Akademik Veri Yönetim Sistemi
23rd International Istanbul Scientific Research Congress on Life, Engineering, Architecture, and Mathematical Sciences, İstanbul, Türkiye, 20 - 22 Kasım 2025, cilt.1, sa.7, ss.1049-1060, (Tam Metin Bildiri)
This study provides a structured evaluation of data augmentation strategies for object detection by using
a custom dataset and two widely used models, Faster R-CNN and YOLOv7. A baseline model without
augmentation is first established, where Faster R-CNN reaches a mean Average Precision value of 40.42
and YOLOv7 reaches 41.33. Building on this reference point, the study examines two augmentation
groups. The first group includes color-based transformations such as hue, saturation, grayscale,
brightness and contrast. The second group includes distortion-based transformations such as noise, blur
and cutout. The results indicate that color-based augmentations consistently improve detection accuracy,
while distortion-based methods yield smaller gains. One of the clearest improvements emerges from the
combined brightness and contrast strategy. Under this setting, Faster R-CNN improves from 40.42 to
44.89 and YOLOv7 improves from 41.33 to 45.93, showing that moderate color variation strengthens
feature learning and model generalization. Distortion oriented techniques produce limited benefits
because they reduce visual clarity or remove informative regions. Larger augmentation combinations
lead to over regularization and reduce convergence stability. The study addresses a gap in the literature
by comparing augmentation strategies under equal conditions and by demonstrating that balanced color-
oriented transformations provide the most reliable performance gains. These findings offer practical
guidance for researchers who aim to design effective augmentation pipelines for object detection
tasks.B