Multiple Projective Dictionary Learning to Detect Plastic Surgery for Face Verification

Researchers have shown that the changes in face features due to plastic surgery can be modeled as a covariate that reduces the ability of algorithms to recognize a person's identity. Traditional dictionary learning methods learn a sparse representation using I0 and I1 norms that are computationally expensive. This paper presents a multiple projective dictionary learning (MPDL) framework that does not require the computation of I0 and I1 norms. We propose a novel solution to discriminate plastic surgery faces from regular faces by learning representations of local and global plastic surgery faces using multiple projective dictionaries and by using compact binary face descriptors. Experimental results on the plastic surgery database show that the proposed MPDL framework is able to detect plastic surgery faces with a high accuracy of 97.96%. To verify the identity of a person, the detected plastic surgery faces are divided into local regions of interest (ROIs) that are likely to be altered by a particular plastic surgery. The cosine distance between the compact binary face descriptors is computed for each ROI in the detected plastic surgery faces. In addition, we compute the human visual system feature similarity score based on phase congruency and gradient magnitude between the same ROIs. The cosine distance scores and the feature similarity scores are combined to learn a support vector machine model to verify if the faces belong to the same person. We integrate our proposed MPDL framework for face verification with two commercial systems to demonstrate an improvement in verification performance on a combined database of plastic surgery and regular face images.