Um sistema de reconhecimento facial escalável baseado em aprendizado profundo para aplicações em tempo real

Abstract

This work discusses the growing relevance of multi-camera surveillance systems, emphasiz ing the need for effective identification of individuals. Facial recognition systems detect and quickly identify and locate people, helping to prevent incidents. Facial recognition, particularly using Deep Learning techniques, appears as a promising solution. However, challenges such as scalability and resilience still persist. It proposes the use of Vector Databases, implementing HNSW, to improve efficiency in the face matching stage. The processing steps around the models were integrated into the models themselves, ensuring batch processing of steps previously carried out in a serial manner. Scaling and monitoring tools were solutions to ensure system resilience. The objective is to develop an efficient, resilient and low-cost facial recognition system, based on deep learning and distributed learning. This study presents a structure that encompasses concepts, system construction methods and test results. The tests demonstrated that the increase in demand was resolved through the scalability of the processing steps. Code and models are avaliable at <https://github.com/vilsonrodrigues/face-recognition>.