You only need haze: Bidirectional disentangled translation network for unsupervised image dehazing

Recently, learning-based methods have achieved notable progress in image dehazing through supervised training on synthetically paired datasets. However, the substantial domain gap between synthetic and real-world hazy images often impairs generalization performance, thereby limiting their effectiveness in practical applications where target domains differ significantly from the training data. Even worse, acquiring sufficient pixel-aligned hazy-clear image pairs in real-world scenarios is costly and challenging. To this end, we introduce a novel Bidirectional Disentangled Translation Network (BDT-Net) for unsupervised dehazing, which regards haze removal as a feature disentanglement task, i.e., separating content-relevant information from the clean factor and haze-related information from the fuzzy factor. Specifically, we design a dual-branch disentanglement framework comprising a Content Recovery Branch (CRB) for extracting structural content information and a Parameter Estimation Branch (PEB) dedicated to capturing haze-related characteristics. Among them, we leverage forward dehazing and reverse rehazing physics-based models to establish haze cycle consistency, thus our BDT-Net can be optimized only needing the hazy image itself. To better distinguish the haze information from the clean content in the latent space, we design an effective Feature-wise Contrastive Representation (FCR), which can not only consider the inherent self-similarity within each information flow but also exploit the mutual exclusivity between different components. Furthermore, a two-way Pixel-wise Contrastive Representation (PCR) is incorporated to enhance the capability of restoring clear image clarity and content. Extensive experimental results on benchmark datasets demonstrate the superiority of our BDT-Net over other compared state-of-the-art methods.