Facial Expression Recognition in Videos using Dynamic Kernels.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition in real-world videos requires analyzing local facial movements.
• Understanding expression formation from diverse facial parts is crucial for accurate recognition.

Purpose of the Study:

• To propose a novel dynamic kernel-based representation for facial expressions.
• To assimilate local facial movements using spatio-temporal representations within a universal Gaussian mixture model (uGMM).

Main Methods:

• Developed a dynamic kernel-based representation for facial expressions.
• Utilized local spatio-temporal representations and a universal Gaussian mixture model (uGMM).
• Evaluated three dynamic kernel types: explicit mapping, probability-based, and matching-based on MMI, AFEW, and BP4D datasets.

Main Results:

• Probability-based kernels demonstrated the highest discriminative power for facial expression recognition.
• Matching-based kernels offered superior computational efficiency compared to other dynamic kernel types.
• The proposed dynamic kernel representation effectively handles global context changes while preserving local similarities.

Conclusions:

• Dynamic kernel representation, particularly probability-based kernels, enhances facial expression recognition accuracy.
• Matching-based kernels provide an efficient alternative for real-time applications.
• The uGMM framework effectively integrates local facial movement statistics for robust expression analysis.