Evaluating the Impact of Data Augmentation on Image Classification Accuracy

Rahman, Muhammad Anis Ur (2025) Evaluating the Impact of Data Augmentation on Image Classification Accuracy. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (8MB) | Preview
Preview	PDF (Configuration manual) Download (765kB) | Preview

Abstract

Data augmentation is the key to the better deep image classification model’s stability and generalization. Although flipping, cropping, and rotating are famous methods, the latest researches in the fields of the most recent generative models and adaptive augmentation plans unveil the latest aspects of the synthesis of variant and task-adaptable training samples. In the present thesis, we investigate the comparative effect of three radically different strategies, namely, GAN-driven generation, Vector Quantized Variational Autoencoders (VQ-VAE), and mix-driven strategies like MiAMix.

The methods are compared on three representative test datasets, namely MNIST, CIFAR-10, and Tiny ImageNet, sharing a same convolutional neural network architecture and test pipeline. The methods are compared on classification accuracy and Frechet Inception Distance (FID), thereby double-click evaluation of predictive potential and sample faithfulness is supported. GAN methods are stable in class imbalance handling and image-richest generation of samples, and VQ-VAE shows stability and reconstruction performance on test sets. The combination methods are generally strong in low-data conditions with the benefit of computational speedup and simplicity of implementation.

The findings indicate that in terms of quantity, no one of the augmentation techniques is superior to others. Rather, the effects of the optimal solution would largely rely on the type of dataset, model sensitivity, and target of augmentation. The study suggests a reproducibility protocol for evaluating augmentation methods and provides constructive recommendations in choosing augmentation methods in real-world machine learning pipeline designs. The study introduces the efficiency of method combination of the generative and mix types for acquiring the scalable and robust classification ability.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Shahid, Abdul UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence > Computer vision Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence > Computer vision Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Tamara Malone
Date Deposited:	19 May 2026 14:36
Last Modified:	19 May 2026 14:36
URI:	https://norma.ncirl.ie/id/eprint/9304

Actions (login required)

View Item