Face Aging Based On Attribute Control

Face aging is a challenging face attribute editing problem where the goal is to generate an aging/rejuvenated face image of a target age given a face image.This requires the model to be able to ensure that the identity of the individual is fixed while generating an aging/rejuvenated image that matches the current identity.Differences in face appearance are particularly evident due to differences in individual aging patterns,e.g.,some individuals age with features such as grey hair and deepening wrinkles,while others age with features such as hair loss and sagging facial muscles.However,the performance is limited in unconstrained scenarios due to the fuzzy matching of aging features with identity features and the limited ability to express conditional vectors in conditional generation models.To address the above issues,the main research work and results of this paper are as follows:1.To address the problem of mismatching the generated aging/rejuvenation images with the practical aging process of individuals,this paper propose the Latent Age Code Decision(LACD)approach for fine-grained face ageing feature matching.Specifically,this paper model the practical aging process and aging personalisation by reinforcement learning.Both the practical aging process and aging personalisation based on the individual are observations of an accurate description of the individual’s target aging features.This paper’s approach describes the matching of aging features in the aging feature latent space as two Markov decision processes and defines two agents to model the practical aging process and the aging personalisation of an individual.The two agents jointly make rational decisions about the target aging features.This paper conducted experiments on two challenging datasets and the experimental results show the effectiveness of our LACD.2.For conditional generation models with limited conditional vector representation capability,traditional face aging work uses one-hot coding or predefined distributions as conditions for face aging image generation,however,for the complete aging pattern of an individual,the semantic-free one-hot encoding and predefined distributions are used as conditions corresponding to the individual ageing process,both of which are difficult to describe the highly non-linear aging process.This has led to existing conditional generation models preferring to generate an average aging appearance for the target age group.In order to achieve a more fine-grained control of the aging appearance of the target age group,this paper propose to quantify the influence of different attributes on the aging appearance using Shapley values,and thus compute aging feature labels that provide a more reasonable description of the aging appearance of the target age group.3.This paper design a face image aging generation system based on the theoretical basis of the latent age code decision method and the principle of simplicity and ease of use.The face image aging generation system is capable of processing face images,including detecting,recognizing,aligning and cropping face images.It also performs image enhancement,background removal and image generation to ensure that the resulting ageing effect is more realistic.