Enhancing Accuracy of Face Recognition in Occluded Scenarios With Occlusion-Aware Module-Based Network

Face recognition technology despite its extensive application across various domains. However, occlusion factors like masks and glasses are significantly impeded by current face recognition models, resulting in limitations to their practical usage. We present Occlusion-Aware Module Network called Occlusion-Aware Module-based Network (OAM-Net), designed to enhance the accuracy of occluded face recognition. OAM-Net comprises two sub-networks: an occlusion-aware sub-network and a key-region-aware sub-network. The occlusion-aware sub-network incorporates an attention module to adaptively modify the weights of convolutional kernels for optimizing the processing of occluded face images. Meanwhile, the key-region-aware sub-network integrates a Spatial Attention Residual Block (SARB) for precise identification and localization of key facial regions. The network’s generalization performance and accuracy are further enhanced by implementing a meta-learning-based strategy to boost the network’s generalization performance and accuracy. Experimental results affirm OAM-Net’s superior performance of OAM-Net over other state-of-the-art methods in occluded face recognition, underlining its significant potential for practical application.