Improving image quality and visualization in complex deep-sea environments is of great importance for underwater scientific research and engineering applications. In order to solve the problems of scarcity of deep-sea datasets caused by the special environment of the deep-sea, as well as the problems of color distortion and low contrast of deep-sea images, a paired deep-sea image dataset DSIEB was created, and on this basis a generative adversarial DM-GAN was built algorithm proposed for networks combining DC attention and MSDR multi-scale dense residuals. First, the DC dual-channel attention mechanism is built in the network hopping connection part to strengthen the connection between channels and extract the texture features of image details. Second, the MSDR multi-scale residual block is embedded in the generator structure to improve the attention is on local information and the ability to reuse features. Finally, a new loss function is reconstructed and the smoothing fidelity SF loss is introduced to guide the network to learn the mapping from the original image to the target image from multiple perspectives. Experiments are carried out on the self-built dataset DSIEB, and compared with seven advanced underwater image enhancement algorithms. The experimental results show that the proposed algorithm has stronger generalization ability and is suitable for various deep-sea images.