A privacy-preserving framework for automatic facial emotion dynamics analysis: a case study in depression detection
Type: dissertation
Abstract: This thesis investigates whether facial emotion dynamics features extracted from videos using a privacy-preserving methodology can serve as reliable indicators of depression.A facial emotion recognition framework was developed using two Neural Network architectures (MLP and FC+RNN) with a grid search over parameter and hyperparameters.Approximately 2000 models were trained and evaluated on the CREMA-D, RAVDESS, and combined datasets.Both MLP and RNN architectures achieved state-of-the-art performance in cross-validation, while the RNN showed stronger generalization in cross-corpus evaluations.The best models on the combined dataset were used to compute dynamic features using E-DAIC dataset, including intensity, variability, instability, inertia, and cross-lag.In line with prior research, emotional cross-lag, intensity, and variability supported the predicted patterns for depression assessment.These findings demonstrate that facial emotion dynamics can be automatically extracted in a privacy-preserving manner and reliably distinguish depressed from non-depressed participants.
Source: Summit (Simon Fraser University)
Author: Zahoor, Maryiam
Institution: Simon Fraser University
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7139197717
