Predicting stress levels for smartphone users using transfer learning induced residual net

Smartphones have become an essential part of our daily lives, especially during the COVID-19 pandemic when most people are forced to stay at home. This resulted in increased reliance on smartphones for education. Restricted body movements and continuous looking at phone-screens started gradually degrading the mental health especially for children and teenagers. The real-time measurement of stress levels (SLs) generated due to chronic exposure to screens becomes interesting study for the psychologists. Still, there exist void in linking the outcomes found by psychologists with automatic facial-expression-classification (FEC) from technical back-ground. In the current study, the phone users' SL is recognized by analyzing their facial expressions using Convolutional Neural Network. But it is observed that the accuracy gets saturated after certain iterations. This problem is addressed by introducing skip-connections into the architecture and implementing Residual Network. After successful completion of the SL classification, the authors delve into analyzing the training and validation errors with respect to human level performance. It is found that the system is suffering from overfitting. One way to address this issue is to feed more data to the system by Transfer Learning. The proposed work has a potential to open up a new avenue of SL measurement.

Automated summary

Smartphone usage has become integral to our daily lives, but prolonged screen exposure may have negative effects on mental health. This study utilizes a Convolutional Neural Network to analyze facial expressions and recognize users' stress levels, while also addressing the issue of overfitting.