Underwater images suffer from quality degradation due to light attenuation and scattering effects, which seriously affects underwater target recognition and scene understanding. In this paper, a mathematical model of underwater image quality degradation is established by analysing the principle of underwater imaging and the quality degradation mechanism. Based on CycleGAN (Cycle-Consistent Generative Adversarial Network, CycleGAN) network, an improved image enhancement method is proposed to optimise the discriminator structure and loss function design. The experimental results show that the method achieves 28.3 dB and 0.892 in PSNR and SSIM metrics, respectively, which is a significant improvement over the traditional method. The model shows good enhancement effect in different depth and turbidity environments, and the processing speed is 92 ms/frame, which has strong practical value.