LDR-Net: Landmark-Guided Diverse Regional Representation Learning for Facial Expression Recognition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Facial Expression Recognition (FER) is crucial for human-computer interaction.
• Current FER methods struggle with real-world variations due to limited spatial feature extraction.
• Overlooking spatial structure hinders representation capabilities in existing FER systems.

Purpose of the Study:

• To propose a novel Landmark-guided Diverse Regional Representation Network (LDR-Net) for robust FER.
• To enhance regional feature representation by preserving structural details and focusing on expression-critical areas.
• To improve FER performance under real-world variations like occlusions and pose changes.

Main Methods:

• Developed a Diverse Regional Feature Extraction (DRFE) module using landmark-guided cropping and cross-level feature integration.
• Introduced a Diverse Representation Learning (DRL) module with a dual-stream mechanism (Transformers and attention) for local and global feature learning.
• Proposed a Hybrid Feature Fusion (HFF) module for optimized joint decision-making through hierarchical fusion.

Main Results:

• LDR-Net demonstrated superior performance on multiple FER benchmarks (RAF-DB, AffectNet, SFEW).
• The network achieved state-of-the-art results on occlusion and pose test sets.
• Validated effectiveness in real-world driving scenarios using the KMU-FED dataset.

Conclusions:

• LDR-Net effectively addresses limitations in spatial structure representation for FER.
• The proposed network offers improved accuracy and robustness for facial expression recognition.
• LDR-Net shows significant potential for real-world applications in human-computer interaction.