Comparative Analysis of Data Augmentation Techniques for ResNet18-based Image Classification on CIFAR-100
International Journal on Advanced Science, Engineering and Information Technology · 25.10.2026 · 16 (5) · 1920-1928
DOI: 10.18517/ijaseit.16.5.12458 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal on Advanced Science, Engineering and Information Technology, online erschienen am 25.10.2026. Verfasst von Xie Shaolin und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
On the CIFAR-100 dataset, the study conducted a series of experiments using an untrained ResNet-18 model to investigate the impact of augmentation type, combination, and intensity on image classification accuracy. We examined eight augmentation techniques, including geometric, photometric, hybrid, and automated augmentation methods. The experiments aim to explore three topics: which single technique is better? which is better, manual combination or automated methods?, and is there a relationship between intensity and accuracy? Our evaluation metrics included Top-1 accuracy, Top-5 accuracy, macro F1 score, test loss, overfitting gap, convergence period, and paired t-tests (including effect size and confidence intervals), among others. CutMix achieved the highest Top-1 accuracy (57.52%) and the smallest overfitting gap, while CutMix+Flip achieved the best overall performance (62.88%), outperforming larger augmentation pipelines and automated methods. The impact of augmentation intensity varied. When enhancement intensity exceeds a reasonable range, category-related image features may become excessively distorted or occluded, increasing learning difficulty and reducing classification accuracy. These findings provide practical guidance for selecting, combining, and calibrating data augmentation strength, but further validation on larger datasets and other convolutional and Transformer-based architectures is still needed.
Zitieren
APA
Shaolin, X., Aris, T. N. M., & Kasmiran, K. A. (2026). Comparative Analysis of Data Augmentation Techniques for ResNet18-based Image Classification on CIFAR-100. International Journal on Advanced Science, Engineering and Information Technology, 16(5), 1920-1928. https://doi.org/10.18517/ijaseit.16.5.12458
Vancouver
Shaolin X, Aris TNM, Kasmiran KA. Comparative Analysis of Data Augmentation Techniques for ResNet18-based Image Classification on CIFAR-100. International Journal on Advanced Science, Engineering and Information Technology. 2026;16(5):1920-1928. doi:10.18517/ijaseit.16.5.12458
BibTeX
@article{Shaolin2026,
author = {Shaolin, Xie and Aris, Teh Noranis Mohd and Kasmiran, Khairul Azhar},
title = {Comparative Analysis of Data Augmentation Techniques for ResNet18-based Image Classification on CIFAR-100},
journal = {International Journal on Advanced Science, Engineering and Information Technology},
year = {2026},
volume = {16},
number = {5},
pages = {1920--1928},
doi = {10.18517/ijaseit.16.5.12458}
}
EUVIMED