DPR-V2S: A Deep Framework for Periocular Recognition in Surveillance Environments

In recent years, biometrics systems have attracted much attention from the biometric community, especially periocular recognition. The periocular region is considered one of the most discriminative and reliable regions of the human face since it suffers less from aging and variations in facial expression. Many studies have shown its potential as a biometric trait that can be used for verification and identification tasks. However, these methods ignore periocular recognition performance between modalities: video-to-still (V2S) and still-to-video (S2V), and surveillance scenarios. Therefore, this research field remains open and it needs to be exploited. Because of that, we decide to perform the periocular recognition in surveillance scenarios and between modalities, i.e., V2S. Our primary goal is to establish a new performance metric under these circumstances so that the periocular region's potential use under uncontrolled conditions and between modalities can be evaluated. Thus, we propose a new approach called DPR-V2S to perform periocular recognition, where the probe data consists of a variable number of frames from a video and the target data consists of a still image. Our method uses information from each frame enhancing its utility in the verification and identification tasks. Results are reported for the Cox dataset and show that our method surpasses the baseline in all evaluation metrics. In the best scenario, our approach achieves 98% of TPR <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$({@}\text{FPR} =1e^{-1})$</tex>, 4.60% of EER, 3.64 of decidability, and 87.84 % of Rank-5.