Improving Face Recognition by Exploring Local Features with Visual Attention

Over the past several years, the performance of state-of-the-art face recognition systems has been significantly improved, due in a large part to the increasing amount of available face datasets and the proliferation of deep neural networks. This rapid increase in performance has left existing popular performance evaluation protocols, such as standard LFW, nearly saturated and has motivated the emergence of new, more challenging protocols (aimed specifically towards unconstrained face recognition). In this work, we employ the use of parts-based face recognition models to further improve the performance of state-of-the-art face recognition systems as evaluated by both the LFW protocol, and the newer, more challenging protocols (BLUFR, IJB-A, and IJB-B). In particular, we employ spatial transformers to automatically localize discriminative facial parts which enables us to build an end-to-end network where global features and local features are fused together, making the final feature representation more discriminative. Experimental results, using these discriminative features, on the BLUFR, IJB-A and IJB-B protocols, show that the proposed approach is able to boost performance of state-of-the-art face recognition systems. The proposed approach is not limited to one architecture but can also be applied to other face recognition networks.