Cross-age Face Generation And Verification Via Generative Adversarial Networks

Face related problems are important but challenging in the field of computer vision.They have drawn many researchers' continuous attention and exploration for several decades.However,most of the existing researches are only applicable to unconstrained environments.The light,pose,expression and age changes of the human face will lead to the decline of face recognition performance.The major challenge of face related problem is mostly attributed to the uncontrollable factor–aging.Nowadays,cross-age face research is playing an important role in practical applications.For instance,it can be applied in finding missing persons,identifying criminals,authenticating and entertainment.Lately,the excellent generative capability of Generative Adversarial Network in image generation and recognition tasks has brought new methods to face-related issues.Based on the related theories and methods of Generative Adversarial Networks and the analysis of the current domestic and international cross-age face synthesis and verification,this thesis researches the cross-age face problem based on the Generative Adversarial Networks from the perspective of generation model.We design a new network which combines Res Net and Adversarial Autoencoder to solve the cross-age face problem about generation and verification.The main contributions of this thesis include:(1)We propose a series of improvements based on the Conditional Adversarial Autoencoder(CAAE).Firstly,considering the facial symmetry,we improve the filter kernel to make full use of the horizontal continuous characteristic information,such as eyes,forehead.Thus,the generated image looks more natural.Secondly,we design a content loss based VGG to make synthetic images more detailed.Moreover,we apply the Res Net to encoder structure to achieve more realistic cross-age face images.The experimental results show the effectiveness of the proposed method.(2)We present a new method to solve the cross-age face verification called Disentangled Representation Learning Residual GAN(DR-RGAN).The encoder-decoder structure of the generator enables our model to learn a representation.We train the encoder to encode images into unstructured information.The age code is specified in a disentangled manner,so that our generator is trained to generate images from unstructured representations using the age characteristics provided by the disentangled part of the representation.Besides,the discriminator is trained not only to distinguish real vs.fake but also to predict the identity and age of a face.Through this the encoder represents the facial feature without age interference to realize the face verification.Meanwhile,considering that the pixel-based error can cause the lack of details,we design the content loss based VGG to improve the performance of the proposed model.Experiments on public available face aging dataset-CACD?VS show the effective-ness of the proposed method(DR-RGAN).