Convolutional neural network for data augmentation

Rachel Kim * and Emily White

Department of Computer Science, Stanford University, California, USA.
 
Research Article
World Journal of Advanced Engineering Technology and Sciences, 2024, 13(02), 870-886.
Article DOI: 10.30574/wjaets.2024.13.2.0528
Publication history: 
Received on 04 October 2024; revised on 20 November 2024; accepted on 24 November 2024
 
Abstract: 
In deep learning, the triumph of Convolutional Neural Networks (CNNs) depends significantly on whether large and varied datasets are available. In most real-world applications, obtaining enormous amounts of labeled data is either time-consuming, costly, or unfeasible. Data augmentation has become a principal technique to artificially enlarge training sets by generating novel data instances by applying various transformations. Traditional augmentation methods, such as rotation, flipping, scaling, and cropping, cannot produce adequately diverse and semantically rich data. To fight this limitation, this study delves into using CNNs as a tool for sophisticated data augmentation.
The primary objective of this research is to explore and evaluate the prospects of CNN-based data augmentation techniques for improving the generalization performance of deep learning models. We propose a CNN-based data augmentation framework using learned features to create synthetic but realistic image data. This involves using deep generative models, transfer learning, and feature-space transformations instead of conventional augmentation techniques for data augmentation.
Experiments were conducted on benchmark image datasets MNIST and CIFAR-10 for comparing models learned from traditional and CNN-augmented data. The result is a remarkable classification accuracy and robustness boost when CNN-based augmentation is applied. Particularly noteworthy in the current context is that the augmented datasets produced more informative and diverse samples, their overfitting suppression was reinforced, and model generalization improved.
Our findings illustrate the potential of CNNs to transform data augmentation and optimization in automating. Not only can this approach improve model performance, but it also reduces the need for human data annotation. The implications are particularly valuable in sparse-annotated data domains, such as medical imaging and autonomous driving systems. Future research will integrate CNN augmentation with adversarial training and semi-supervised learning to improve learning efficiency and robustness in low-data regimes.
 
 
Keywords: 
Convolutional Neural Networks; Data Augmentation; Deep Learning; Image Generation; Synthetic Data; Generalization
 
Full text article in PDF: 

Click here

 
Copyright information: 
Copyright © 2024 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0