Deep Neural Networks for Depression Recognition Based on Facial Expressions Caused by Stimulus Tasks

Abstract

With the growth of the global population, the proportion of individuals with depression has rapidly increased; it is currently the most prevalent mental health disorder. Although existing studies on depression have mainly examined the several databases, which comprise facial images and videos of non-Chinese subjects, there are few effective databases for a Chinese population. In this study, we first create a depression database by asking participants to perform five mood-elicitation tasks. After each task, their facial expressions are collected via a Kinect. In the depression database, the facial feature points (FFP) and facial action units (AU) are obtained. We build a range of deep belief network (DBN) models based on FFPs and AUs to extract facial features from facial expressions, named 5DBN, AU-5DBN and 5DBN-AU. We evaluate all proposed models in our built database, and the results demonstrate that (1) the recognition performance of the AU-5DBN model is higher than that of the 5DBN-AU model, and that of the single feature model is the lowest; (2) The performance of depression recognition in the positive and negative emotional stimuluses are higher than that of neutral emotional stimulus; (3) The classification rate for females is generally higher than that for males. Most importantly, the constructed database is from a real environment, i.e., several psychiatric hospitals, and has a certain scale. The experimental results show higher recognition performance in the database; thus, the proposed method is validated as effective in identifying depression.