Facial expression recognition based on multi-domain norm-referenced encoding.

Area of Science:

• Computer Vision
• Neuroscience
• Machine Learning

Background:

• Human facial expression recognition is robust across varied head shapes.
• Current machine learning models struggle with out-of-domain transfer for facial expression recognition without extensive training data.

Purpose of the Study:

• To integrate neuroscience principles into computer vision models for improved facial expression recognition.
• To develop a method that facilitates the transfer of learned facial expressions to novel head shapes with high data efficiency.

Main Methods:

• Proposed a biologically inspired mechanism based on norm-referenced encoding.
• Represented inputs as deviations from a domain-specific reference vector.
• Assumed that deviations from a reference are preserved across domains to enable generalization.

Main Results:

• The proposed model generalized to new head shapes using only a single additional training image.
• Demonstrated generalization abilities and data efficiency on datasets with highly varying head shapes.
• Showcased the scalability and effectiveness of norm-referenced encoding in computer vision models.

Conclusions:

• Norm-referenced encoding, inspired by neuroscience, significantly enhances the generalization capabilities of computer vision models for facial expression recognition.
• This approach offers a data-efficient solution for transferring learned expressions to novel domains.
• Paves the way for broader applications in facial pattern recognition and other related fields.