<![CDATA[This research develops an emotional detection system for students during learning using YOLOv8 model N and Computer Vision technology. The main objective measures accuracy of facial expression recognition across happy, sad (murung), neutral (normal), and confused classes while evaluating lighting and camera position impacts. Dataset of 1500 Kaggle images was labeled on Roboflow with augmentations including flip, crop, blur, and noise, then trained on Google Colab for 100 epochs using pre-trained YOLOv8n weights. Model validation employed confusion matrices, precision-recall curves, and real-time webcam testing on 10 students at 5-15 cm distances. Key findings show accuracies of 75% (happy), 62% (sad), 69% (neutral), and 33% (confused), averaging 59.75% under >39 lux frontal lighting. Optimal performance occurred at 0°-10° angles, but backlit conditions reduced efficacy by 25-35% due to shadow occlusion. The system enables real-time classroom monitoring with 4.2 ms inference latency, supporting cognitive engagement assessment despite bingung class limitations from dataset imbalance. Future improvements recommend expanded ambiguous samples and lighting augmentations.]]>
Copyrights © 2026
