Intelligent System Based on YOLOv9 for Emotion Detection in Educational Environments

Abstract

Automatic emotion detection in academic settings presents a significant challenge for the development of intelligent systems focused on student well-being. This study comparatively evaluates YOLOv8 and YOLOv9 on a balanced dataset of over 8,000 facial images across five emotional categories: happiness, neutral, anger, sadness, and surprise. The solution was implemented on a web platform that integrates real-time capture, training, evaluation, and inference using Node.js, React, and FastAPI. Experimental results show that YOLOv9 achieves superior performance, with an accuracy of 0.7110, a relative improvement of 8.01%, a 38.72% reduction in localization error, and a total inference time of 152.1 ms, representing a 44% decrease compared to YOLOv8. Taken together, these findings confirm that PGI and GELAN offer measurable advantages in accuracy and efficiency, particularly in emotion recognition tasks.