Facial Prior Guided Micro-Expression Generation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial micro-expression (FME) generation is crucial for augmenting limited labeled datasets.
• Existing image animation methods struggle to capture subtle, short-term facial motion essential for FMEs.

Purpose of the Study:

• To develop a novel framework for facial micro-expression generation by leveraging facial priors.
• To improve the accuracy and realism of synthesized FMEs and enhance micro-expression recognition.

Main Methods:

• A facial-prior-guided framework was proposed, estimating action unit (AU) geometric locations from facial landmarks.
• An adaptive weighted prior (AWP) map was developed to mitigate AU estimation errors and capture subtle motions.
• The facial prior module was integrated into mainstream image animation frameworks to guide motion representation and generation.

Main Results:

• The proposed facial prior module significantly enhances the performance of image animation frameworks for FME generation.
• Experiments on three benchmark datasets validated the effectiveness of the facial prior module.
• The FME generation technique successfully enlarged datasets, improving the performance of action recognition models.

Conclusions:

• The facial-prior-guided framework offers a robust solution for FME generation.
• The method effectively addresses the limitations of existing techniques in capturing subtle facial dynamics.
• This approach contributes to advancing FME research and applications, particularly in dataset augmentation and recognition tasks.