Towards a Better Performance in Facial Expression Recognition: A Data-Centric Approach

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition is crucial for various applications but faces challenges in real-world scenarios.
• Current research predominantly focuses on model-centric improvements, with insufficient attention to dataset quality.
• Misclassification in facial image datasets hinders the performance of automatic facial expression recognition systems.

Purpose of the Study:

• To propose a novel data-centric method to address misclassification issues in facial image datasets.
• To enhance the accuracy and robustness of facial expression recognition models.
• To improve the quality of facial expression datasets without modifying or augmenting images.

Main Methods:

• A data-centric strategy involving progressive dataset refinement through successive training of a fixed Convolutional Neural Network (CNN) model.
• Utilizing correctly predicted facial images from previous training iterations to progressively refine the dataset.
• Implementing automatic reclassification of the entire dataset after the final training iteration.

Main Results:

• Significant improvements in validation accuracy on FER2013 (20.45%), NHFI (14.47%), and AffectNet (39.66%).
• Achieved state-of-the-art recognition rates on reclassified datasets: 86.71% (FER2013), 70.44% (NHFI), and 89.17% (AffectNet).
• Demonstrated the effectiveness of the data-centric approach without image modification, deletion, or augmentation.

Conclusions:

• The proposed data-centric method effectively tackles misclassification in facial expression datasets.
• This approach enhances facial expression recognition accuracy and achieves state-of-the-art performance.
• Focusing on dataset quality is a promising direction for advancing facial expression recognition technology.