Micro-expression recognition based on multi-scale 3D residual convolutional neural network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate micro-expression recognition is crucial for applications like psychotherapy and criminal interrogation.
• Challenges include capturing subtle, fleeting facial features and improving recognition performance.
• Existing methods struggle with the effective extraction of micro-expression subtle features.

Purpose of the Study:

• To propose a novel architecture for enhanced micro-expression recognition.
• To address the challenge of capturing weak and fleeting facial features.
• To improve the accuracy of high-precision micro-expression recognition.

Main Methods:

• Developed a multi-scale 3D residual convolutional neural network (3D-ResNet50).
• Utilized micro-expression optical flow feature maps as input.
• Incorporated multi-scale convolutional modules and an attention mechanism for feature fusion.
• Employed a discriminative network structure with multiple output channels.

Main Results:

• Achieved recognition accuracies of 74.6% (SMIC), 84.77% (SAMM), and 91.35% (CASME II).
• Demonstrated substantial improvement over existing mainstream methods.
• Effectively enhanced micro-expression recognition performance and accuracy.

Conclusions:

• The proposed multi-scale 3D residual convolutional neural network offers superior micro-expression recognition.
• The novel architecture effectively integrates spatial and temporal features for improved contextual awareness.
• This work provides a valuable reference for high-precision micro-expression recognition research.