Impact of data augmentation strategies on deep learning-based image classification

Abstract

This paper explores how data augmentation techniques influence the performance and generalization of deep learning-based image classifiers. Augmentation is essential for enhancing robustness and accuracy, especially when training data are limited or imbalanced. The study discusses various augmentation methods, from basic geometric transformations to advanced approaches such as Mixup, CutMix, and GAN-generated samples. The analysis highlights the improvement these methods bring to image classification tasks and their role in preventing overfitting.