Otimização do reconhecimento facial em ambiente hospitalar com redes pré-treinadas

Abstract

Facial Recognition has established itself as a robust and essential technology for access control and individual identification, being widely applied in security and authentication contexts. However, factors such as variations in lighting, facial expressions, and capture angles still represent relevant challenges to model accuracy. To address these limitations, this research proposes a Computer Vision methodology focused on facial recognition in a hospital setting, with the goal of validating activity execution by healthcare professionals.Initially, the MTCNN (Multi-task Cascaded Convolutional Networks) architecture was utilized for face detection and alignment in the images. Subsequently, the FaceNet model, accessed through the DeepFace library, was employed to extract facial embeddings, which are compact feature vectors that represent the identity of each individual in a 128-dimensional space. These normalized embeddings served as input for the comparative analysis of two classifiers designed to handle vectorial data: a One-Dimensional Convolutional Neural Network (CNN 1D) and a Multi-Layer Perceptron (MLP).The models were trained on a subset of the LFW dataset (Labeled Faces in the Wild), applying regularization techniques, Dropout, EarlyStopping, ReduceLROnPlateau, and evaluation via metrics such as accuracy, precision, recall, F1-score, FAR (False Acceptance Rate), and FRR (False Rejection Rate). The FaceNet + MLP model demonstrated the best overall performance, achieving 94.32% accuracy, 92.24% balanced accuracy, and an F1-score of 92.38%, indicating a high capacity for generalization and a low error rate. In contrast, the FaceNet + CNN 1D model obtained 80.53% accuracy, with FAR and FRR both equal to 0.1947, showing consistent, yet inferior, performance compared to the MLP, especially in underrepresented classes.Despite not reaching the state-of-the-art results reported for the complete LFW (>99%), the findings of this study prove the viability of combining FaceNet with lightweight classifiers. This approach shows promise for applications in hospital environments, where traceability and identity validation are essential for safety and operational compliance.