Bidirectional Feature Disentangled Translation Network for Unsupervised Image Dehazing

Recently, learning-based dehazing methods have achieved remarkable performance by supervised training on synthetic paired data. However, the huge domain gap between synthetic data and real-world hazy scenes limits the generalization ability of the model, making the performance of models unsatisfactory in practical applications. To this end, this paper proposes an unsupervised dehazing method, called bidirectional feature disentangled translation network (BFDT-Net), which regards haze removal as a feature disentanglement task, separating content-related information from clean factors and extracting haze-related information from fuzzy factors. Specifically, a two-branch feature disentanglement framework is constructed to learn different feature distributions. Among them, the content recovery branch (CRB) focuses on extracting clear content information, while the parameter estimation branch (PEB) is responsible for extracting haze-related features. Moreover, by combining the forward dehazing and reverse rehazing paradigms of the atmospheric scattering model (ASM), the haze cycle consistency is established, so that BFDT-Net can be optimized only by relying on the hazy image itself. Experimental results on a large number of benchmark datasets demonstrate that the proposed BFDT-Net outperforms the existing state-of-the-art dehazing methods.