Cost-Sensitive Two-Stage Depression Prediction Using Dynamic Visual Clues

详细信息    查看全文

• 刊名：Lecture Notes in Computer Science
• 出版年：2017
• 出版时间：2017
• 年：2017
• 卷：10112
• 期：1
• 页码：338-351
• 丛书名：Computer Vision ? ACCV 2016
• ISBN：978-3-319-54184-6
• 卷排序：10112

文摘

This paper presents a novel and effective approach to depression recognition in the visual modality of videos, which automatically predicts the depression level through two cost-sensitive stages. It delivers an improved solution in two ways compared with other vision based methods. On the one hand, current techniques regard depression recognition as either a classification or a regression problem, which tends to incur overfitting due to the high complexity of the model and the limited number of training samples. To handle such an issue, we propose a two-stage framework consisting of a coarse classifier and a fine regressor. The former makes use of a set of linear functions, corresponding to different depression intensities, to approximate the complex non-linear model, where a coarse range of the test sample is preliminarily located. The latter then predicts its precise depression level within the given range. On the other hand, depression recognition is different from the general classification and regression tasks, since its analysis is cost-sensitive as the diagnosis of heart diseases and cancers. However, this critical cue is not taken into account in the previous investigations, thus making their results problematic. To address this drawback, we embed the indicator of medical risk assessment into both the two stages by constraining the classifier using a weight matrix and loosening the regressor to an expanded range of depression level. The proposed method is evaluated on the Audio and Video Emotion Challenge (AVEC) 2013, and the performance is superior to the best one so far reported using the visual modality. Furthermore, it proves complementary to the audio based methods, and their joint use further ameliorates the accuracy. These facts clearly highlight the effectiveness of the proposed method on depression recognition.