistent style translation module for real-world image dehazing.

Abstract: Domain transfer-based methods for image dehazing effectively address the limited generalization inherent to physics-based priors. However, these approaches still exhibit two major limitations: (i) insufficient restoration and enhancement of fine details, and (ii) the introduction of semantically inconsistent artifacts due to semantic ambiguity in heavily hazed regions. To overcome these challenges, we propose SCDE-Dehaze, an image-dehazing framework explicitly designed for simultaneous haze removal and detail enhancement. Our framework comprises two components: Anti-Perturbation Cycle-Consistent learning (APCC) for detail enhancement and Semantic-Consistent Domain Transfer (SCDT) for haze removal. APCC enhances fine image details by enforcing cycle-consistency against intentionally injected perturbations. Meanwhile, SCDT guides semantically consistent domain translation by injecting semantic priors derived from clear images. Extensive experiments demonstrate the superiority of our method over state-of-the-art methods in real-world dehazing while achieving ~1.5 dB improvement in PSNR on challenging O-HAZE dataset.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: true

Submission Guidelines: true

Anonymous Url: true

No Acknowledgement Section: true

Submission Number: 3027