Comparação de métodos de reconhecimento facial utilizando Fisherface

Abstract

Humans have the natural ability to recognize the faces of familiar individuals and identify them by name. This process of facial recognition occurs daily, whether when greeting known individuals on the street by name or when viewing an image of a familiar face. While it is a trivial task for humans, automated facial recognition is an extremely complex task for a computer to execute. Over the years, various authors have proposed multiple facial recognition algorithms with the goal of increasing prediction accuracy. Therefore, this master's dissertation explores two approaches to the Fisherface facial recognition algorithm: (i) a segmented approach and (ii) performing face alignment for each image in the face database. Both approaches were tested by analyzing three performance indicators: accuracy, total processing time, and the overall computational load during algorithm execution. The results demonstrated that it is essential to perform the face alignment approach to achieve better accuracy rates. The segmented approach required low memory consumption during testing but exhibited a lower average accuracy rate when executed on a normalized face database.