Towards haze removal with derived pseudo-label supervision from real-world non-aligned training data

Abstract

Single-image dehazing seeks to restore clear images by addressing degradation issues caused by hazy conditions, such as detail loss and color distortion. However, since collecting large-scale and precisely aligned hazy/clear image pairs is unrealistic in real-world scenarios, existing data-driven learning-based dehazing algorithms are often affected by data authenticity and domain gaps between synthetic and real-world scenes, resulting in unsatisfactory performance. To this end, we propose a novel haze removal framework derived from real-world captured non-aligned training data. Specifically, our framework can be divided into two components: a pseudo-label supervision generation stage and an image dehazing stage. For one thing, the former stage explores clean-related style information from the haze-free image and transfers it to its corresponding hazy counterpart, thus generating fine-aligned training image pairs. More clearly, we relieve domain divergence and pixel misalignment through a well-designed Cross-Modulation Align Network (CMA-Net), which includes the Domain Transfer Module (DTM) and Feature Alignment Module (FAM). For another, the later stage focuses on constructing an effective dehazing architecture with a pseudo-label pixel-wise supervision training paradigm. Therefore, we propose a standard U-shape dehazing network with Physics-related Feature Unit (PFU) and Gate Attentive Fusion (GAF). Furthermore, we establish a new non-aligned hazy/clear dataset named Hazy-JXBIT by our camera devices, to further evaluate the proposed dehazing framework. Extensive experimental results demonstrate that these fine-aligned pseudo-label training pairs generated by CMA-Net can be beneficial for building a steady dehazing network USD-Net, and prompt us to obtain superior performance over existing state-of-the-art methods.