Effect of Data Augmentation on the Accuracy of Convolutional Neural Networks

Abstract

Data augmentation is one of the most famous techniques to increase the efficiency of a deep learning model. It involves using conceptual techniques to generate new data samples from existing ones. In the context of image processing, there are a plethora of options for performing data augmentation, for example, rotating images, changing hue, altering greyscale, etc. The main reason why data augmentation is used to reduce overfitting and increase the accuracy of a model. Convolutional neural networks (CNNs), conversely, are the most famous deep learning models used in the domain of image processing, image classification in particular. However, not much is known about the impact of various data augmentation techniques on the performance on various CNNs. In this paper, we analyse the effect of data augmentation on two different types of CNNs of varying dimensions. The aim of this study is to observe the impact of data augmentation techniques on the performance of CNNs of varying complexities. In order to bolster our findings, we used three different datasets for the use case of image classification and tried to compare the results for both CNNs. We analysed the performance of both types of CNNs over varying learning rates and epochs in order to compare our findings across varying hyperparameters.