Face Image Inpainting Algorithm Based on Generative Adversarial Network

Traditional image inpainting algorithms extract effective information from single source images, but the lack of high-level context information often leads to poor restoration results. In this paper, we propose face image inpainting algorithm based on generative adversarial network, which repairs the missing area by high-level context information from multi-source correlated images. The context loss and the perception loss are introduced to find the closest potential distribution space of the image to be repaired. The pixel consistency and content similarity can be guaranteed through the game between the generator and discriminator. Finally, the best mapping vector is used to generate the repaired content and complete the image repair task. Experiments show that this method can successfully predict up to 80% of the semantic missing of facial images on LFW face data set, generate complete and realistic images, and improve the image inpainting effect.