Face Recognition System for Unconstrained Condition

Face recognition systems have become an integral part of various applications, including security, surveillance, and identity verification. These systems predominantly rely on deep learning-based approaches, leveraging Convolutional Neural Networks (CNNs) and facial feature extraction techniques. While such systems have shown promising results under controlled conditions, their performance tends to degrade significantly when facing with challenges like face mask occlusion or hard recognition poses. This poses a significant concern, especially in the context of recent global events where mask wearing has become prevalent. This paper aims to propose a novel face recognition system to handle unconstrained conditions and provide experimental results to validate its effectiveness.To address these limitations, our proposed system incorporates two key advancements. Firstly, we leverage the Deep3DFace approach to reconstruct a detailed 3D face model. By utilizing this model, we can overcome the partial occlusion caused by masks, allowing the system to extract relevant facial features more accurately, thereby improving recognition performance. Secondly, we use DREAM (Deep Residual Equivariant Mapping), a novel technique that focuses on handling hard recognition poses. DREAM utilizes equivariant mappings to capture pose variations effectively, enabling the system to recognize faces regardless of challenging pose configurations. This approach enhances the system’s ability to handle unconstrained conditions and improves recognition accuracy in real-world scenarios.We deployed our proposed system on the DGX-A100 system, a state-of-the-art hardware platform specifically designed for deep learning tasks. We conducted comprehensive experiments using benchmark datasets containing individuals wearing masks and exhibiting challenging poses. The results demonstrate the effectiveness of our approach, showcasing improved recognition accuracy and robustness compared to existing systems.