Towards Enhanced Security and Suspect Identification: Optimizing Face Views for Real-Time Facial Recognition in a Semi-unconstrained Multi-camera Environment

The field of facial recognition has shown significant advancements over the years. However, these systems are still prone to factors such as varying pose and angles, lightning conditions, and occlusion. The rapid adoption and deployment of multi-camera system in various environments pose to proffer a potential solution to enhance the performance of face recognition systems. This research embarked on a journey to leverage multi-camera environments by focusing on optimizing the input to face recognition system. The research delved into an innovative methodology centred on enhancing face recognition by meticulously selecting the best ‘face view’ across different camera views. To achieve this, the study incorporated a multi-metric evaluation system with consideration on image quality, pose estimation, and occlusion, to develop a comprehensive ‘face view’ scoring system. The CelebA and LfW datasets were leveraged as they emulate semi-unconstrained environments and contain multiple images of individuals. The proposed methodology had the highest precision, recall, and F1-score of 89.05%, 88%, and 87.79%, respectively. This comparative evaluation proves the impact selecting the best ‘face view’ in multi-camera environments. This research shows the potential and impact of optimizing the input of facial recognition systems.