PLGSA-Transformer: Landmark-Guided Attention and Occlusion-Adaptive Thresholding for Masked Face Recognition

The paper introduces PLGSA-Transformer for masked and unmasked face matching. It combines MediaPipe-derived Gaussian heatmaps over the periocular region with EfficientNetB3 features, a residual attention gate, and a two-layer multi-head self-attention branch. A jointly trained Occlusion-Adaptive Cosine Threshold head adjusts the matching threshold according to predicted occlusion severity. On 858 images drawn from Zenodo MDMFR, a masked CelebA-HQ collection, and author-collected data, the authors report 97.22% pair-verification accuracy and ROC AUC 1.0000. The small, heterogeneous evaluation warrants scrutiny.