CRS-CONT: A Well-Trained General Encoder for Facial Expression Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Current facial expression recognition (FER) methods require task-specific training data and encoders.
• This leads to significant training burdens and limitations in generalizability.

Purpose of the Study:

• To develop a general facial expression encoder pre-trained for versatile feature extraction.
• To eliminate the need for fine-tuning on specific FER datasets.

Main Methods:

• Extended self-supervised contrastive learning for pre-training a general FER encoder.
• Introduced coarse-contrastive (CRS-CONT) learning using coarse-grained labels and data augmentation.
• Developed a weight vector to manage feature distribution constraints for fine-grained accuracy.

Main Results:

• The pre-trained general encoder demonstrated superior or comparable performance on various FER datasets.
• Achieved strong results in cross-dataset evaluations and on unseen facial expressions.
• The encoder, with frozen weights, adapted effectively to different facial expression tasks.

Conclusions:

• The proposed CRS-CONT method offers a robust solution for general facial expression analysis.
• Reduces training burden and overcomes limitations of fully-supervised learning.
• Provides a foundation for more adaptable and efficient FER systems.