Abstract: Face recognition is widely used in security, authentication, and surveillance. However, recognizing partially occluded faces remains a significant challenge due to missing facial features. This project proposes a 3D Partially Occluded Face Recognition System Using Hybrid Deep Learning Techniques, integrating 3D geometric facial structure, texture analysis, and advanced deep learning models to improve recognition accuracy in occluded scenarios. The system employs ResNet50 for robust 2D feature extraction, while PointNet++ processes 3D facial point cloud data. To mitigate the impact of occlusions such as masks, sunglasses, and scarves, selfattention mechanisms and transformer-based CNNs are used to focus on unoccluded facial regions. Additionally, feature-level fusion combines 3D structural features with facial texture to enhance performance. A diverse dataset, including BU-3DFE, FRGC v2.0, Bosphorus, and FaceWarehouse, is used for training and evaluation. The system is tested across various occlusion types to ensure robustness, reliability, and high recognition accuracy. Performance metrics such as recognition accuracy, F1 score, and occlusion robustness score are used for evaluation. For real-world deployment, the system is integrated into a web-based application the proposed system significantly improves face recognition accuracy under occlusions, making it a practical solution for security and authentication applications
Keywords: 3D Face Recognition, Occlusion Handling, Deep Learning, Feature Fusion, Hybrid Recognition Techniques, Web Deployment.
|
DOI:
10.17148/IJARCCE.2025.14655
