TriCAFFNet: A Tri-Cross-Attention Transformer with a Multi-Feature Fusion Network for Facial Expression Recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) has advanced, but real-world applications remain challenging.
• Existing methods struggle with subtle expressions and varying environmental conditions.

Purpose of the Study:

• To propose an advanced model, TriCAFFNet, for robust facial expression recognition in challenging environments.
• To improve the accuracy and reliability of FER systems.

Main Methods:

• Developed TriCAFFNet, a multi-feature fusion network incorporating Local Binary Pattern (LBP), Histogram of Oriented Gradients (HOG), landmark, and convolutional neural network (CNN) features.
• Implemented tri-cross-attention blocks to enable feature interaction and mutual guidance for capturing salient attention.

Main Results:

• Achieved state-of-the-art (SOTA) performance on benchmark datasets.
• Obtained 92.17% accuracy on RAF-DB, 67.40% on AffectNet (7 classes), and 63.49% on AffectNet (8 classes).

Conclusions:

• TriCAFFNet demonstrates superior performance in facial expression recognition, particularly under challenging real-world conditions.
• The proposed multi-feature fusion and attention mechanisms are effective for enhancing FER accuracy.