Research on masked face recognition by fusing multi-level features of PVT

Abstract

The prevalence of respiratory diseases has made masks play an important role, which has brought new challenges to face recognition algorithms. Inspired by the multi-scale feature fusion model, a Pyramid Vision Transformer (PVT) based face mask feature extraction model is proposed. The model introduces self-attention mechanism to extract rich face information, and realizes multi-scale attention to mask faces by fusing multi-level feature vectors of PVT. Compared with traditional feature fusion model, the model has higher recognition accuracy and fewer parameters. In addition, the model adopts Sub-center ArcFace loss function to improve robustness. The model was trained on a large scale simulated mask face dataset, and tested and evaluated on ordinary face, simulated mask face and real mask face dataset respectively. The experimental results show that the proposed method has higher recognition accuracy than other mainstream methods, and is an effective mask face recognition method.

Keywords

masked face recognition / Transformer / self-attention / feature fusion