Published October 11, 2026 | Version v1

Asfalt Yuzeylerdeki Catlaklarin Tespitinde Veri Artirma Paradoksu: Kokeni ve Erken Durdurma Etkilesimi

Description

This study compares AlexNet, VGG16, and ResNet50 for automated asphalt surface crack classification on a publicly available dataset of 5565 images using three random seeds, and investigates the data augmentation paradox we reported earlier. Counter-intuitively, geometric augmentation degrades accuracy by up to 4.56 percentage points across all architectures examined. Here we trace the origin of this paradox through controlled experiments and show that the effect stems not from augmentation itself but from its interaction with a tight early stopping budget: augmentation delays convergence and destabilises validation loss, so the patience criterion expires before the model matures and under-trained weights are frozen. When the epoch budget is widened, the paradox disappears. The decisive role of transfer learning and the performance differences between architectures are preserved, although the ordering itself changes between in-domain and out-of-domain evaluation. We further show that, given the class imbalance, evaluation must rely on MCC and balanced accuracy rather than raw accuracy.

Files

124.pdf

Files (541.4 kB)

Name Size
md5:5610f879fc511b72d718015039171c0c
541.4 kB Preview Download

Additional details

Additional titles

Translated title (English)
The Data Augmentation Paradox in Asphalt Surface Crack Detection: Its Origin and Interaction With Early Stopping