StressNet: A Deep Neural Network Based on Dynamic Dropout Layers for Stress Recognition

Stress is a body response to the changing of environmental conditions, such as facing time pressure, threats, or scary things. Being in a stressful state for a long time affects our physical and mental health. Therefore, we need to regularly monitor our stress. In this paper, we propose a deep neural network with novel dynamic dropout layers to address the stress recognition task through thermal images. Dropout regularization has been widely used in various deep neural networks for combating overfitting. In the task of stress recognition, overfitting is a common phenomenon. Our experiments show that our proposed dynamic dropout layers speed up both the training process and alleviate overfitting, but also make the network focus on the important features while ignoring unimportant features at the same time. The proposed approach was evaluated in comparison with the baseline models [5, 10] over the ANUStressDB dataset. The experimental results show that our model achieves 95.8% classification accuracy on the test set. The code publicly is available at https://github.com/onehotwh/StressNet.