Aplicação de técnicas de deep learning para autenticação de usuários através do reconhecimento facial em notebooks

Abstract

The present work aims to apply techniques of digital image processing and neural networks to authenticate and unlock notebooks. For this, knowledge of the Viola-Jones algorithm was used on this project, which applies Haar's rectangular features to make complex images simpler for analysis, and of the Adaptive Boosting Algorithm, to classify the weight of the features identified in the images. A deep learning approach associated with artificial neural networks was also conducted to reinforce the correct identification of the faces used to verify the system, and to better understand the operation behind the manipulation of the digital image signal for its analysis. For validation and testing of the system, a dataset of 700 publicly available images from 50 distinct users was employed, utilizing evaluation metrics such as accuracy, precision, sensitivity, specificity, and F-score. According to the results, a minimum accuracy of 98,6% and a maximum of 100% were achieved, along with a minimum precision of 62,5% and a maximum of 100%. Sensitivity ranged from a minimum of 50% to a maximum of 100%, specificity ranged from a minimum of 98,5% to a maximum of 100%, and F-score ranged from a minimum of 66,7% to a maximum of 100%