Research On Face Progression Based On Deep Learning

Face aging(FA)is the process of aesthetically rendering a given facial image to present the effects of aging,which is also known as age progression or age synthesis.Face aging technology is widely used in face recognition,criminal investigation,entertainment,forensics etc.Compared to the traditional FA methods,the deep learning-based ones,which aim at gaining high-level aging features from the basic pixel values of image,have achieved more natural,realistic and reasonable aging effects.However,there are many challenges which are the description of aging features,unimodal representation of aging results,and the distortion in synthesized faces caused by low-resolution inputs.In response to these problems,the researches and innovation achievements of this dissertation are as follows:(1)A young child face aging algorithm based on divide-and-conquer strategy is proposed.In the growth of young children,both shape deformation and texture variation,which are caused by internal factors,impact the aging facial appearances.In the dissertation,two decoupled networks are designed to capture aging effects.It explicitly models geometric and textural transformations using two components: GD-GAN,which simulates the Geometric Deformation using Generative Adversarial Network;TV-GAN,which reconstructs the Textural Variations guided by age-related saliency map.With a warping operation,this work demonstrates the advantages over the state-of-the-arts in terms of synthesizing visually plausible aged images for young children,as well as the personalized features perfectly preserved.(2)A multimodal FA framework via facial disentangled technique of age-specific and age-irrelevant information is proposed.Under the influence of different external factors,the aging appearances have diversities for adults,which reveal the multimodality of face aging.In the dissertation,facial disentangled technique is utilized to separate the age-specific and age-irrelevant information.Specifically,a Variational Autoencoder(VAE)-based encoder is designed to represent the diversity and prevent mode collapse in inference time.Moreover,a cycle-consistency loss is designed to utilize the unpaired facial data thoroughly.The extensive experimental results demonstrate that diverse age-specific features can be sampled logically,which are combined with the instinct age-irrelevant features to yield multimodal aging effects.(3)A face aging algorithm based on super-resolution is proposed.The deep-learning based face aging methods have high demands on the quality of the inputs.In the dissertation,a two-step method is calculating via face super-resolution for low qualified facial images,and then synthesizing faces with face aging methods.For face super-resolution,a component semantic prior guided generative adversarial network(CSPGAN)is proposed.Semantic probability maps of facial components are exploited to modulate features in the CSPGAN through affine transformation.Meanwhile,the discriminative network is designed to perform multiple tasks which predict semantic category and distinguish authenticity simultaneously.Experiments show that the reconstructed super-resolved faces can effectively prompt the authenticity of age progressed images by reconstructing the high frequency information.In summary,this dissertation proposes effective solutions to the difficulties in face aging based on deep learning,which has important reference value for the research of face aging synthesis.