Improved Automatic Stress Detection Through Compound Facial Expressions Using Multi Scale Residual Network

This study proposes an approach to automatic stress detection using compound facial expression images. A deep learning framework is employed to develop automatic stress detection. The proposed approach utilizes compound facial expressions provided in the iCV-MEFED dataset which is a comprehensive compound facial expressions image dataset. The dataset is used to carry out a psychologist supervised mapping of facial expressions that are indicators of stress. The stress indicator facial expression classes are trained and analyzed using a Multiscale Residual Network (MSRN). The study also investigates the inclusion of contempt emotion related compound facial expressions as potential stress indicators. The results were compared with the experiments carried out on stress classification without contempt as well. Experiments using the proposed approach are implemented using a 5-fold cross validation and evaluated using the metrics of accuracy, precision, specificity, sensitivity, and F1-score. The proposed approach achieved accuracy of up to 78.06%. The findings of this study demonstrate potential efficiency of the use of compound facial expressions for automatic stress detection. Also, the study highlights the challenges and complexities of compound facial expressions' classification. The paper forwards a potential contribution to the growing field of human emotion recognition and its applications in psychological analysis.