This work presents a robust three-dimensional (3D) facial recognition framework designed to address challenges posed by pose variation and partial occlusion. The method integrates multi-feature face maps, lightweight deep learning, and advanced preprocessing to improve recognition accuracy. Pose normalization is performed using the Iterative Closest Point (ICP) algorithm, while depth, azimuth, and elevation maps are derived as geometric representations. MobileNetV2 with ArcFace is employed to ensure computational efficiency and strong inter-class discrimination. To manage occlusions, a multi-feature thresholding strategy detects affected regions, followed by grid-based interpolation to reconstruct missing data. The recognition network is trained using dual-stage transfer learning, where a pretrained backbone is first adapted to 2D facial data and later fine-tuned on 3D inputs. A custom CNN further enhances feature extraction and classification. Evaluated on the Bosphorus dataset, the framework achieves 99.12% accuracy, demonstrating strong performance under realistic constraints.