The Effect of Image Augmentation Techniques on the Accuracy of Deep Learning Models: A Systematic Review

Abstract

Image augmentation is an important technique for increasing the quantity and diversity of training data, thereby improving accuracy and reducing overfitting in deep learning models, especially when data is limited such as in the medical imaging domain. However, the variety of available augmentation techniques poses a challenge in selecting the most effective approach. This study aims to conduct a systematic review of the influence of various image augmentation techniques on the accuracy of deep learning models. The method follows the PRISMA protocol, searching the Scopus, IEEE Xplore, ScienceDirect, and arXiv databases from 2022 to 2026; of 264 initial records, 26 studies met the inclusion criteria. The synthesis shows that basic geometric techniques (flip, rotation, translation) provide consistent accuracy improvements at low computational cost; mixing-based techniques such as Mixup and CutMix yield greater gains in generalization and robustness, with CutMix reported to improve clean accuracy by an average of 1.54% and robust accuracy by 3.06%; while generative techniques based on GANs and diffusion models effectively address data scarcity in medical imaging but demand greater resources. This review concludes that the combination of techniques and their suitability to domain characteristics matter more than the selection of a single technique.