Comparativa de extractores de características para clasificación de rostros

Abstract

This work proposes realizing a comparative between two image feature extraction automatized methods applied to the face classification problem. The methods proposed are: first, extraction of feature vectors from images through a convolutional neural network, and then compare and determine if refinement of these features through an unsupervised machine learning algorithm actually improves the posterior performance of the classifier algorithms. The algorithms to be used in the feature extraction are the following: a deep neural network with the GoogLeNet arqchitecture in it’s InceptionV3 version, KMeans for unsupervised learning and two supervised learning algorithms to validate the quality of the features extracted. At the end a comparison will be made between the four results, and it will be verified if the feature refinement through the unsupervised machine learning algorithm really enhances the performance of supervised machine learning algorithms.