In order to solve the problem that paired-based datasets are difficult to obtain during low-illumination image enhancement and the image quality after enhancement is poor, the implementation of unpaired low-illumination image enhancement is studied by improving the cyclic generative adversarial network. The main method is to improve the structure of the generator and discriminator and improve the quality of the enhanced image. The generator part uses the U-NET model integrating the Vision Transformer structure to replace the original generator model, which improves the periodic consistency and content retention of image transformation, and effectively deals with the long-distance spatial correlation problems common in image research. In view of the characteristics of image research, PatchGAN is selected instead of the traditional discriminator to improve the ability to discriminate image details and improve image quality. The results show that compared with the traditional method, the improved model has better subjective visual effects, and the objective evaluation index is also improved. It is concluded that the effectiveness of the model improvement in this paper is displayed.