Avaliação dos métodos FaceNet e LBPH de reconhecimento facial através de uma detecção híbrida de Yolo com MTCNN

Abstract

There is currently a growing demand for safety and comfort, which is reflected in our daily lives and in our surroundings. Whether it is to unlock a cell phone or enter a restricted access area, biometric readings are increasingly present in our daily lives, and the expectation is that technology will evolve to make these methods reliable, practical and non-invasive. This fits well when we enter the area of computer vision and come across facial recognition, an area that works on identifying people, whether in photos, videos or even in real time, and which has great potential in security, authentication and monitoring. However, this area also contains a series of challenges. Lighting, position, angle, accessories that can hide parts of the face and even aging can present difficulties for a facial recognition system. To overcome these obstacles, a series of promising studies have brought modern and sophisticated methods based on machine learning that have proven to be robust and efficient systems. Among these systems, there is a more classical model known as Local Binary Pattern Histograms, which uses the comparison of pixels with their edges to transform images into histograms and compare them to identify people. Another very promising method uses convolutional neural networks to analyze the most important features of a face across all its neurons and generate highly discriminative representations, as is the case with FaceNet. To evaluate these models, this study will test both in different situations and provide results showing how technology has advanced and how these recent convolutional neural network methods are proving increasingly effective in dealing with the various obstacles of facial recognition.