Attentive Semantic and Perceptual Faces Completion Using Self-attention Generative Adversarial Networks

We propose an approach based on self-attention generative adversarial networks to accomplish the task of image completion where completed images become globally and locally consistent. Using self-attention GANs with contextual and other constraints, the generator can draw realistic images, where fine details are generated in the damaged region and coordinated with the whole image semantically. To train the consistent generator, i.e. image completion network, we employ global and local discriminators where the global discriminator is responsible for evaluating the consistency of the entire image, while the local discriminator assesses the local consistency by analyzing local areas containing completed regions only. Last but not least, attentive recurrent neural block is introduced to obtain the attention map about the missing part in the image, which will help the subsequent completion network to fill contents better. By comparing the experimental results of different approaches on CelebA dataset, our method shows relatively good results.