Unsupervised multi-branch network with high-frequency enhancement for image dehazing

Hang Sun, Zhiming Luo, Dong Ren, Bo Du, Laibin Chang, Jun Wan

Abstract: Highlights•An Unsupervised Multi-branch Dehazing Network (MBDN) is designed to explore how to avoid the learning confusion caused by the input of different data distributions in CycleGAN-based methods, and explore effective ways to enhance the feature representation consistency of multi-branch networks, further improve the dehazing performance of the model.•Based on the observation that hazy images and their clear counterparts exhibit only subtle differences in high-frequency information, a High-Frequency Enhancement Module (HFEM) is proposed to efficiently extract high-frequency details (e.g., edges and textures) from hazy images. This module enhances the ability to capture high-frequency features by adaptively fusing high-frequency information into the network. Moreover, the efficient utilization of high-frequency information by the HFEM enhances network robustness when images are heavily contaminated with haze.•A novel method called UME-Net is developed to seamlessly integrate the designed MBDN and HFEM module to handle image dehazing in various challenging scenes.