Learning Deep Global Multi-Scale and Local Attention Features for Facial Expression Recognition in the Wild

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial Expression Recognition (FER) in the wild faces significant challenges due to occlusion and pose variations.
• Existing methods often struggle with these real-world complexities, impacting recognition accuracy.

Purpose of the Study:

• To propose a novel network, MA-Net (Multi-scale and Local Attention Network), for robust FER in the wild.
• To enhance FER performance by effectively handling occlusion and pose variations.

Main Methods:

• Developed MA-Net with three core components: a feature pre-extractor, a multi-scale module, and a local attention module.
• The multi-scale module fuses features with diverse receptive fields to mitigate occlusion and pose issues.
• The local attention module focuses on salient local features, reducing interference from occlusions and non-frontal poses.

Main Results:

• MA-Net achieved state-of-the-art performance on several in-the-wild FER benchmarks.
• Specific accuracies include: CAER-S (88.42%), AffectNet-7 (64.53%), AffectNet-8 (60.29%), RAFDB (88.40%), and SFEW (59.40%).

Conclusions:

• The proposed MA-Net demonstrates superior performance for FER in the wild.
• The network's architecture effectively addresses key challenges like occlusion and pose variation, paving the way for more reliable facial expression analysis.