Abstract: Abstract：The use of deep learning methods， specifically encoder-decoder networks， has obtained exceptional performance in image dehazing. However， these approaches often solely rely on synthetic datasets for training the models， ignoring prior knowledge about hazy images. It presents significant challenges in achieving satisfactory generalization of the trained models， leading to compromised performance on real hazy images. To address this issue and leverage insights from the physical characteristics associated with haze， this paper introduces a novel dual-encoder architecture that incorporates a prior-based encoder into the traditional encoder-decoder framework. By incorporating a feature enhancement module， the representations from the deep layers of the two encoders are effectively fused. Additionally， Transformer blocks are adopted in both the encoder and decoder to address the limitations of commonly used structures in capturing local feature associations. The experimental results show that the proposed method not only outperforms state-of-the-art techniques on synthetic data but also exhibits remarkable performance in authentic hazy scenarios.

Key words: Key words： image dehazing, image restoration, Transformer