Lifespan age synthesis on human faces with decorrelation constraints and geometry guidance

Abstract

It is challenging to use a single portrait as the reference and synthesize matching facial appearances throughout the lifetime. The following issues more or less plague previous attempts at this task: the loss of identity information and unnatural and fragmented changes in age-related patterns. To alleviate these problems, we propose a new method for lifespan age synthesis with decorrelation constraints and geometry guidance. In particular, orthogonality is imposed on two branches of features extracted from the source face so that they encode different kinds of facial information. Additionally, we develop a hybrid learning strategy based on joint supervision of landmarks and age labels, which guides the model to learn facial shape and texture transformation simultaneously. Qualitative and quantitative evaluations demonstrate that our approach outperforms state-of-the-art competitors. Relevant source code is available at https://github.com/zlq1z2l3q/GGDC.