Optimizing Facial Expression Recognition with Lightweight Hybrid CNN via Feature Fusion
Abstrak
FER (Facial Emotion Recognition) is crucial to safety, human-computer interaction, healthcare, and marketing applications. However, challenges like occlusions, subtle variations in expressions, and class imbalance hinder its applicability in the real world. This new FER model will combine the complementary strengths of CNNs (Convolutional Neural Network) of YOLOv11 (You only Look Once) and EfficientNet-B0. By utilizing the spatial feature extraction capability of YOLO with the global semantic feature representations of EfficientNet, the model achieves enhanced classification performance, particularly for under-represented emotions. To counter the class imbalance in the dataset SMOTE (Synthetic Minority Oversampling Technique) and a class-weighted loss is used. Evaluated on the RAF-DB (The Real-world Affective Faces Database) dataset, this approach outperforms single-model architecture consistently with significant gains in classification accuracy for a large spectrum of expressions.
Identifier
- OpenAlex: https://openalex.org/W4414271033
- DOI: https://doi.org/10.1109/prml66062.2025.11160313

