Recently, face image inpainting has become a fascinating research area in the field of deep learning. However, the existing methods have the disadvantage that the image inpainting results are not clear enough. Therefore, we propose a new face image inpainting method based on GAN (Generative Adversarial Network) in this thesis. Firstly, a deformation network based on GAN is designed. Then we add an identical autoencoder to the generative part of this generative adversarial network. Two loss functions of mean square error (MSE) loss and GAN loss are combined in the training process. Finally, through the analysis of results based on the CelebA dataset, the average of the new model's PSNR (Peak Signal-to-Noise Ratio) is 36.74dB, the average value of SSIM (Structural SIMilarity) is 0.91. Compared with the previous method, the new model has improved the effect of face image inpainting.