Real-time Facial Recognition Using Multi-Task Learning on a Raspberry Pi

This paper investigates multi-task learning for facial recognition using the Raspberry Pi, a popular single-board computer, to demonstrate how this inexpensive platform can perform deep learning tasks complexity in real time. We used MobileNet, MobileNetV2, and InceptionV3 as base models due to their efficiency and accuracy. The MTL models training were performed on a database built from photos of known individuals and celebrities from the VGGFace2 dataset, divided into three tasks: identifying individuals (9 classes), age estimation (3 groups), and ethnicity prediction (3 groups). Multitask learning enables the simultaneous execution of these tasks using shared layers between deep learning models. The results show a high accuracy rate: MTL InceptionV3 models achieved 93.3% person identification, 95.6% age estimation, and 97.5% ethnicity prediction. The MTL MobileNet model achieved the highest accuracy with 99% person identification, 99.3% age estimation, and 99.5% ethnicity prediction. The MTL MobileNetV2 model achieved 98.3% results in person identification, 97.3% in age estimation, and 99% in ethnicity prediction. These results demonstrate the significant potential of Raspberry Pi-based facial recognition systems in real-world applications such as security systems, personalized customer experiences, and demographic analytics. This study shows that multitask learning on the Raspberry Pi is practical, demonstrating that complex deep learning models can run efficiently even with limited resources. This opens opportunities for innovation, making facial recognition systems more flexible and easier to use. It also enables better resource utilization, thereby reducing computational load and energy consumption in real-time applications.