Face Synthesis Of Different Ages Based On Conditional Adversarial Autoencoder

As the application of intelligent technology in human daily life is more and more extensive,the development of face recognition technology is also rapid.The demand for research on face attributes is becoming urgent in areas such as security surveillance,human-computer interaction,targeted commercial publicity,and television entertainment This paper analyzes and summarizes the domestic and foreign research status of face aging technology and age estimation,introduces the development process of generative adversarial networks and autoencoders,and studies the face aging based on conditional adversarial autoencoder using two public face datasets,.The main research contents of this article are as follows:(1)A face age estimation algorithm based on transfer learning is proposed as one of the evaluation methods of face aging algorithm in this paper.The convolutional neural network model for face recognition is used as the basic framework,the classification layer of the original model is removed,and several new fully connected layers are added to classify the age.These models have been fully trained on large data sets and have excellent ability to extract facial features.During the training process,the parameters of the original model are frozen.The new fully connected layer needs to update the parameters according to the optimizer and learn the face age pattern(2)A face aging algorithm based on Conditional Adversarial Autoencoder is proposed The structure of the generator is an autoencoder.The encoder maps the high-dimensional RGB image space to the low-dimensional hidden space to obtain a vector characterizing the personalized features of the face.The one-hot age and gender labels are connected to the feature vector of the hidden space as the input of the decoder.It is hoped that the decoder will generate face images under different ages according to the specified age label.The task of the generator is to generate pictures that can confuse the discriminator,so that the discriminator cannot distinguish whether its input data is the original face picture or the synthetic picture of the generator.Its loss function includes pixel loss and reconstruction loss.The network structure of the discriminator and the generator is the same,the task of the discriminator is to distinguish whether the input data is the original face image,and its loss function includes reconstruction loss and discriminant loss.Combined with the proportional control theory in Boundary Equilibrium Generative Adversarial Networks(BEGAN),it can generate reasonable face images of different ages.