Hybrid Domain Consistency Constraints-Based Deep Neural Network for Facial Expression Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) performance is hindered by factors like uneven illumination, occlusions, and annotation subjectivity.
• Traditional FER methods struggle with these real-world complexities, necessitating more robust approaches.

Purpose of the Study:

• To propose a novel Hybrid Domain Consistency Network (HDCNet) for improved facial expression recognition.
• To enhance FER accuracy by integrating spatial and channel domain consistency through feature constraints.
• To develop a method that does not require additional labeled data for its attention-consistency constraints.

Main Methods:

• The HDCNet utilizes a feature constraint method combining spatial and channel domain consistency.
• It mines attention consistency features by comparing original and augmented facial expression images.
• A mixed domain consistency loss function constrains feature expression and optimizes network weights.

Main Results:

• Experiments on RAF-DB and AffectNet datasets demonstrated improved classification accuracy.
• The proposed HDCNet achieved accuracy improvements ranging from 0.3% to 3.84% over existing methods.
• The attention-consistency constraints function effectively without requiring additional labels.

Conclusions:

• The Hybrid Domain Consistency Network (HDCNet) offers a significant advancement in facial expression recognition.
• The proposed method effectively addresses common challenges in FER datasets, leading to higher accuracy.
• HDCNet provides a robust and label-efficient approach for facial expression recognition tasks.