Reconhecimento facial utilizando transferência de aprendizado e redes neurais

Abstract

In recent decades, the field of machine learning has progressed rapidly due to increased research efforts and the emergence of new neural network architectures. This document aims to present an analysis of the performance of various convolutional neural networks trained using transfer learning, specifically focusing on their performance in tasks such as facial recognition and classification. To achieve this, publicly available datasets were combined with a set of data containing images of new individuals, which needed to be integrated with the existing datasets and identified individually. Five different architectures of convolutional neural networks were trained (VGG, DenseNet, Xception, ResNet, ResNext). Following the training process, the Grad-Cam algorithm was also employed to generate a heat map, with the purpose of examining how the networks extract features prior to classification.