Facial Expression Recognition Based on Fine-Tuned Channel-Spatial Attention Transformer

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) is crucial for human-computer interaction.
• Real-world FER is challenged by factors like occlusion and head pose variations.
• Existing methods struggle with these real-world complexities.

Purpose of the Study:

• To develop an advanced model for accurate facial expression recognition in the wild.
• To address the limitations of current FER systems in uncontrolled environments.
• To improve the robustness of FER against occlusions and head pose changes.

Main Methods:

• Proposed a novel model: fine-tuned channel-spatial attention transformer (FT-CSAT).
• Incorporated a channel-spatial attention module to focus on relevant facial features.
• Utilized a fine-tuning strategy to optimize performance and manage parameters.

Main Results:

• Achieved high accuracy on benchmark datasets: 88.61% on RAF-DB and 89.26% on FERPlus.
• Demonstrated superior performance on Occlusion-RAF-DB and Pose-RAF-DB datasets.
• FT-CSAT effectively handles facial occlusion and head pose variations.

Conclusions:

• The FT-CSAT model significantly advances the state-of-the-art in facial expression recognition.
• The proposed attention mechanism and fine-tuning approach enhance robustness in real-world scenarios.
• FT-CSAT offers a promising solution for reliable FER in challenging conditions.