Facial Expression Recognition with Contrastive Learning and Uncertainty-Guided Relabeling.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial Expression Recognition (FER) is crucial for human-computer interaction.
• Existing deep learning (DL) methods for FER struggle with extracting discriminative semantic information and handling annotation ambiguity.
• The padding erosion problem also affects FER performance.

Purpose of the Study:

• To develop an end-to-end FER network that efficiently and accurately recognizes facial expressions.
• To mitigate the impact of annotation ambiguity in FER datasets.
• To improve the extraction of discriminative expression features.

Main Methods:

• Proposed an end-to-end network incorporating supervised contrastive loss (SCL) for enhanced feature learning.
• Introduced an uncertainty estimation-based relabeling module (UERM) to address annotation ambiguity.
• Integrated an amending representation module (ARM) to tackle padding erosion.

Main Results:

• Achieved state-of-the-art (SOTA) performance on public benchmarks.
• Reached 90.91% accuracy on RAF-DB, 88.59% on FERPlus, and 61.00% on AffectNet.
• Demonstrated significant improvements in FER performance due to proposed modules.

Conclusions:

• The proposed method effectively extracts discriminative expression features and handles annotation ambiguity.
• The novel network architecture significantly advances the state-of-the-art in facial expression recognition.
• This approach offers a robust solution for accurate and efficient FER systems.