Reconhecimento facial baseado no algoritmo de aprendizado profundo YOLO

Abstract

Technological development has made it possible for man in recent decades to make a great leap in computational processing techniques aimed at Artificial Neural Networks (ANN). Facial biometrics is a growing area in the world, and its applications are interesting for both private and public companies. Its growing ability to authenticate in security and entertainment systems makes it one of the best means of identity validation and/or people flow monitoring. However, we still come across systems that have a certain slowness in recognition and that demand high processing power. The objective is, then, to supply this demand with a light and robust recognition architecture, and that at the same time presents optimized performance indexes, such as average precision (mAP - Mean Average Precision), inference, low response time and intersection over union (IoU - Intersection Over Union). The present work presents a supervised learning process, in which an emerging architecture is used that has been highlighted in the scenario of pattern recognition systems in image processing. It is programmed in Python and C languages, using OpenCV and Darknet frameworks in YOLOv4 architecture (version 4). The ANN is worked on in a cloud virtual machine with NVIDIA Tesla T4 GPU in Colab environment, training it in 3 different databases: OIDv4, Personal and Wider Face. Our system can also operate locally on Linux systems such as Ubuntu Minimal, which in turn requires basic configuration adjustments in the Jupyter IDE. The results show the optimized sorting/detection capability in the YOLO architecture, as well as achieving improved indices on the job. The dataset OID obtained a mAP of 69.23% for object class, with an average inference of 82.2%, detection time of 16 seconds and reached an IoU of 52.63%. The second dataset Personal achieved an incredible 99.11% mAP in individual recognition facial biometrics, with an average inference of 98%, detection time of 3 seconds and achieved an IoU of 82.56%. And finally, the third dataset Wider Face, obtained a mAP of 86.04% in multivariate facial biometrics, with an average inference of 91%, detection time of 5 seconds and reached an IoU of 61.32% . The results, therefore, demonstrate the quality of the neural network developed in virtue of the objectives initially proposed, in which they are promising in relation to comparisons with other models in the literature in the state of the art in recent years.