Abstract: Due to the complex and variable underwater environment, the collected underwater images usually produced serious problems such as color distortion, low clarity and low contrast. For this, an underwater image enhancement model based on generative adversarial network was designed by this paper. The generator adopted a U-shaped network architecture and utilized multi-dimensional differential deep fusion convolution to improve the feature extraction capability. Based on the efficient additive attention, a multiscale feature fusion Transformer module was built in the jump connection part to acquire multiscale pixel and channel information. In the bottleneck layer, a feature reconstruction Transformer module based on multi-scale dilated attention has been built to capture deeper image information and construct global features. Experiments and tests were carried out on four underwater image datasets, and the experimental results show that the proposed model effectively solves the problems of color bias, detail blurring and low contrast of underwater images. The enhancement performance on the LSUI dataset is optimal, with PSNR, SSIM, and UIQM reaching 25.22, 0.86 and 3.09, respectively.