CNN-Based Facial Expression Recognition with Simultaneous Consideration of Inter-Class and Intra-Class Variations

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition is vital for understanding human emotions and nonverbal cues.
• Current facial recognition technology often overlooks the significance of the loss function in deep learning models.
• Existing methods primarily focus on novel network architectures, neglecting loss function optimization.

Purpose of the Study:

• To introduce a new loss function for Convolutional Neural Network (CNN) based facial expression recognition.
• To simultaneously address inter-class and intra-class variations for improved recognition accuracy.
• To enhance the performance of facial expression recognition systems by optimizing the loss function.

Main Methods:

• Developed a novel loss function designed to minimize intra-class variations by pulling deep features towards their class centers.
• Increased inter-class variations by pushing deep features away from non-corresponding class centers and maximizing distances between different class centers.
• Integrated the proposed loss function into a CNN architecture for facial expression recognition tasks.

Main Results:

• The proposed loss function demonstrated superior performance compared to existing methods on benchmark datasets.
• Evaluated on Cohn-Kanade Plus, Oulu-Casia, MMI, and FER2013 datasets, showing significant improvements.
• Effectively reduced intra-class variations and increased inter-class variations, leading to more robust feature representations.

Conclusions:

• The novel loss function offers a significant advancement in facial expression recognition accuracy and efficiency.
• This approach provides a more effective way to train deep learning models for emotion recognition.
• The method shows strong potential for real-world applications requiring precise facial expression analysis.