Robust face recognition model based sample mining and loss functions

Abstract

Traditional face recognition algorithms rely on margin-based softmax loss functions merely. However, these algorithms tend to perform poorly with low-quality images owing to the varying hardness of the datasets. To address this issue, we introduce a face recognition algorithm based on sample mining (FRABSM), an innovative face recognition algorithm that improves performance by integrating sample mining with conventional margin-based methods. Sample mining selectively focuses on specific samples during model training. FRABSM prioritises information-dense samples with more distinctive features. In this study, we present a probability-driven mining strategy that enhances the ability of the model to handle hard samples, thereby significantly increasing its robustness and adaptability. Mathematical evaluations demonstrate the effectiveness of FRABSM. An accuracy of 94.70% is achieved on the CPLFW. Additionally, the experimental results show that our approach achieves improvements over state-of-the-art methods on three renowned datasets (CPLFW, IJB-B, and TinyFace), highlighting its potential and efficiency. The source code is available at https://github.com/Xkf0/FRABSM.