Lightweight ViT Model for Micro-Expression Recognition Enhanced by Transfer Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Micro-expressions are subtle emotional cues vital for applications in interrogation and healthcare.
• Current neural network approaches for micro-expression recognition often require substantial computational resources.
• Vision transformers offer competitive accuracy but face challenges with parameter efficiency without image-specific biases.

Purpose of the Study:

• To develop an efficient and accurate micro-expression recognition system.
• To optimize the MobileViT model for micro-expression recognition using transfer learning.
• To address the computational demands of existing micro-expression recognition methods.

Main Methods:

• Combined CASME II, SAMM, and SMIC datasets, including macro-expression samples for pre-training.
• Utilized transfer learning to train the MobileNetV2 block of MobileViT as a facial expression feature extractor.
• Fine-tuned the MobileViT model and employed grid search for hyperparameter optimization, followed by SVM classification.

Main Results:

• Achieved an accuracy of 84.27% in micro-expression recognition.
• Demonstrated a processing time of only 35.4 ms per sample.
• Showcased comparable accuracy to state-of-the-art methods with improved efficiency.

Conclusions:

• The proposed fine-tuned MobileViT approach offers a computationally efficient solution for micro-expression recognition.
• This method balances high accuracy with reduced processing time, enhancing practical applicability.
• Transfer learning and model optimization are effective strategies for improving micro-expression recognition systems.