Development of a Deep Learning Model with a Novel Loss Function for Facial Recognition based on ArcFace

This thesis focuses on the development of a novel loss function for deep learning-based facial recognition models. Building upon the foundation of ArcFace, the proposed loss function integrates key concepts from recent advancements, including MagFace, CurricularFace, and AdaFace, to enhance model performance. The ResNet architecture, specifically leveraging pre-trained models, is fine-tuned to accommodate this new loss function. The effectiveness of the resulting models is rigorously evaluated on the IJB-C dataset, with performance metrics highlighting the improvements in recognition accuracy and robustness. This research contributes to the field of facial recognition by offering a more refined loss function that balances identity separation and intra-class compactness, thereby improving model generalization.