Application of Multiple Deep Learning Architectures for Emotion Classification Based on Facial Expressions

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) is crucial for understanding human emotions.
• Applications span big data analytics, healthcare, security, and user experience.
• Deep learning models offer advanced capabilities for FER tasks.

Purpose of the Study:

• To comprehensively evaluate ten state-of-the-art deep learning models for FER.
• To analyze model performance based on accuracy, training time, and file size.
• To identify optimal architectures for specific FER application requirements.

Main Methods:

• Utilized the FER2013 dataset for facial expression recognition.
• Evaluated ten deep learning models: VGG16, VGG19, ResNet50, ResNet101, DenseNet, GoogLeNet V1, MobileNet V1, EfficientNet V2, ShuffleNet V2, and RepVGG.
• Assessed key performance metrics including test accuracy, training time, and weight file size.

Main Results:

• EfficientNet V2 and ResNet50 demonstrated superior performance with high accuracy and stable convergence.
• DenseNet, GoogLeNet V1, and RepVGG showed strong results but with slower initial convergence.
• Lightweight models (MobileNet V1, ShuffleNet V2) offered computational efficiency but lower accuracy on challenging emotions.

Conclusions:

• A critical trade-off exists between computational efficiency and predictive accuracy in FER models.
• Model selection for FER should align with specific application needs and constraints.
• This research advances deep learning for FER by detailing model performance and trade-offs.