Depression recognition based on dynamic facial and vocal expression features using partial least square regression

Depression is a typical mood disorder, and the persons who are often in this state face the risk in mental and even physical problems. In recent years, there has therefore been increasing attention in machine based depression analysis. In such a low mood, both the facial expression and voice of human beings appear different from the ones in normal states. This paper presents a novel method, which comprehensively models visual and vocal modalities, and automatically predicts the scale of depression. On one hand, Motion History Histogram (MHH) extracts the dynamics from corresponding video and audio data to represent characteristics of subtle changes in facial and vocal expression of depression. On the other hand, for each modality, the Partial Least Square (PLS) regression algorithm is applied to learn the relationship between the dynamic features and depression scales using training data, and then predict the depression scale for an unseen one. Predicted values of visual and vocal clues are further combined at decision level for final decision. The proposed approach is evaluated on the AVEC2013 dataset and experimental results clearly highlight its effectiveness and better performance than baseline results provided by the AVEC2013 challenge organiser.