Combining GAN with reverse correlation to construct personalized facial expressions.

Area of Science:

• Computer Vision
• Cognitive Science
• Affective Computing

Background:

• Deep learning enables facial expression manipulation but lacks personalization and struggles with non-basic emotions.
• Current methods rely on large labeled databases, limiting expression range and complexity.
• Existing technologies cannot model nuanced or personalized emotional expressions.

Purpose of the Study:

• To develop a novel, interdisciplinary approach for personalized facial expression generation.
• To overcome limitations of existing methods in handling non-basic emotions and data dependency.
• To enable fine-grained control over facial expression synthesis.

Main Methods:

• Combining Generative Adversarial Networks (GANs) with psychophysical reverse correlation.
• Utilizing reverse correlation to extract personalized mental representations of facial expressions.
• Developing a system that does not require large labeled databases or expert knowledge.

Main Results:

• Successful generation of personalized facial expression prototypes for basic and non-basic emotions.
• Demonstrated ability to manipulate facial expressions using generated prototypes.
• Introduced concepts of dominant and complementary action units to describe expression prototypes, challenging universality.

Conclusions:

• The proposed interdisciplinary approach effectively generates personalized facial expressions, including complex emotions.
• The method overcomes the reliance on large datasets and offers fine-grained control.
• The findings suggest a new way to represent and generate facial expressions, moving beyond universal prototypes.