Combining ArcFace and Visual Transformer Mechanisms for Biometric Periocular Recognition

Abstract

In the last decades, advances in Biometrics have resulted in the popularization of biometric identification applications in different scenarios. However, biometric recognition techniques can exhibit sub-par performance in undesirable or restricted scenarios. Therefore, there is still a need to investigate better recognition techniques and more appropriate biometric traits. Studies have shown that attention is an important mechanism present in biological vision systems, including the human vision system, that can improve significantly the correct recognition rates in computer vision systems. Studies have also shown that periocular characteristics suffer less from environmental changes than faces in undesirable scenarios, achieving similar performance using only 25% of all the data in the face. Motivated by these findings, this paper proposes a new method for periocular recognition based on attention mechanisms that incorporates a recent ViT architecture together with the ArcFace loss function. Experimental results obtained on UBIPr and FRGC, two popular datasets, showed that the proposed method obtained lower error rates when compared to other state-of-the-art periocular recognition methods, in addition to being able to provide the visualization of attention weights for a better understanding of the most important periocular regions used by the neural network for biometric recognition.