Effect of Data Augmentation Levels on MobileNetV2 Performance for Facial Expression Recognition Using FER-2013 Dataset

Abstract

This research is conducted to assess how data augmentation techniques influence the performance of deep learning models in the task of facial expression classification. A critical issue commonly encountered in facial image analysis is the scarcity of available datasets, which frequently results in overfitting and limits the model’s ability to generalize effectively to unseen data. In response to this limitation, the study implements two distinct levels of augmentation, referred to as light augmentation and complex augmentation, and evaluates their performance against a baseline condition where no augmentation is applied. The model employed in this investigation is MobileNetV2, trained using the FER-2013 dataset that contains seven distinct emotion categories. The findings from the experimental evaluation indicate that the application of light augmentation yields the highest validation accuracy at 22.7%, outperforming both the no-augmentation scenario (10.8%) and the complex augmentation approach (8.3%). Although the use of complex augmentation results in lower loss values, it does not translate into improved classification accuracy. This outcome suggests that overly intensive augmentation strategies may hinder the model’s capacity to effectively learn and extract meaningful features. Overall, these results highlight the necessity of carefully determining appropriate augmentation methods to enhance the performance of deep learning models.