Detecção facial automática baseada em redes neurais convolucionais

Abstract

Given the large number of applications that use facial recognition, such as access authorizations in web applications, use of public services, public safety, entertainment, etc., this work implemented the detection and facial recognition in Python of the neural networks MobileNet, VGG19 and VGG16 applied to the dataset labeled faces in the wild (LFW). The data augmentation and Transfer Learning techniques were employed. The former assists in the development of a robust training set and the latter in the use of models with pre-adjusted weights. To obtain the best results, the training was carried out with the cross-validation technique, which aims to obtain robust models. As for the results, accuracies of 61.0%, 87.0% and 98.0% were obtained for the networks MobileNet, VGG19 and VGG16, respectively. Neural networks available natively in Python were used, which have their weights adjusted according to previous training with the set of images imagenet. The architecture of the models can be easily adapted by adding new layers. The VGG16 network proved to be effective for facial recognition with good performance in terms of prediction time.