Images of Stress: Investigating Image Encodings and Transfer Learning for Stress Detection

Stress is a widespread phenomenon in our modern society. Each individual experiences stress differently, and a person's prolonged exposure to it may seriously affect their health. It is therefore essential to identify high stress scenarios and develop techniques to increase stress resilience. This paper focuses on stress detection based on physiological signals, namely the blood volume pulse, which can be measured using of the shelf commercial devices. In recent research many deep learning approaches have been introduced within the field of stress and emotion detection that have led to remarkable performance gains. This work explores the performance of benchmark deep learning architectures for stress classification using image encodings of the blood volume pulse time series. However, data and label scarcity pose a major challenge. Transfer learning is an often used technique to counteract this problem. Still, several open questions remain regarding the choice of well performing computer vision models and image encodings, as various researchers come to different conclusions. This paper hence focuses on a detailed comparison analysis of different pre-trained convolutional neural network architectures and several image encodings, such as the symmetric projection attractor reconstruction. To the best of our knowledge, the latter has not been considered for stress detection yet. While our analysis of the model's prediction quality for the often used WESAD dataset reveals a significant inter-subject variability, EfficientNetV2L, using scalogram image encodings as input, achieves the most stable classification results and outperforms the other networks regarding prediction quality.