Aiming at the problem that the existing dehazing algorithms lack attention to the noise concentration in different regions of the hazy image and the distinction between far and near features， this paper proposes a new generative adversarial network model. In the model， two UNet3 + networks are used to realize the full-scale jump connection and depth supervision， and multi-scale feature fusion is used to extract the high and low-level semantics in different scale feature images. The addition of deep supervision can better learn the near-far level representation in the image. At the same time， the multi-scale pyramid feature fusion module integrating the self-attention mechanism is added to the generator structure to better retain the multi-scale structure information of the feature map and improve the attention to different haze concentration regions. The experimental results show that the algorithm network can obtain better visual effects than other advanced algorithms such as BPPNET on NTIRE 2020， NTIRE 2021， O-Haze datasets， and Dense-Haze datasets. The peak signal-to-noise ratio and structural similarity index on the Dense-Haze dataset are， respectively， 24.82 and 0.769.