A Robust Facial Expression Recognition Algorithm Based on Multi-Rate Feature Fusion Scheme

Area of Science:

• Artificial Intelligence
• Computer Vision
• Affective Computing

Background:

• The growing importance of artificial intelligence (AI) necessitates advanced methods for understanding human emotions.
• Facial expression recognition (FER) is a key component in AI for interpreting human affective states.

Purpose of the Study:

• To propose a robust multi-depth network for efficient and accurate facial expression classification.
• To enhance the spatio-temporal information processing for improved emotion recognition.

Main Methods:

• A multi-depth network utilizing minimum overlapped frames for increased spatio-temporal data.
• Implementation of a multirate-based 3D convolutional neural network (CNN) and adaptive image normalization.
• Reinforcement of features using a self-attention module and classification via a joint fusion classifier.

Main Results:

• Achieved 96.23% accuracy on the CK+ database.
• Outperformed state-of-the-art models on MMI (96.69%) and GEMEP-FERA (99.79%) databases.
• Demonstrated performance on the challenging AFEW database with 31.02% accuracy.

Conclusions:

• The proposed multi-depth network offers a robust approach to facial expression recognition.
• The method shows competitive and superior performance across various standard and challenging datasets.
• Further research may focus on improving performance in highly variable real-world environments.