TIR/VIS cross-modality modelling via correlative subspace learning

A novel TIR/VIS cross-modality model is presented. It has a live signature and is illumination invariant for the thermal infrared spectrum. It can be used to solve several difficult problems in face biometrics, e.g. face recognition under weak illumination while lacking a thermal infrared face template of enrollment, liveness assurance against photograph and video face spoofing, face disguise identification, etc. In the presented approach, correlative subspace for TIR/VIS cross-modality is learned by canonical correlation analysis and correlation coefficients between TIR and VIS face variables on different face patches are presented by leave-one-out cross-validation. The experiment results show that it is effective to model TIR/VIS cross-modality using canonical correlation analysis via patch correlation coefficient based weighting.