Complex Emotion Recognition via Facial Expressions with Label Noises Self-Cure Relation Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning for facial expression recognition typically focuses on basic emotions, requiring extensive annotated data.
• Recognizing complex emotions is challenging due to the difficulty in obtaining large, high-quality annotated datasets.

Purpose of the Study:

• To address the limitations of data scarcity and annotation difficulty in complex facial expression recognition.
• To introduce a novel few-shot learning approach for robust emotion classification from facial expressions.

Main Methods:

• Proposed Self-Cure Relation Networks (SCRNet), a metric-based few-shot model for complex emotion recognition.
• SCRNet utilizes deep features from convolutional neural networks to learn a distance metric.
• A mechanism for correcting noisy labels using class prototypes in external memory was implemented during meta-training.

Main Results:

• SCRNet effectively classifies facial images of new emotion classes using only a few examples per class.
• The model demonstrated robustness to label noise, a common issue in real-world datasets.
• Experiments on public and synthetic datasets validated the method's effectiveness.

Conclusions:

• Few-shot learning offers a viable solution for complex facial expression recognition with limited data.
• SCRNet provides a robust and effective approach for classifying nuanced human emotions from facial cues.
• The developed method has implications for advancing affective computing and human-computer interaction.