Research On Depression Detection Based On Multiple Instance Learning Of Facial Expressions

Depression is a serious mental health disorder that poses a significant threat to individuals and families in today’s society.Early detection and treatment are crucial for preventing and managing depression.However,the current ratio of psychiatrists to patients is too low to meet the growing demand for medical care among depression patients.Therefore,finding an objective and effective method for identifying depression is an urgent issue in the field of depression detection research.In this regard,this article focuses on facial expressions as an objective behavioral signal for detecting depression.In the current facial expression-based depression detection work,each single frame or short time-sequences clip instance that makes up the video is assigned the same category label.Since it is difficult for us to determine which instances reflect information related to the category label,instances that are not related to the category label will negatively affect the supervised learning and thus affect the model performance.In addition,many depression detection methods obtain spatiotemporal information by extracting facial expression changes in short time sequences,while ignoring the behavioral representation of subjects under long time sequences information.Long-term sequence information contain contextual information of video sequences,which can better describe depressive behaviors.In this paper,aiming at the deficiencies in the above research on facial expression-based depression detection,multiple instance learning is applied to deep neural network modeling,and a more reasonable and effective depression detection model is proposed.The main work and innovations of this paper are as follows:(1)At present,the facial expression data used for depression detection only contains coarse-grained depression information labels,and the training method of supervised learning may reduce the performance of the model.Therefore,this paper proposes a depression detection model based on multiple instance learning.The method uses a convolutional neural network to extract emotion representations from facial images within each video frame.Subsequently,these representations are aggregated by dual attention pooling,and the bag-level feature information for the time window is obtained,which is used for the classification of the final model.This method achieved an accuracy rate of 0.727 and a recall rate of 0.729 on the 150 balanced sample data sets constructed in Chapter 3 of this paper.Compared with the supervised learning method that does not include a multiple instance learning module,the accuracy rate and recall rate improved by 4.98%and 7.18%,respectively.It shows that the multiple instance learning method has better robustness and generalization ability in the depression detection task.(2)In order to capture long-term behavioral representations in the facial expressionbased depression detection task,this paper proposes a dual-stream multiple instance learning method.The method first uses the max pooling method to find the key instance that best represents the depression information,and then obtains the attention weight by calculating the distance between other instances and the key instance.Finally,by aggregating attention weights with instance representations,a long-term video bag-level representation is obtained.This method can mine the internal relationship between facial expressions and long-term depression category labels,and has the ability to represent the emotional changes of the subjects.Compared with other latest depression detection research results,the method proposed in this paper achieved 0.747 accuracy and 0.745 recall rate on the constructed data set,improving the baseline method by 15.63%and 15.14%,respectively.This paper proposes weakly supervised methods for multiple instance learning to identify depression through raw facial expressions.Compared with supervised learning training methods,our method can make better use of instance data and has higher robustness and generalization.At the same time,the methods proposed in this paper can be applied to the detection of other mental diseases,providing new ideas and methods for related research.