ScholarWorks@숙명여자대학교

초록

Automatic depression recognition is essential to depression diagnosis. In this paper, we investigate the problem of depression recognition from facial images, each of which is labeled with one Beck Depression Inventory (BDI-II) score. Because of the ambiguity between one facial image and the depression score, the annotators may not present the accurate score but tend to give those around the ground-truth one. To solve the problem, this paper adopts label distribution to annotate each image, in which each (score) label has a relevance degree. First, we apply the Gaussian distribution to generate the depression score distributions, in which the ground-truth score attains the highest degree, while the neighborhood scores also have degrees to some extent. Thus, each image can contribute to not only its ground-truth score but also neighborhood scores. Second, we generate the depression severity level distributions from the score distributions according to the mapping relationship between BDI-II score and severity level. Finally, we propose a novel method to learn joinT depression scoRE And level distribuTion, termed as TREAT. In the experiments, we compare TREAT with several state-of-the-art methods on three publicly released datasets AVEC 2013, AVEC 2014, and AVEC 2019, and the experimental results justified that TREAT achieves the best performance.

키워드