A Generative Face Completion Method Based On Associative Memory

Associative memory is an important brain function human being owns, which could realize storage and cognition of information. Recurrent neural networks possess the function of associative memory, and they can realize image storage and generation basically. Based on stacked recurrent neural networks, a generative face completion method is proposed to utilize their associative memory function to realize face completion. According to different masked regions, the stacked recurrent neural network can memorize the relationship among row pixels from four directions of images, and associate lost pixels of the masked region in face images. In our method, a parallel multi-streams recurrent architecture with context constraints consider both global and local context information to associate the memorized images efficiently. Moreover, we propose an image hybrid strategy to optimize face images, which merges the associated pixels generated from different angles. Experiments on the Wild Face data set (CelebA) reveal that our generative face completion method can get the state-of-the-art results on the face completion problem.