Attention-aware ensemble learning for face-periocular cross-modality matching

Abstract

Face and periocular regions serve as complementary biometric modalities in identity recognition. The face-periocular cross-modality matching (FPCM) provides a versatile solution, especially when traditional face recognition systems encounter challenges due to occlusions or the presence of sunglasses, which can obscure the periocular region, rendering it less effective in periocular recognition systems. This paper introduces a novel approach based on attention-aware ensemble learning (AEL) to address these challenges. This notion is embodied in AELNet, which features an attention-aware shared-parameter encoder and multiple classifier heads. AELNet is designed to harness the complementary features of the face and periocular regions, enhancing the quality of joint embeddings. A key aspect of AELNet is its ability to foster diversity among the classifier heads through unique embedding techniques and batch sampling strategies, ultimately boosting FPCM performance. We demonstrate the effectiveness of the AELNet by conducting extensive experiments on five unconstrained periocular-face datasets as a benchmark. Codes are publicly available at https://github.com/tiongsikng/ael_net.