Realistic Processing For Reconstructed Face And Age Editing Based On Generative Adversarial Networks

Craniofacial reconstruction is one of the main methods for identifying the skulls of unknown body.Craniofacial reconstruction technology,which can extend the knowledge of anthropology,forensic science and other fields,has attracted much attention from many subjects from home and abroad.There are some limitations in the craniofacial models made by existed Craniofacial reconstruction technology because the available craniofacial data is less.The reconstruction can only contain physical geometry information without texture,opening state of eyes and other realistic information,which will influence the further skull identification.A kind of P-GAN based on generative adversarial Network as well as a kind of age editing of the craniofacial photo after realistic processing by Adversarial Autoencoder Network are proposed in this paper,and a prototype system of realistic craniofacial processing and age editing is designed and developed.This paper includes the following research contents:1)In view of the insufficient realistic effect of the reconstructed face made by Pix2Pix(Image-to-Image Translation)Network,a kind of P-GAN Network based on Pix2 Pix Network,which can be used to do realistic craniofacial processing,is proposed in this paper.P-GAN Network adds craniofacial information constraint network P-net on the basis of Pix2 Pix,in which the craniofacial features are captured by VGG16(Visual Geometry Group),network convergence is strengthened by SE-Block(Sequeze and Excitation)and the most similar samples are selected by Face++.The reconstructed craniofacial photo and real human face photo are inputted into the network,and Contrastive loss is combined with Triplet loss to form a loss function.In this paper,the data of reconstructed craniofacial photo after Frankfurt correction and normalization is inputted into P-GAN Network to get the face with realistic processing.According to the experiment result,through the processing of P-GAN,the reconstructed face become more realistic and its uniqueness of identity is also maintained.By computing the Euclidean distance and cosine distance between two facial photos through Baidu Face API,the experiment results can judge the similarity between the real photo and the other one with realistic processing,which is better than the result of Pix2 Pix.2)As the age information of the reconstructed face may be inconsistent with that in the database,which may result in low matching rate,the adversarial autoencoder network which integrated the generative adversarial network and autoencoder is applied in this paper.The adversarial autoencoder network inputs the facial photo,age and target age.The encoder captures information with nothing to do with age features,the modulator adjusts the age features into one-hot features,and the decoder outputs the age features,reconstructs the facial photo of the targeted age,and does quantitative analysis of the preservation rate of the identity,emotion and background of the reconstructed photo.3)The realistic processing for reconstructed face and age editing system is designed and realized in this paper by the methods mentioned above.The results indicate that,this system can make the reconstructed craniofacial photo infinitely close to the real face photo,and it can do the age editing and reconstruction to the reconstructed craniofacial photo after realistic processing at the same time.