The identity-level angular triplet loss for cross-age face recognition

Despite promising progress has been achieved on face recognition problems, cross-age face recognition remains a challenging task due to its age variations. Human appearances change along with the age growing process, which increases the difficulty of recognition tasks. Existing methods mainly focus on synthesizing new facial images according to different age levels or isolating age-related features and identity related features. In this paper, we propose an identity-level angular triplet loss for cross-age face recognition. The facial images are projected to an embedding space where the angle between feature embeddings can represent similarities of images. Different from Euclidean distance metric, the angular metric used in our method guides the model to learn discriminative features under large intra-class discrepancy. Angles between intra-class embeddings are reduced while that between inter-class are enlarged. The selection of good triplets is conducted on an identity-level rather than instance-level with a moderate positive mining strategy. Experiments are conducted on cross-age databases and results prove the effectiveness of our method.

Automated summary

This paper proposes an identity-level angular triplet loss for cross-age face recognition, which adjusts the angles between feature embeddings in an embedding space to represent similarities of images and improves the accuracy of recognition by learning discriminative features.