Abstract
Micro-expression recognition faces significant challenges due to the scarcity of labeled data and the subtlety of facial movements. This paper presents a multiview representation framework that leverages transfer learning from macro-expressions (auxiliary domain) to micro-expressions (target domain). By projecting features into a common space and employing sparse dictionary reconstruction alongside multiview feature selection, the proposed method effectively addresses the small sample problem. Experimental results on benchmark databases demonstrate superior performance compared to existing state-of-the-art techniques.
Introduction
Micro-expressions are fleeting, involuntary facial movements that reveal suppressed emotions, typically lasting less than 1/25th of a second. Their detection is vital for fields like clinical psychology and security. However, automated recognition is hindered by the Small Sample Problem—the high cost of labeling makes datasets small—and Feature Uncertainty, as no single descriptor perfectly captures these subtle movements. This research explores whether abundant macro-expression data can serve as a "mentor" to improve micro-expression recognition.
Methodology
The proposed framework utilizes a three-stage pipeline to bridge the gap between macro and micro expressions.
1. Feature Projection and Dictionary Reconstruction
To handle the domain shift between high-intensity macro-expressions and subtle micro-expressions, the model projects features from both domains into a shared common space. A dictionary of representative patterns is learned for macro-expressions and then used to linearly reconstruct the micro-expression dictionary. This process relies on sparse representation techniques to identify core commonalities.
2. Multiview Feature Selection
Rather than relying on a single feature type, the framework extracts four distinct descriptors: LBP-TOP, LOCP-TOP, VLBP, and Optical Flow. To avoid the "curse of dimensionality" and noise associated with simple concatenation, the model employs a multiview representation strategy. Specifically, an $F$-norm block sparse selection process is used to identify the most representative features across these different "views."
Experimental Results
The framework was evaluated using the CK+ database for macro-expressions and the CASME/CASME II databases for micro-expressions, focusing on happiness, sadness, and surprise.
- Performance: On the CASME II dataset, the framework achieved a recognition rate of 78.2%, outperforming the FDM method (68.5%).
- Multiview Advantage: The multiview approach (77.0% accuracy) significantly outperformed single-feature baselines like Optical Flow (61.7%), proving that intelligent feature selection is superior to simple stacking.
- Efficiency: Sparse learning reduced computational overhead while enhancing accuracy by filtering out redundant information.
Conclusion
This multiview transfer learning framework successfully utilizes macro-expression data to overcome the data scarcity in micro-expression research. While the results are state-of-the-art, future work will focus on automating dictionary size selection and improving robustness against real-world environmental factors like lighting and head pose.
Summary updated