Subspace based face synthesis method for cancelable face biometrics

Projection based methods such as Principle Component Analysis (PCA) or Non-nagative Matrix Factorization (NMF) are widely being used in many face recognition systems. However, original facial images can be reconstructed when the person’s coefficient feature vector of subspace and a set of basis vectors are known. In this sense, the coefficient feature vector of subspace is same as the person’s identity. Accordingly, coefficient feature vector of subspace should be concealed. Because, similar as the original biometric data, once a coefficient feature vector of subspace is compromised, that cannot be replaced. In order to enhance security of biometrics, cancelable biometrics has been recently introduced. Cancelable biometrics can generate new templates using noninvertible transforms. The transformed template can be used as template of biometric recognition systems. Because the transformed cancelable template does not reveal the original biometric template, the coefficient feature vector of subspace can be concealed by cancelable biometrics scheme. However, previous cancelable face transformations cannot provide facial images that can be recognized by humans. Hence, ‘human inspection’, a particular property of face biometrics cannot be provided by previous cancelable face biometric methods. In this paper, we propose a facial image synthesis method which allows human inspection of cancelabe face templates. Using the proposed method, human-recognizable facial images can be synthesized while providing inter-class discrimination and intraclass similarity.