Learning to invert local binary patterns

Abstract

In this work, we have proposed to invert the local binary patterns (LBP) descriptor. The success of the inversion gives rise to two applications: face de-appearance and re-appearance. The de-appearance, based on image-LBP forward mapping, is thorough in the sense that not only the identity information but also the soft-biometric information of the subject is removed. The re-appearance yields face reconstruction with high fidelity and also enables secure application with a unique encryption key. The reappearance is achieved by learning the inverse mapping of the LBP descriptors through an 0-constrained coupled dictionary learning scheme that jointly learns two overcomplete dictionaries in both the pixel and the LBP domains such that inverse mapping from the LBP domain to the pixel domain is made possible without knowing the mapping function explicitly. The procedure also comes naturally with high selectivity when reconstructing the faces with various LBP encryption keys. We have shown the effectiveness of our proposed approach on the FRGC ver 2.0 database which involves large-scale fidelity test and face verification experiments using the state-of-the-art commercial and academic face matchers.