Cross-Spectral Periocular Recognition by Translating Thermal Images into Visible Light Images

Periocular recognition can be used even if the target person is wearing a mask. However, the accuracy of authentication using visible light images can be degraded due to the lighting environment or changes in facial expressions. One way to overcome this problem is to use thermal images, but switching an existing system from using visible light images to thermal ones is costly in terms of registration again. In this paper, we propose a periocular recognition method that translates thermal images into visible light images and performs matching in the visible light domain. We introduced an identity loss into a Pix2Pix to realize our cross-spectral periocular recognition. We demonstrated that the accuracy using our method was higher than authentication using a Pix2Pix.