Abstract—Prosopagnosia diminishes an individual’s ability to recognize familiar faces, under pose variations, occlusions, and when faces appear visually similar. This study develops a deep learning framework, ResRAMACL, with ResNet-101, Region-Aware Attention Mechanism (RAAM), and a novel Hybrid Angular-Cosine Margin Loss (HACML), for supporting prosopagnosia patients to recognize human faces. Facial images are aligned using 2D facial landmark plotting, which separates key semantic regions from the images. ResNet-101 extracts feature maps for each region, which are weighted by RAAM to highlight cognitively crucial areas. These features are fused into a single embedding vector and trained using HACML, and the facial embeddings are stored in a Structured Query Language (SQL) database. The prediction phase generates an embedding using a trained model, checks the similarity with the stored embeddings, and predicts the final result. Compared to the existing loss functions, the proposed HACML achieved up to 2.16 units higher inter-class separation and up to 0.44 units lower intra-class variance. Ablation study proved higher performance compared with variations of the pipeline. The generalizability analysis proved minimal variation in performance when tested with other datasets. Overall, this proves the significance of the approach for integration with Augmented Reality (AR)-based or mobile cognitive assistance technologies, enabling prosopagnosia patients to recognize faces.
 
Keywords—face recognition, prosopagnosia, Region-Aware Attention Mechanism (RAAM), ResNet-101, Structured Query Language (SQL) database


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).