Abstract: To address the issues of ineffective treatment of non-uniform haze and susceptibility to overfitting due to small-scale training sets in current dehazing algorithms, a dual-branch dehazing network based on multi-scale feature extraction and domain transfer optimization is proposed. The multi-scale feature extraction branch is utilized to learn the color and structure mapping from hazy images to clear ones, the multi-scale estimation based on channel attention is achieved through a multi-scale residual dense module. The domain transfer branch introduces pre-trained ConvNeXt, which enables the model to obtain additional prior information and improve its generalization ability. The experimental results demonstrate the effectiveness of the proposed algorithm in dehazing on both synthetic and real datasets. The proposed algorithm not only has good performance in removing non-uniform haze, but also has superior generalization ability. Moreover, it has achieved satisfactory results in objective evaluation indicators peak signal-to-noise ratio (PSNR) and structure similarity index measure (SSIM).

Key words: image dehazing, multi-scale extraction, domain transfer, channel attention