Face Completion Based On Generative Adversarial Networks

With the development of the Internet era,the application of face completion has become more and more extensive.Face completion has gradually gained the attention of the academic community and has become a research hotspot in the field of computer vision.Mature face image restoration technology has very important significance and a wide range of applications,which can be applied to media social,entertainment,security,and archaeology.In order to use the face geometry prior and obtain long-range correlation,we propose two effective face completion algorithms based on generative adversarial networks.1.A face completion algorithm based on a multi-discriminator generative adversarial networks.In order to make full use of the face geometry prior,we train multiple discriminators on specific areas of the face image,combined with perceptual loss,adversarial loss and reconstruction loss.The model can specifically repair the key parts of the face,making the details of the reconstructed face image more realistic.The multiple discriminators used in the method only update the parameters in the training phase,which reduces the calculation amount and parameter amount of the network in the prediction phase,and greatly improves the repair efficiency of the network on the face image2.A face completion algorithm based on a self-attention generative adversarial networks.The method can obtain the long-range correlation between the global and local features of the image in spatial and channel dimensions.The method has a plurality of parallel self-attention modules,including a position attention module,a channel attention module,a foreground attention module,and a foreground cross background attention module.The self-attention module can robustly learn the feature correlation between the foreground and the background of the image,and solves the problem that the convolutional layer in the network is limited by the convolution kernel being too small to obtain the long-range correlation between features.In order to verify the effectiveness of the proposed two methods for face completion tasks,we have carried out sufficient experimental verification on both algorithm models.We performed experiments on two face image databases and compared them with several states of the art face image restoration algorithms.The experimental results show that the above two algorithms have significantly improved the quantitative and qualitative results compared with state of the art.