Due to the attenuation and scattering of light in an underwater environment， the images directly captured by imaging equipment suffer from significant quality degradation. Although learning-based underwater image enhancement methods improve the original image imaging quality to a certain extent， most of the existing methods use artificially synthesized or model-generated paired datasets for training. Meanwhile， there is a large domain difference between artificial or model-generated images and real underwater images in distribution， which leads to problems of excessive enhancement and no obvious removal of color shift in the enhancement results. Focusing on these problems， an underwater image enhancement model that integrates domain transfer and attention mechanisms was proposed in this paper. First， an image generation network with domain transfer was designed and combined with the physical imaging model and the water type classifier. In this way， the feature description mapping between images in different domains and scenarios could be learned， thereby reducing the difference between the generated images and the real images. Furthermore， a multi-scale hybrid attention encoder-decoder network was designed. With the help of efficient feature connections and different attention-fused structures， the model's ability to recover local image details was improved. Finally， a global domain association consistency loss function was proposed to better train the network model parameters and improve the quality of image enhancement by constructing content and structure consistent associations of the generated images at each stage of the domain transfer. The proposed model achieved accuracies of 3.140 1， 0.602 1 and 3.076 8， 0.612 4 for the UIQM and UCIQE metrics on the underwater real datasets UIEB and EUVP， respectively. The experiments show that the proposed model could effectively improve the color recovery ability of underwater images， and more details could be recovered.