Occluded Face Restoration Based on Generative Adversarial Networks

Abstract

In recent years, the combination of Convolutional Neural Networks and Generative Adversarial Networks has played a huge potential in the field of face restoration. In order to effectively repair the large area of random occlusion face, this paper constructs an improved Generative Adversarial Networks model based on the Context Encoder, and proposes a self-localization occlusion face image restoration algorithm. Firstly, the occluded part of the face is marked by occlusion locator, and then the marked face image is sent to the generator of Generative Adversarial Networks for restoration. The model generator uses the Convolutional Neural Networks of the Variational Autoencoder structure, and adds the Batch Normalization layer in the model to enhance the information prediction ability of the generator. At the same time, the discriminator is constructed by combining with VGG19, and the discriminator is trained against the generator. Through the experiment on CelebA face data set, this algorithm is significantly better than other methods in the aspect of random large area occlusion face image restoration.