REAL-TIME FACE RECOGNITION BASED ON MULTI-TASK LEARNING WITH RASPBERRY PI

Aboluhom A. A. A., Kandilli İ.

7th INTERNATIONAL ANTALYA CONGRESS OF SCIENTIFIC RESEARCH AND INNOVATIVE STUDIES , Antalya, Türkiye, 11 - 13 Mayıs 2024, cilt.1, ss.103-116

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Cilt numarası: 1
  • Basıldığı Şehir: Antalya
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.103-116
  • Kocaeli Üniversitesi Adresli: Evet

Özet

This paper uses the Raspberry Pi, a popular single-board computer to investigate an approach to multi-task learning for face recognition. The paper aims to demonstrate how this low-cost platform can execute complex deep-learning tasks in real-time. To accomplish this, we utilized MobileNet, MobileNetV2, and InceptionV3 as the core models for shared layers, owing to their balance of efficiency and accuracy. The training was conducted on the VGGFace2 dataset, a widely recognized source of facial images for machine-learning applications. The multi-task learning approach enabled the simultaneous execution of three tasks: identifying individuals, estimating their ages, and predicting their races. The system demonstrated impressive accuracy across all tasks by leveraging shared layers among these deep learning models. Testing results were exceptional, with 97% accuracy for identifying people, 97% for age estimation, and 98% for ethnicity prediction. These high success rates suggest that Raspberry Pi-based face recognition systems have significant potential in realworld applications, such as security systems, personalized customer experiences, and demographic data analysis. The study concludes by emphasizing the significance of multitask learning on compact hardware. The high accuracy rates achieved indicate that complex deep learning models can operate efficiently in a resource-constrained environment. This finding opens the door to further innovation, allowing developers to create more adaptable and accessible face recognition solutions. The multi-task approach also suggests a path towards more efficient resource usage, potentially reducing the overall computational load and energy consumption in real-time applications. Overall, the paper provides a compelling case for implementing multi-task learning for face recognition on Raspberry Pi, with impressive accuracy and versatility. This work points towards future opportunities for extending this approach to other low-cost platforms and multi-functional applications, driving innovation in edge computing