Sistema para Controle de acesso em ambientes restritos utilizando reconhecimento facial através de Redes Neurais

Abstract

Advances in the field of facial identification and facial detection effectively go through the creation of models that allow the automatic extraction of useful representations in images. In this context, this work intended to explore models of artificial intelligence, that promote the extraction of millions and even billions of representations in images containing facial characteristics. The model developed in this work consisted in a siamese neural network, formed by a pair of identical convolutional neural networks used to calculate the difference between images. The model was developed through the TensorFlow library, recently released by Google, and it was trained with approximately 700,000 images. It was possible to obtain 84.78% accuracy in the MUCT dataset. However, as the specific application for this work consisted in enabling the access of users in a restricted environment, upon previous registration, for a real time system, this projects also sought to use a pretrained neural network, which consisted in a implementation of the ResNet model, trained with 3 million images, available in a library for the Python programming language called Face Recognition. Through this method, it was possible to reach a 97.81% accuracy on the MUCT dataset. Thus, it was presented two different approaches for facial identification, and using the Raspberry Pi microcomputer, this project resulted in the simulating of the access of different individuals in a controlled environment, through the research and the development of neural networks with applications in computer vision and facial recognition.