Micro-Expression Analysis Based on Self-Adaptive Pseudo-Labeling and Residual Connected Channel Attention Mechanisms.

Area of Science:

• Computer Science
• Artificial Intelligence
• Psychology

Background:

• Micro-expressions reveal genuine emotions, crucial for fields like psychotherapy and criminal interrogation.
• Current pseudo-labeling methods for micro-expression analysis suffer from inaccurate labeling due to sliding windows and neglect subtle features by focusing on overall ones.

Purpose of the Study:

• To propose a novel micro-expression analysis method, the Spot-Then-Recognize Method (STRM), that addresses limitations in existing approaches.
• To improve labeling accuracy by dynamically assigning pseudo-labels based on micro-expression proportion within video sequences.
• To enhance the extraction of subtle micro-expression features through a specialized network architecture.

Main Methods:

• Developed the Self-Adaptive Pseudo-labeling Method (SAPM) for dynamic and accurate pseudo-label assignment.
• Designed the Multi-Scale Residual Channel Attention Network (MSRCAN) for effective extraction of subtle micro-expression features.
• Integrated MSSN, Spotting Network, and Recognition Network within MSRCAN for refined feature processing.

Main Results:

• STRM achieved a significant overall performance of 58.24% on micro-expression analysis.
• Demonstrated a 19.62% improvement over existing methods.
• Achieved a 1.51× gain compared to the baseline across multiple datasets.

Conclusions:

• The proposed STRM method significantly enhances micro-expression analysis accuracy.
• SAPM and MSRCAN effectively address limitations in labeling accuracy and feature extraction.
• STRM shows superior performance on both short and long video datasets, outperforming existing techniques.