EfficientNet B0 Feature Extraction with L2-SVM Classification for Robust Facial Expression Recognition

Facial expression recognition (FER) remains a challenging task due to the subtle visual variations between emotional categories and the constraints of small, controlled datasets. Traditional deep learning approaches often require extensive training, large-scale datasets, and data augmentation to achieve robust generalization. To overcome these limitations, this paper proposes a hybrid FER framework that combines EfficientNet B0 as a deep feature extractor with an L2-regularized Support Vector Machine (L2-SVM) classifier. The model is designed to operate effectively on limited data without the need for end-to-end fine-tuning or augmentation, offering a lightweight and efficient solution for resource-constrained environments. Experimental results on the JAFFE and CK+ benchmark datasets demonstrate the proposed method’s strong performance, achieving up to 100% accuracy across various hold-out splits (90:10, 80:20, 70:30) and 99.8% accuracy under 5-fold cross-validation. Evaluation metrics including precision, recall, and F1-score consistently exceeded 95% across all emotion classes. Confusion matrix analysis revealed perfect classification of high-intensity emotions such as Happiness and Surprise, while minor misclassifications occurred in more ambiguous expressions like Fear and Sadness. These results validate the model’s generalization ability, efficiency, and suitability for real-time FER tasks. Future work will extend the framework to in-the-wild datasets and incorporate model explainability techniques to improve interpretability in practical deployment Keywords: Facial Expression Recognition, EfficientNet, SVM, Deep Features, Emotion Classification