Face Aging Simulation Based On Hierarchical Model And Sparse Representation

Human face is the primary cue for identification in our dairy life and the study of face images analysis has received the widespread attention in the fields of pattern recognition, computer vision, etc. However, aging often causes the significant and inevitable deformations in face appearance, and these deformations make the performance of face recognition system degenerate dramatically. Computer-aided automatic aging simulation technology can not only improve the robust performance of face recognition system, but also play an important role in finding missing children, digital entertainment and criminal investigation, etc.The research focuses on face representation and aging simulation. We propose hierarchical model for face representation and an algorithm for face aging simulation based on hierarchical and sparse representation. The research contributions are as follows:(1) A hierarchical face representation model is proposed. After analyzing many kinds of information in face images and various face representation model, we propose a novel hierarchical method for face representation. This method includes global representation layer, local representation layer and detailed representation layer. In global layer, we use a statistical appearance model to represent face in shape space and texture space, respectively. In local layer, the local aging features are represented via sparse representation. At the detailed layer, we divide the face into several facial components zones and skin zones to represent them. The experimental results demonstrate that the method is compact and it can express rich information in face images effectively.(2) A novel face aging simulation method based on the proposed face representation model is presented. Through studying and analyzing many kinds of face aging simulating methods based on different face features, we present a hierarchical model for face aging, which includes three parts: global aging layer, local aging layer and detailed aging layer. In global aging layer, we learn age trajectories to simulate aging result in shape and texture space, respectively. In local aging layer, we train a set of age-specific coupled dictionaries and the faces are aged via sparse representation and face hallucination technique. In detailed aging layer, we calculate the similarity of testing images and training images based on the LBP features and Hu invariant moments of them, and choose the most similarity training images to simulate the aging effects. Finally, the Poisson image editing can be used to synthesize the final aging results. Experiments show that the method can obtain accurate and realistic aging simulations, and preserve the identity information efficiently at the same time.