Face Verification with Disguise Variations via Deep Disguise Recognizer

The performance of current automatic face recognition algorithms is hindered by different covariates such as facial aging, disguises, and pose variations. Specifically, disguises are employed for intentional or unintentional modifications in the facial appearance for hiding one's own identity or impersonating someone else's identity. In this paper, we utilize deep learning based transfer learning approach for face verification with disguise variations. We employ Residual Inception network framework with center loss for learning inherent face representations. The training for the Inception-ResNet model is performed using a large-scale face database which is followed by inductive transfer learning to mitigate the impact of facial disguises. To evaluate the performance of the proposed Deep Disguise Recognizer (DDR) framework, Disguised Faces in the Wild and IIIT-Delhi Disguise Version 1 face databases are used. Experimental evaluation reveals that for the two databases, the proposed DDR framework yields 90.36% and 66.9% face verification accuracy at the false accept rate of 10%.