Expression-Based Depression Recognition Combined With Attention Mechanism

Depression is a common mental illness,which has caused great harm to patients,families and society.The traditional way of diagnosing depression is highly dependent on the doctor’s experience and the patient’s cooperation,and has strong subjectivity.Therefore,the development of objective and effective depression recognition methods has become a current research hot-spot.Among them,facial expression-based depression recognition has the advantages of non-invasive,non-contact,easy access,and low cost,which has attracted the attention of many researchers.At present,one of the main problems facing the research on depression recognition based on facial expressions is that the constructed features lack the ability to represent depression,resulting in low accuracy of depression recognition.In order to solve this problem,from the perspective of constructing effective features and reducing redundant information,this paper proposes a depression recognition model based on key spatiotemporal information and a depression recognition model based on local relational attention,and verifies it on self-built and public datasets the validity of the two models.The main innovations and contributions of this paper are as follows:(1)A depression recognition model based on key spatio-temporal information(LMB+Temporal-Attention)is proposed.The model divides facial regions of interest based on prior knowledge,then extracts segment-level and video-level spatio-temporal features based on key facial regions,and finally focuses on salient temporal features through improved temporal attention.The model strengthens the focus on key spatiotemporal information,removes redundant information,and improves the recognition effect of depression.The results of ten-fold cross-validation show that the accuracy,precision,recall,and F1 score of the model are 0.757,0.767,0.786,and 0.761,respectively,achieving the current optimal performance.(2)A part and relation attention-based depression recognition model(PRA-Net)is proposed.The model divides the middle-level feature map to obtain semantic-rich local features,and then uses self-attention and relation attention to calculate the local feature weights and re-tuning,and finally aggregates to form a compact depression representation.The model focuses on the spatial distribution of depression cues,constructs an effective representation of depression,and improves the recognition performance of the model.The results of ten-fold cross-validation show that the accuracy,precision,recall,and F1 score of the model reach 0.736,0.772,0.786,and 0.741,respectively.This model utilizes spatial relationship information in an end-to-end manner to achieve recognition performance comparable to the model in(1)above,while reducing engineering complexity.This paper starts with how to reduce the redundancy of features,proposes a depression recognition model based on key spatio-temporal information and a depression recognition model based on local relational attention,constructs a discriminative depression representation to predict depression,and verifies this paper through experiments Effectiveness of the proposed model.