VaBTFER: An Effective Variant Binary Transformer for Facial Expression Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Transformer models excel at facial expression recognition (FER) by capturing long-range dependencies in facial muscle movements.
• Large Transformer models present deployment challenges due to their size and difficulty training on limited FER datasets.

Purpose of the Study:

• To develop an effective and lightweight Transformer variant for FER.
• To address the computational and data limitations of existing Transformer-based FER models.

Main Methods:

• Proposed VaTFER (variant Transformer for FER) using action unit (AU) tokens with Histogram of Oriented Gradient (HOG) features.
• Introduced a spatial-channel feature relevance Transformer (SCFRT) module with multilayer channel reduction self-attention (MLCRSA) and dynamic learnable information extraction (DLIE).
• Incorporated an excitation module for prediction and a binary quantization mechanism for a lightweight variant (VaBTFER).

Main Results:

• The proposed VaTFER and VaBTFER models demonstrate effectiveness in facial expression recognition.
• Extensive experiments on multiple datasets validate the performance of the developed methods.
• The SCFRT module, MLCRSA, and DLIE contribute to improved learning and reduced parameter count.

Conclusions:

• The developed VaTFER and VaBTFER models offer a lightweight and effective solution for facial expression recognition.
• The novel architectural components address key limitations of pure Transformer models in FER.
• The findings suggest a promising direction for efficient and accurate FER systems.