Research On Depression Recognition Based On Facial Expression

In recent years,the development of depression-assisted diagnosis based on facial videos has been rapid due to its advantages of objectivity,efficiency,low cost,and the absence of the need for subjects to interact with others.However,due to individual expression differences and the diverse nature of depressive behaviors,most existing algorithms struggle to effectively capture the irregular distribution of depressive features in videos while ignoring the problem of interference from non-depressive features.In this paper,the following work is conducted:(1)In response to the insufficient data on depression based on facial videos,this study collected and constructed a facial video database for depression identification research.Facial expression recognition and quantitative analysis were conducted on video data of individuals with depression and a control group from the database to validate the effectiveness of the dataset and explore the characteristics of facial emotional expressions in the two groups.(2)To effectively capture features related to depression,this study constructed a long-short-term difference network model tailored for depression detection.The model employed a two-stage video learning strategy,enabling it to learn facial representations associated with depression over a longer temporal range.In the shortterm facial behavior modeling stage,the study designed a short-term feature extraction module based on image differencing and attention mechanisms.This module learned depression-related facial spatiotemporal features across multiple short-term intervals.In the long-term facial behavior modeling stage,a Long ShortTerm Memory(LSTM)network was employed.This network summarized all the short-term facial features at the video level and further learned their dependencies.By learning the behavioral patterns of individuals with depression in both short-term and long-term contexts,the model achieved depression identification.(3)To mitigate the interference from irrelevant features in depression detection,this study proposes a feature enhancement module.This module is designed based on mutual attention and global attention mechanisms to assign different weights to the extracted short-term facial spatiotemporal features.This enhances the expressive power of depression-related facial features while reducing the interference from irrelevant features,thereby achieving more effective depression identification.Extensive experiments were conducted in this study on the self-collected dataset NCUDID and the publicly available dataset AVEC2014.The results demonstrate that the proposed depression identification framework based on facial emotional expressions achieves high accuracy and reliability.It is capable of distinguishing individuals’ depressive states and serves as an auxiliary diagnostic tool for depression.