Feature Selection on 2D and 3D Geometric Features to Improve Facial Expression Recognition

Area of Science:

• Computer Science
• Human-Computer Interaction
• Artificial Intelligence

Background:

• Facial expression recognition is crucial for human-computer interaction.
• Selecting relevant geometric features is key for accurate classification of expressions.
• Existing methods often involve large feature sets, impacting computational efficiency.

Purpose of the Study:

• To identify and select optimal geometric features for classifying six basic facial expressions.
• To compare feature selection methods for maximizing classification accuracy while minimizing feature set size.
• To investigate the efficiency of reduced feature sets for facial expression recognition.

Main Methods:

• Proposed an initial set of 89 normalized distances and angles from 22 facial landmarks, inspired by FACS and MPEG-4.
• Applied Principal Component Analysis (PCA) and a Genetic Algorithm (GA) for feature selection.
• Evaluated feature sets using four classifiers on the Bosphorus and UIVBFED datasets.

Main Results:

• PCA yielded 39 features, while GA resulted in 47 features.
• Achieved median accuracies of 86.62% on the Bosphorus dataset and 93.92% on the UIVBFED dataset.
• The GA-derived feature set was the smallest among methods with comparable accuracy.

Conclusions:

• The genetic algorithm provides an effective method for selecting a minimal yet accurate set of geometric features for facial expression recognition.
• Reduced feature sets significantly decrease recognition time, enhancing real-time applications.
• This research contributes to more efficient and accurate human-computer interaction systems.