Stress Detection from Facial Expressions Using Transfer Learning Techniques

In today's fast paced world stress has become a significant concern, impacting the individual mental health and overall well being. Various models are designed for detecting stress from facial expressions one among them is ResNet-101 architecture that is designed to detect stress in real-time from video surveillance using the symptoms of facial cues with an accuracy of 80.4%. The limitation of the model is that the minute motions are not detected. To overcome these challenges a comprehensive evaluation is made, evaluating the capacity of deep learning architectures in capturing facial cues associated with stress. Transfer learning is a proven technique which reuses the weights of pre-trained model enhancing the model capabilities. In this research project, we propose the development of a real-time stress detection system using pre-trained Mini Xception, and VGG-16 models. Following a number of tests, it was shown that VGG16 performed the best at recognizing tense emotions when combined with a convolutional layer-based classifier with an accuracy of 97.5%.