With the rapid development of deep neural networks, underwater vision plays an increasingly important role in the underwater robotic operation. However, the scarce underwater datasets greatly limit the performance of deep learning on underwater visual tasks, further hindering the applications of underwater operation. To solve this problem, we propose an underwater image synthesis method, which can directly convert the natural light image into the synthetic underwater image end-to-end. Particularly, a pixel-level self-supervised training strategy is designed to maximize the structural similarity between the synthesized and real images, through training the real underwater images. Finally, extensive experiments are carried out, and the obtained results demonstrate the effectiveness and superiority of our methods by quantitative and qualitative comparisons. The proposed underwater image synthesis method offers a valuable sight for underwater vision and manipulating control.