Estimação de idade e reconhecimento de pele em Imagens digitais a partir de deep learning

Abstract

With the technological evolution of digital media, such as social networks and online communities, privacy and security on the Internet have become issues of debate regarding infrastructure and regulation by legislative bodies. The easy access provided to users, along with the exacerbated number of information shared by the network, sustains one of the problems faced by criminal authorities, the spread of illegal content involving child pornography. Over time, computational methods were used to assist federal experts in assessing the seized content. Computer vision is one of the main areas of development of these methods, using deep learning techniques in artificial neural networks to recognize characteristics associated with the classification of child pornographic content. These neural networks use mathematical processes capable of simulating the cognitive process of the human brain. Through the study of deep learning techniques, applied in digital images and using pre-trained neural structures to perform age estimation and skin segmentation in pre-selected image bases, this study obtained an accuracy percentage of 60.9%, 82.67% and 95.9% for the age estimation experiments, surpassing the state of the art. For the skin segmentation experiments, the neural model generated 83.65% and 65.03% accuracy, lower than the related works studied. In conclusion, both the neural model and the database defined for the age experiments were satisfactory, while the skin segmentation tests showed an inadequacy of the neural network determined with the database used.