Research On Stress Detection Algorithm Based On Image Sequence

Stress is becoming more and more common in modern society,and its impact on people's life and mental health is increasing.It needs to be discovered and adjusted in time.Therefore,stress detection has important research significance.Traditional stress detection methods based on physiological signals rely on the collection of physiological signals by the device,and the application scenarios are limited.In recent years,with the deployment of more and more contactless cameras in life and the advancement of data collection and analysis technology,image-based stress detection methods have been developed,but there are still the following problems in application deployment: manual extraction of facial features and use shallow machine learning algorithms cannot fully extract stress-related features;the image preprocessing steps of most stress detection algorithms are more complicated,which is not conducive to hardware deployment.Therefore,in view of these key issues,based on the image sequence,this paper studies the stress detection algorithm.The main research work of this paper is as follows:(1)A stress detection algorithm based on convolutional neural network is proposed.Since a single image is used as input,it is not possible to effectively extract the action features related to stress.At the same time,the R channel of the image is related to the heart rate characteristics.Therefore,for the problem of stress-related feature extraction,this paper sets the input as an Rchannel image sequence,and builds a feature extraction module based on local binary convolution to extract pressure-related features more effectively.In addition,this article uses the self-attention mechanism to strengthen the attention to changes in its own characteristics.Since there is currently no public stress data set to study the relationship between stress and human face,the laboratory has cooperated to establish a multi-paradigm and multi-modal pressure data set,including six paradigms,image,voice and heart rate three modalities.The results on this self-built data set show that the algorithm has good results.(2)Aiming at the problem of hardware deployment,a lighter student network that is more conducive to deployment is proposed.The student algorithm network separates the feature extraction of spatial features and temporal features,builds a spatial feature extraction module based on convolutional neural networks,and builds a temporal feature extraction module based on gated recurrent unit.In addition,in order to better perform knowledge distillation,this paper proposes to use the KL divergence loss function as the loss function of the soft label loss.The results of the self-built data set show that the student network can achieve the effect of the teacher network after knowledge distillation,which proves the effectiveness of the student network in pressure detection and the effectiveness of the knowledge distillation method proposed.Finally,the algorithm network was successfully deployed on the hardware platform,and a good pressure detection effect was obtained.