Department of Computer Engineering

The face is a powerful channel for non-verbal communication. Anatomically-based facial action units (AUs), both individually and in combinations, convey a wide range of facial expressions, emotions, and intentions. Traditional human-observer methods for measuring these actions are labor-intensive, qualitative, and impractical for real-time applications or large datasets, highlighting the need for automated, objective, reliable, and efficient approaches. Additionally, synthesizing realistic facial expressions—using advanced techniques like diffusion models—is essential for generating large, balanced datasets and training personalized models. However, these generative models come with limitations, particularly regarding training speed and the overall quality of the generated outputs. In this talk, I will present my work on deep learning-based methods for automated facial action detection and synthesis, with a focus on improving detection performance and generating more realistic images efficiently. I will also present several applications of our models in detecting and monitoring obsessive-compulsive disorder and analyzing parent-child interactions to explore child development and anxiety-related behaviors, among others.