HDUD-Net: heterogeneous decoupling unsupervised dehaze network

Haze reduces the imaging effectiveness of outdoor vision systems, significantly degrading the quality of images; hence, reducing haze has been a focus of many studies. In recent years, decoupled representation learning has been applied in image processing; however, existing decoupled networks lack a specific design for information with different characteristics to achieve satisfactory results in dehazing tasks. This study proposes a heterogeneous decoupling unsupervised dehazing network (HDUD-Net). Heterogeneous modules are used to learn the content and haze information of images individually to separate them effectively. To address the problem of information loss when extracting the content from hazy images with complex noise, this study proposes a bi-branch multi-hierarchical feature fusion module. Additionally, it proposes a style feature contrast learning method to generate positive and negative sample queues and construct contrast loss for enhancing decoupling performance. Numerous experiments confirm that the proposed algorithm achieves higher performance according to objective metrics and a more realistic visual effect when compared with state-of-the-art single-image dehazing algorithms.