Facial Micro-Expression Recognition Using Double-Stream 3D Convolutional Neural Network with Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Micro-expressions (MEs) are brief, involuntary facial expressions indicating concealed emotions.
• Automatic ME recognition is challenging due to short duration, low intensity, and limited datasets.
• Existing methods struggle with feature extraction and data scarcity.

Purpose of the Study:

• To develop an effective deep learning model for automatic micro-expression recognition.
• To address the challenges of feature extraction and data insufficiency in ME datasets.
• To improve the accuracy and robustness of micro-expression recognition systems.

Main Methods:

• A novel double-stream 3D convolutional neural network (DS-3DCNN) was proposed.
• The model utilizes two streams for spatiotemporal feature extraction and facial motion variation analysis.
• Subtle facial motions were amplified, and a macro-expression dataset augmented training data.
• Supervised domain adaptation bridged the gap between ME and macro-expression datasets.

Main Results:

• The DS-3DCNN model demonstrated superior performance on public ME datasets.
• It outperformed existing state-of-the-art models, achieving over 6% improvement on MEGC2019.
• The model effectively extracts discriminative features from subtle and short-duration facial expressions.

Conclusions:

• The proposed DS-3DCNN offers a promising solution for accurate automatic micro-expression recognition.
• Leveraging macro-expression data and domain adaptation enhances model performance with limited ME data.
• This advancement has significant implications for fields utilizing emotion recognition technology.