CAAM: A calibrated augmented attention module for masked face recognition

Abstract

Along with other aspects of daily life, the COVID-19 pandemic has a substantial impact on the performance of facial recognition (FR) systems installed in various locations for identity verification. To address this pivotal issue, we propose an attention-guided masked face recognition (MFR) method, named Calibrated Augmented Attention Module (CAAM), which consists of two core components: Recursive Attention Gate (RAG) and an Augmented Feature Calibration Block (AFCB). In the first stage, RAG guides the backbone network to pay attention to non-occluded face regions for feature learning by calibrating multi-layer features while progressively reducing the network’s response to mask-occluded regions in a recursive manner. In the second stage, a dual-branch AFCB first augments the attention map generated by RAG to incorporate cross-dimensional interactions, which are then calibrated to build spatial and inter-channel dependencies across informative spatial locations for MFR. Experiments conducted on various masked face datasets validate the superior performance of CAAM.