Look Across Elapse: Disentangled Representation Learning and Photorealistic Cross-Age Face Synthesis for Age-Invariant Face Recognition

Despite the remarkable progress in face recognition related technologies,reliably recognizing faces across ages still remains a big challenge. Theappearance of a human face changes substantially over time, resulting insignificant intra-class variations. As opposed to current techniques forage-invariant face recognition, which either directly extract age-invariantfeatures for recognition, or first synthesize a face that matches target agebefore feature extraction, we argue that it is more desirable to perform bothtasks jointly so that they can leverage each other. To this end, we propose adeep Age-Invariant Model (AIM) for face recognition in the wild with threedistinct novelties. First, AIM presents a novel unified deep architecturejointly performing cross-age face synthesis and recognition in a mutualboosting way. Second, AIM achieves continuous face rejuvenation/aging withremarkable photorealistic and identity-preserving properties, avoiding therequirement of paired data and the true age of testing samples. Third, wedevelop effective and novel training strategies for end-to-end learning thewhole deep architecture, which generates powerful age-invariant facerepresentations explicitly disentangled from the age variation. Moreover, wepropose a new large-scale Cross-Age Face Recognition (CAFR) benchmark datasetto facilitate existing efforts and push the frontiers of age-invariant facerecognition research. Extensive experiments on both our CAFR and several othercross-age datasets (MORPH, CACD and FG-NET) demonstrate the superiority of theproposed AIM model over the state-of-the-arts. Benchmarking our model on one ofthe most popular unconstrained face recognition datasets IJB-C additionallyverifies the promising generalizability of AIM in recognizing faces in thewild.