Performance Analysis of Emotion Recognition Using YOLO

Abstract

This paper presents facial emotion detection system built using YOLOv8 architecture that recognizes seven distinct emotions: angry, disgust, fear, happy, neutral, sad, and surprise. We trained our model using the FER2013 dataset, converting traditional classification data to an object detection format to leverage YOLOv8's capabilities. Unlike conventional emotion recognition approaches that simply categorize expressions, our system introduces three advanced analysis components: emotion intensity quantification that measures each emotion on a 0-100% scale, explainable AI visualization using Grad-CAM that reveals which facial regions influence the model's decisions, temporal emotion tracking that monitors emotional changes over time. Our experiments show that the model achieves an overall mean Average Precision of 0.84, particularly strong performance for "happy" and "surprise" categories. Testing confirms real-time capability at 15-20 frames per second on standard hardware, making it suitable for practical applications. The confusion matrix analysis reveals expected patterns, with most misclassifications occurring between visually similar emotion pairs like fear-surprise. The system successfully detects emotions in both webcam input and uploaded images, demonstrating robustness to varying conditions. This research contributes a more nuanced approach to emotion recognition that goes beyond binary classification, applications in psychological research, HCI, affective computing where understanding emotional context and intensity is crucial.