Quantifying the Dynamics of Data Augmentation Hyperparameters for Facial Expression Recognition
Year: 2023
Type: article
Abstract: Automated recognition of facial expressions is a central component of systems used in an expanding array of domains.For a computer to automatically recognize affect, copious amounts of data are required to successfully train the model.It can often take a lot of work to collect and label data.In recent years, researchers have applied numerous data augmentation strategies to increase the diversity of the data within training datasets.Here, I examined the most common data augmentation strategies to determine which strategies result in higher performance for the facial expression recognition machine learning model.I first tested each data augmentation technique by itself and compared their performances.I next ran an ablation study with the augmentation strategies.I then analyzed the effect of dataset size on the marginal contribution of data augmentation.I find that augmentation does not always improve performance.When the dataset size is small, it results in a degradation of model performance.The accuracy of models with data augmentation starts to outperform the models with no data augmentation when the training dataset size is greater than a certain threshold.These results highlight the importance of considering dataset size when applying data augmentation to computer vision.
Source: Journal of Student Research
Authors: Ryan Lin, Peter Washington
Language: English
Topic: Emotion and Mood Recognition
Subfield: Experimental and Cognitive Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: diamond
OpenAlex ID: W4391761887