Stress Detection using Facial Emotion Recognition: Comparative Analysis of VGG-16, ResNet-50, and a Custom CNN Architecture

Stress is an important psychological factor that can significantly impact overall health and productivity; hence, timely detection and management of stress are essential to prevent adverse effects on mental and physical well-being. In this study, facial emotion recognition is employed to detect stress using the CK+ dataset, which is a widely used dataset for facial expression recognition, containing anger, disgust, contempt, fear, happiness, sadness, and surprise emotions. Initially, well-established architectures VGG-16 and ResNet-50 are tested on the CK+ dataset. Although these models delivered competitive accuracies, a custom hybrid model, based on Convolutional Neural Networks (CNN), achieved superior performance. The custom model used fewer parameters leading to faster training time while still maintaining a higher accuracy of 99%. The model outperformed already existing solutions in both accuracy and training time making it more effective for real-word stress detection through facial expressions. Notably, the custom model achieved comparatively higher accuracy than ResNet-50 and faster training time compared to VGG-16 model.