Decorrelated Adversarial Learning for Age-Invariant Face Recognition

There has been an increasing research interest in age-invariant facerecognition. However, matching faces with big age gaps remains a challengingproblem, primarily due to the significant discrepancy of face appearancescaused by aging. To reduce such a discrepancy, in this paper we propose a novelalgorithm to remove age-related components from features mixed with bothidentity and age information. Specifically, we factorize a mixed face featureinto two uncorrelated components: identity-dependent component andage-dependent component, where the identity-dependent component includesinformation that is useful for face recognition. To implement this idea, wepropose the Decorrelated Adversarial Learning (DAL) algorithm, where aCanonical Mapping Module (CMM) is introduced to find the maximum correlationbetween the paired features generated by a backbone network, while the backbonenetwork and the factorization module are trained to generate features reducingthe correlation. Thus, the proposed model learns the decomposed features of ageand identity whose correlation is significantly reduced. Simultaneously, theidentity-dependent feature and the age-dependent feature are respectivelysupervised by ID and age preserving signals to ensure that they both containthe correct information. Extensive experiments are conducted on popularpublic-domain face aging datasets (FG-NET, MORPH Album 2, and CACD-VS) todemonstrate the effectiveness of the proposed approach.