Research On Face Aging Based On Deep Learning

Image synthesis is a hot research topic in the field of computer vision and face aging belongs to the task of style transfer in image synthesis.Traditional face aging algorithms often require image pair or image sequence for learning.Besides,image features should be designed manually.This makes the design of the algorithm very difficult.In recent years,due to the advent of conditional generative adversarial networks,the distribution of face image conditional on age is easier to learn,and the quality of synthetic image is further improved.The face aging algorithm based on autoencoder and generative adversarial networks can automatically extract the features of the image,and then map the modified features back to the aged image.However,there are still some problems in this kind of algorithm,such as low accuracy of age,complex network architecture,loss of identity information and lack of image details in high age group.To address these issues,this thesis makes the following contributions:(1)In order to solve the problem of low age accuracy of synthetic images,clustering supervision is added to the image features extracted by the encoder,so that the identical pictures can be coded into adjacent space to filter out the identity-independent information in the features.At the same time,the form of clustering loss is deduced in detail,and the corresponding age estimation and face recognition experiments are carried out to verify the effectiveness of the algorithm.The experimental results show that the algorithm improves the age accuracy of the synthetic images.(2)In order to solve the problem of complex network architecture and loss of identity information,a feature-based learning generative adversarial network is used.Meanwhile,the distance loss of image in feature space is introduced.The loss is insensitive to noise,and can better retain identity information,thus increase the robustness of the model.This method is concise,and the function of each module is clear.Besides the training process is stable.The experimental results show that this method can improve the age accuracy of the synthetic images,while better retaining the identity information of the input image.(3)In order to solve the problem of insufficient texture details of synthetic images in older age groups,we try to increase the complexity of the network,and use the deep residual network to carry out exploratory experiments.At the same time,in order to better retain identity information,the Pearson correlation coefficient loss between image features is introduced.By enhancing the correlation between input feature vectors and output feature vectors,the image information can be better preserved.The experimental results show that the residual module improves the age accuracy of the older age group,and the loss of Pearson correlation coefficient can help the network better retain identity information.