Real-time facial recognition via multitask learning on raspberry Pi.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Embedded Systems

Background:

• Facial recognition often requires high-end hardware, limiting its accessibility.
• Multi-task learning (MTL) offers efficiency but typically relies on powerful computational resources.
• Resource-constrained devices like the Raspberry Pi present challenges for complex deep learning tasks.

Purpose of the Study:

• To investigate the feasibility of deploying efficient MTL for facial recognition on a Raspberry Pi.
• To evaluate the performance of different base models (MobileNet, MobileNetV2, InceptionV3) for MTL tasks on this device.
• To demonstrate real-time deep learning capabilities on low-cost, embedded hardware.

Main Methods:

• Trained MTL models using MobileNet, MobileNetV2, and InceptionV3 as base architectures.
• Utilized a custom database derived from the VGGFace2 dataset.
• Focused on three facial recognition tasks: person identification, age estimation, and ethnicity prediction.

Main Results:

• MobileNet achieved superior accuracy: 99% for person identification, 99.3% for age estimation, and 99.5% for ethnicity prediction.
• Demonstrated successful real-time performance of MTL models on the Raspberry Pi.
• Achieved high accuracy comparable to systems on high-end hardware.

Conclusions:

• Efficient MTL models can be successfully deployed on resource-constrained devices like the Raspberry Pi for facial recognition.
• This approach significantly reduces computational load and energy consumption.
• Facial recognition systems become more accessible and practical for real-world applications.