A facial expression recognition network based on attention double branch enhanced fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expressions are crucial indicators of human emotional, cognitive, and physiological states.
• Automatic facial expression recognition (FER) is vital for applications in healthcare, business, security, education, and human-computer interaction.
• Current FER methods face challenges with insufficient feature extraction, loss of local information, and suboptimal accuracy.

Purpose of the Study:

• To address limitations in existing facial expression recognition techniques.
• To propose an advanced FER network that enhances feature extraction and accuracy.
• To improve the comprehensive understanding of facial cues through fused global and local information.

Main Methods:

• Developed a novel facial expression recognition network utilizing an attention double branch enhanced fusion strategy.
• Employed two parallel branches to independently capture global enhancement features and local attention semantics.
• Implemented decision-level fusion to integrate and leverage the complementarity of global and local information.

Main Results:

• The proposed network effectively extracts more complete features by fusing and enhancing both global and local information.
• Achieved high expression recognition accuracy: 89.41% on the RAF-DB dataset and 88.84% on the FERPlus dataset.
• Demonstrated superior performance compared to numerous existing methods in facial expression recognition.

Conclusions:

• The attention double branch enhanced fusion network significantly improves feature representation for facial expression recognition.
• The method's effectiveness and superiority are validated by its excellent performance on benchmark datasets.
• This work contributes a robust model for advancing automatic facial expression recognition technology.