In order to realize the image translation from microwave image to optical image and facilitate the visual understanding of microwave image, an optical image inversion method based on generative adversarial network is proposed. In this paper, through the designed MIMO antenna array and the improved bistatic RM algorithm based on spectrum separation, the curved square plate with patches is used for microwave imaging, and the microwave image and corresponding optical image are collected to construct a dataset. The generator network of Cycle GAN model is improved by introducing hourglass module, SE module and AdaLIN normalization method, which strengthens the feature extraction and feature fusion ability of the model for image internal semantics. Model comparison and ablation experiments show that the FID, KID and LPIPS score of the improved Cycle GAN model are 29.84, 0.1725 and 0.2011, respectively. And it has good optical image inversion ability. In this paper, the influence of dataset size and training epoch on the inversion performance of the improved Cycle GAN model is also studied. The results show that the improved Cycle GAN model trained on the large dataset with 200 epochs achieves a better optical image inversion effect.