Convolutional Neural Networks are widely used for image recognition tasks. Data augmentation is a technique used to improve spatial invariance and reduce overfitting in Convolutional Networks. This article provides an empirical comparison of two common data augmentation techniques: RandAugment (a stochastic technique that applies geometric transforms to images) and Conditional Generative Adversarial Networks (models that generate synthetic samples from the same distribution as the training set). Combinations of these data augmentation techniques are also investigated. Three models – a base model developed for this study, and finetuned pre-trained versions of ResNet50 and InceptionV3 – are evaluated on benchmark datasets. Results indicate that RandAugment is more effective.