Optimization of transfer learning for facial emotion classification on the FER-2013 dataset

Abstract

Facial expressions play a key role in non-verbal communication by naturally reflecting human emotions. Facial emotion recognition (FER) using computer vision has gained attention with advances in deep learning. However, deep learning models require large datasets to perform well, posing a challenge for FER tasks with limited data. Transfer learning is a promising approach to address this issue, but a standardized method for FER is yet to be established. This study optimizes three transfer learning models ResNet-50, Inception V3, and Xception on the FER-2013 dataset. Experiments include testing input image sizes, hyperparameter tuning, data augmentation, layer addition, and training methods. Results show each model requires different input sizes for best accuracy. Hyperparameter tuning improves accuracy by 6.35%, 4.69%, and 1.04% for ResNet-50, Inception V3, and Xception, respectively. Augmenting only the disgust class yields better accuracy than augmenting all classes. The freeze fine-tuning method is less effective than fine-tuning alone on datasets with thousands of samples but outperforms the freeze layer method. The best accuracies achieved are 64.89% (ResNet-50), 65.83% (Xception), and 66.40% (Inception V3). These findings provide insights into freeze fine-tuning limitations and guidance for optimizing transfer learning in FER with limited data.