Reconhecimento de expressões faciais em tempo real com YOLO: estudo comparativo e desenvolvimento de um protótipo web.

Abstract

This work addresses automatic facial expression recognition under uncontrolled conditions, a relevant challenge for applications in human-computer interaction, security, healthcare, and real-time interactive systems. In this context, this study investigated the use of models based on the You Only Look Once (YOLO) architecture, which performs detection and classification in a single stage, combining speed and accuracy. The methodology was conducted in two complementary stages. First, four versions of the YOLO family were trained and comparatively evaluated, specifically YOLOv8n, YOLOv10n, YOLOv11n, and YOLOv12n, using the public dataset “9 Facial Expressions for YOLO”, composed of images captured in uncontrolled scenarios and covering nine facial expression categories. The evaluation considered metrics such as precision, recall, mAP, FPS, GFLOPs, and confusion matrices. In the second stage, the selected model was integrated into a web prototype for real-time facial expression recognition, including webcam capture, backend inference, prediction display in the interface, user feedback registration, and data visualization through a dashboard. The results indicated a slight predictive advantage of YOLOv12n in the experiments performed, while YOLOv10n presented the highest inference speed. However, YOLOv8n was selected for the applied stage because it provided the best balance between accuracy, speed, and practical feasibility, achieving an mAP50 of 0.837, an mAP50–95 of 0.681, and 56.17 FPS. The developed prototype demonstrated adequate functionality in the complete flow of capture, inference, prediction display, feedback registration, and aggregated result analysis. Nevertheless, it was observed that more subtle or ambiguous expressions, pose variations, lighting conditions, and camera quality still influence the perceived performance of the system. Therefore, this work contributes both with a comparative analysis of YOLO architectures applied to facial expression recognition and with the implementation of a functional prototype for real-time inference.