Advancing Facial Emotion Intensity Classification Through Fuzzy Ensemble Learning with Variable Intensity Levels: A Deep Dive Into Model Dynamics

In this research, we explore the complexities of Facial Emotion Intensity Classification, leveraging the FER2013 dataset to enhance adaptability to real-world scenarios. Employing a cross-subject research approach, our study investigates the impact of variable emotion intensity levels on model performance, revealing challenges in capturing nuanced facial expressions. Our methodology integrates a sophisticated Fuzzy Ensemble Learning framework, combining EfficientNetV2B0, InceptionResNetV2, and MobileNetV2. Two comprehensive experiments scrutinize the model’s response to varying emotion intensities. The methodology includes using facial landmarks to generate masks, directing focus to the region of interest and mitigating redundant data. In the first experiment, classifying intensities into “High” and “Low”, the ensemble achieves commendable F1-score: 76% for “Low” intensity and 80% for “High”. The second experiment introduces a “Medium” intensity class, maintaining high accuracy for “High” and “Low” (78 and 76%, respectively) but faces a challenge with “Medium” intensity, resulting in 41% F1-score. These results highlight the intricacies of interpreting variable emotion intensities in facial expressions, revealing limitations in pretrained models for nuanced emotional dynamics. The findings emphasize the need for caution when applying such models to variable intensity facial emotion scenarios, urging a conscientious approach for robust and reliable real-world applications.