Age-Puzzle FaceNet for Cross-Age Face Recognition

摘要

Cross-Age Face Recognition (CAFR) has drawn increasing attention in recent years. Technically however, due to the nonlinear variation of face aging and insufficient datasets covering a wide range of ages, it remains a major challenge in the field of face recognition. To address this problem, we propose a novel model called Age-Puzzle FaceNet (APFN) based on adversarial training mechanism. The model we propose can be subdivided into two networks consisting of three elementary parts: (a) a Generator G: our core part for generating age-invariant identity features; (b) an Identity Classifier which forms the first identity recognition network (namely IRN) with the generator G to enhance identity recognition performance; (c) an Age Discriminator which attempts to retrieve age information from the generated features and forms the second network (namely Age Verification network AVN) with the same generator. Our extracted features achieve improvement on age invariance via adversarial training in AVN while remaining identity discriminative utilizing joint training in IRN. Apart from achieving state-of-the-art performance, APFN has demonstrated two other distinct characteristics as follows. First, identity-labeled dataset and age-labeled dataset are used respectively for above two networks such that no more effort to search for training data labeled by both age and identity. As a consequence, more training data is available to give a better recognition performance. Second, the strategy we adopt to handle age and identity attributes can provide a new insight on other robust recognition domain with respect to multi-attributes or attributes separation. We conducted comprehensive experiments on two publicly available datasets called Cross-Age Celebrity and Cross-Age LFW. The results of our proposed architecture demonstrate a better performance and effectiveness.