A Method for Face Image Inpainting Based on Autoencoder and Generative Adversarial Network

Face image inpainting has great value in the fields of computer vision and digital image processing. In this paper, we propose a face image inpainting method based on autoencoder and Generative Adversarial Network (GAN). The neural network for image inpainting consists of two parts, a generator and a discriminator. The autoencoder is used twice in the discriminator part, after the final inpainted image is generated by local discriminator and global discriminator. The final loss function is obtained by combining Generative Adversarial Loss and Mean Squared Error (MSE) Loss [20]. We improve and implement an image inpainting model with two evaluation metrics, namely, Peak Signal-tonoise Ratio (PSNR) and Structural similarity index measure (SSIM) [27], respectively. The proposed model for image inpainting is much more suitable for face image inpainting.