Evaluating the impact of AI-generated synthetic noisy data on human emotion recognition models

Human emotion recognition remains a challenging task due to the complexity and variability of human emotions in real-world scenarios. This study investigates the impact of AI-generated synthetic data on enhancing Facial Expression Recognition (FER) model performance. Using the Juggernaut XL model, we generated 300 synthetic images per emotion category from the FER2013 dataset and incorporated them into the training process of a VGG-19-based FER model. Experimental results revealed that the synthetic data did not improve key performance metrics, with the originally trained model achieving an accuracy of 65%, compared to 63% for the augmented dataset. Precision, recall, and F1-score also exhibited fluctuations across different emotion categories, as illustrated by confusion matrices. The findings suggest that the quality of synthetic images plays a crucial role in model effectiveness, as insufficient diversity may introduce noise rather than beneficial augmentation. Factors such as limited training epochs and potential dataset biases may have also influenced the outcomes. This study highlights the importance of optimizing synthetic image realism to improve FER models and offers practical insights for future AI-driven applications using data augmentation.