Deep learning approach for face recognition with disguise variations

Abstract

With the increased interest in face recognition across di erent applications, the research in this area has ourished over the past few decades. However, face recognition with disguise variations has gained little attention. Faces in unconstrained settings with disguise as an additional covariate makes this problem challenging. It includes alterations in facial appearance using disguise accessories. In this thesis, we propose deep learning based transfer learning approach to handle the problem of disguise, with the network being ne-tuned on the proposed loss function termed as \Disguised Loss". We have evaluated our network on Disguised Faces in Wild (DFW) 2018 dataset [1] where the proposed algorithm is able to produce competitive results. We have also introduced a new dataset termed as \DFW2019" which is an extension of DFW2018 dataset [1]. Apart from the addition of 250 subjects with 3140 images, 250 plastic surgery image pairs and 100 bridal image pairs have also been added. Additional protocols for plastic surgery face recognition have also been introduced. We have presented baselines for all the protocols along with the results of the proposed approach.