Micro-Expression Recognition via LoRA-Enhanced DinoV2 and Interactive Spatio-Temporal Modeling.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Micro-expression recognition (MER) faces challenges due to brief duration, low intensity, and complex spatial frequency patterns.
• Existing MER methods often require extensive preprocessing and significant computational resources.
• Subtle facial texture variations are difficult to capture and analyze effectively.

Purpose of the Study:

• To introduce a novel MER architecture that improves accuracy while reducing computational cost.
• To integrate frequency-domain transformation and graph-based temporal modeling for efficient MER.
• To enhance sensitivity to subtle facial movements through adaptive spatial frequency decomposition.

Main Methods:

• Fine-tuning a large feature extraction model with LoRA (Low-Rank Adaptation) for computational efficiency.
• Implementing a Spatial Frequency Adaptive (SFA) module to decompose and dynamically weight spatial frequencies.
• Utilizing a Dynamic Graph Attention Temporal (DGAT) network, combining Graph Attention Networks and LSTM for temporal modeling.

Main Results:

• The proposed MER architecture achieved superior performance on SAMM, CASME II, and SMIC datasets.
• On the SAMM 5-class setting, the approach yielded an unweighted F1 score (UF1) of 81.16% and an unweighted average recall (UAR) of 85.37%.
• The method outperformed existing approaches by significant margins in UF1 and UAR.

Conclusions:

• The novel MER architecture effectively addresses challenges of brief duration, low intensity, and spatial frequency variations.
• Integrating frequency-domain analysis and graph-based temporal modeling offers an efficient and accurate approach to MER.
• The proposed method demonstrates significant potential for real-world micro-expression analysis applications.