N-Step Pre-Training and Décalcomanie Data Augmentation for Micro-Expression Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Facial expressions are categorized into micro-expressions (brief, low-intensity) and macro-expressions (longer duration).
• Collecting sufficient micro-expression data is challenging due to the need for emotion suppression during video observation.
• Insufficient data can lead to decreased performance in machine learning models trained for micro-expression recognition.

Purpose of the Study:

• To address the data scarcity issue in micro-expression datasets.
• To propose novel methods for enhancing micro-expression recognition model performance.
• To improve the accuracy and robustness of models dealing with subtle emotional cues.

Main Methods:

• N-step pre-training: Employing multiple transfer learning steps from action recognition datasets to facial expression datasets.
• Décalcomanie data augmentation: A novel technique utilizing facial symmetry to generate synthetic data by combining facial halves.

Main Results:

• The proposed N-step pre-training effectively transfers knowledge from related domains, mitigating data limitations.
• Décalcomanie data augmentation successfully expands the micro-expression dataset by leveraging facial symmetry.
• Both methods collectively overcome the data shortage problem, leading to significantly improved model performance.

Conclusions:

• The study demonstrates the efficacy of N-step pre-training and Décalcomanie augmentation for micro-expression analysis.
• These techniques provide a viable solution for training high-performance models with limited micro-expression data.
• The findings contribute to advancing the field of emotion recognition through more robust and data-efficient methods.