TL;DR
This paper introduces a novel "Two-Level Spatio-Temporal Feature Fused Two-Stream Network" (TFT) for micro-expression recognition. By effectively fusing multi-level spatial and temporal features from dedicated encoders (ResNet for spatial, Swin Transformer for temporal), the TFT achieves state-of-the-art performance on challenging datasets, outperforming previous methods by a significant margin.
What Problem Does This Paper Address?
Micro-expressions (MEs) are fleeting, involuntary facial movements that reveal true emotions, making them crucial for understanding human affect. However, their short duration and subtle nature pose a significant challenge for automated recognition. Existing deep learning approaches, particularly those using Vision Transformers (ViTs), often involve computationally intensive processes and struggle to capture the fine-grained, multi-level spatio-temporal information essential for accurate ME detection. Traditional two-stream architectures also fall short by only utilizing the deepest features for classification, neglecting valuable information from earlier network stages.
Methods
The proposed TFT architecture tackles these challenges with a two-stream design:
- Spatial Encoder (Modified ResNet): This stream focuses on extracting static texture features from the apex frame of a micro-expression. A modified ResNet architecture is used, optimized for computational efficiency and adapted to align with the temporal encoder's stages.
- Temporal Encoder (Swin Transformer): This stream captures the dynamic facial muscle movements by processing optical flow information. The Swin Transformer is chosen for its efficient, hierarchical approach to self-attention, which balances local and global feature learning while significantly reducing computational complexity compared to standard ViTs.
- Two-Level Feature Fusion: A novel fusion algorithm is introduced to integrate multi-level spatio-temporal features from both streams. This algorithm leverages the similarity in output feature shapes across different stages of the encoders, allowing for a more comprehensive capture of subtle ME details.
- Temporal Aggregation (Weighted Average Operator): Recognizing that the apex frame holds the most salient information, a weighted average operator is employed to aggregate temporal features. This operator assigns higher weights to frames closer to the apex, effectively capturing the most informative parts of the ME sequence.
- Transfer Learning: To overcome the scarcity of ME data, the model utilizes transfer learning, pre-training both encoders on the CK+ dataset before fine-tuning on ME datasets. This approach ensures consistency in input and preprocessing, leading to better adaptation of learned features.
Datasets
The proposed method was evaluated on three widely used spontaneous micro-expression datasets:
- SMIC-HS (Spontaneous Micro-Expression Corpus - High Speed): A dataset of high-speed spontaneous micro-expressions.
- CASME II (Chinese Academy of Sciences Micro-Expression II): Another prominent dataset for spontaneous micro-expression recognition.
- SAMM (Spontaneous Actions and Micro-Movement database): A dataset containing spontaneous micro-movements.
The evaluation was conducted using a combined corpus from these datasets, relabeled into three classes (negative, positive, and surprise) for a comprehensive assessment.
Key Insights and Ideas
- Multi-level Fusion is Crucial: The paper strongly emphasizes that simply fusing features from the deepest layers of a two-stream network is insufficient for ME recognition. Integrating features from multiple stages is vital for capturing the subtle, fine-grained information characteristic of MEs.
- Swin Transformer's Efficiency: The adoption of the Swin Transformer for temporal feature extraction is a key innovation, offering a significant reduction in computational cost compared to ViTs while maintaining or improving performance. This is attributed to its shifted window mechanism.
- Weighted Temporal Aggregation Matters: The simple mean operator is outperformed by a weighted average operator that prioritizes apex frames. This highlights the importance of understanding the temporal dynamics of MEs and assigning importance accordingly.
- Transfer Learning for Data Scarcity: The success of transfer learning from macro-expression datasets (CK+) demonstrates a practical strategy to mitigate the challenge of limited ME data.
Key Results
The proposed TFT architecture achieved state-of-the-art results across the evaluated datasets. Notably:
- It achieved an Unweighted Average Recall (UAR) exceeding 0.905 on CASME II, placing it among the top-performing frameworks in the literature.
- The feature fusion algorithm improved both the F1 score and UAR by approximately 4% compared to simpler fusion methods.
- The TFT demonstrated significantly shorter inference times (1.423s for four video samples) compared to a previous state-of-the-art method (SLSTT-LSTM, 2.363s).
Limitations and Future Work
The authors acknowledge a few limitations:
- Cumbersome Training Process: The pre-training and fine-tuning stages require careful management of input formats and optical flow extraction methods to maintain consistency.
- Reliance on Apex Frames: Like many existing ME recognition methods, the proposed approach relies on apex frames. The integration of apex frame detection directly into the recognition network is suggested as a future research direction.
- Real-World Gap: While evaluated on established datasets, the authors note that these datasets are collected under strict experimental conditions, and a gap exists between these controlled environments and real-world scenarios. Future work could explore more robust methods for real-world applications.
The paper suggests exploring more efficient training strategies and investigating methods that better integrate apex frame detection with the recognition network.
Summary updated