<![CDATA[Facial expression recognition remains a challenging task in computer vision, particularly due to class imbalance in datasets such as FER2013, where the Happy class accounts for 25.05% and Disgust only 1.52%, leading to biased predictions. This study proposes a Multi-Level Ensemble approach that integrates data diversity (bootstrap sampling in ELM), model diversity (CNN and CNN-ELM), and classifier diversity (ELM with different random seeds). The method is evaluated using Stratified 5-Fold Cross-Validation on 35,887 FER2013 images. Results show that pure ELM achieves 36.37% accuracy, CNN baseline 66.90%, CNN-ELM 67.28%, and ELM ensemble 67.34%. The proposed method achieves the best performance at 68.23%, improving the CNN baseline by +1.33%. Diversity analysis reports a disagreement rate of 21.2%, Q-statistic of 0.9313, and double-fault of 25.5%. These results indicate that the proposed framework effectively improves FER performance under class imbalance conditions.]]>
Copyrights © 2026
