Systematic Evaluation of Data Augmentation Strategies for CNN-Based Image Classification
4th Cognitive Models and Artificial Intelligence Conference, AICCONF 2026, Prague, Çek Cumhuriyeti, 24 - 25 Nisan 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/aicconf69182.2026.11600638
- Basıldığı Şehir: Prague
- Basıldığı Ülke: Çek Cumhuriyeti
- Anahtar Kelimeler: CIFAR-10, CNN, data augmentation, deep learning, grid search
- Kayseri Üniversitesi Adresli: Evet
Özet
Convolutional Neural Networks (CNNs) have achieved remarkable results in image classification studies. However, their performance is significantly dependent on training strategies and architectural choices, in particular data augmentation and regularization choices. This paper presents a systematic grid search study of augmentation hyperparameters across three CNN architectures with increasing capacity (Compact V1, Standard V2, and Advanced V3) on CIFAR-10. We evaluate commonly used geometric and photometric transformations, including rotation, vertical/horizontal flipping, scaling, contrast adjustment, and translation. The findings reveal that architecture-specific data augmentation policies meaningfully and measurably improve the model's generalization ability. The most effective Standard V2 configuration achieved an accuracy rate of 90.41%, providing a 3.16% performance increase compared to the baseline model without augmentation. Furthermore, the effects of interactions between augmentation density and parameters on model performance were examined in detail, providing a practical guide for selecting effective data augmentation policies under different model capacities.