Dep-Emotion: Suppressing Uncertainty to Recognize Real Emotions in Depressed Patients.

In depression, affective and emotional dysfunction are important components of the clinical syndrome. At present, doctors mainly judge the real emotions of depressed patients through the naked eye, with a strong subjective consciousness. We collected images of seven expressions voluntarily imitated by 168 subjects, and then recruited 9 raters to recognize these images. The study found that depressed patients have deficits in Facial Emotion Expression, resulting in great uncertainty in their facial expressions. Therefore, we propose the Dep-Emotion to solve this problem. For the depression expression dataset with uncertainty, we use Self-Cure Network to correct the sample label to suppress the uncertainty. At the same time, the input part and downsampling block of ResNet18 are adjusted to better extract facial features. The input image is regularized by Cutout, which enhances the generalization ability of the model. The results show that Dep-Emotion achieves the best accuracy of 40.0%. The study has important implications for automatic emotion analysis and adjunctive treatment of depression.