A systematic image augmentation framework based on dataset limitations for small object detection

Small object detection in aerial images remains a significant challenge due to low-resolution instances, sparse object distribution, and the lack of large, diverse datasets. These limitations reduce the accuracy and generalization ability of modern object detection algorithms. To address this, we first propose a three-tier systematic framework for optimizing data augmentation strategies, which includes (1) dataset analysis, (2) definition of augmentation objectives, and (3) systematic selection of techniques based on identified challenges. Building upon this framework, we propose a data augmentation pipeline designed to enhance small object visibility, improve dataset balance, and increase model robustness. The proposed pipeline consists of two primary stages. The first stage performs image segmentation and cropping, dividing images into smaller segments and applying localized zoom-in and center-cropping techniques. This stage selectively retains only segments with high object density, discarding empty or irrelevant regions, thereby increasing the object-to-image ratio. The second stage involves object-centric image resizing and augmentation, which standardizes input dimensions while maintaining enhanced object clarity and scale, ensuring compatibility with convolutional neural networks (CNNs). The effectiveness of the pipeline is validated using YOLOv8m on two datasets AI-TODv2 and SODA-A. The results demonstrate substantial improvements in performance metrics, including a $$+$$ + 15.9% AP@0.5 on AI-TODv2 and a 23.6% increase in SODA-A using YOLOv8m. But only the class ‘ship’ in AI-TODv2 showed a decline. Notably, in both datasets, the average precision (AP) of the classes has improved, highlighting the effectiveness of the proposed pipeline in enhancing the detection of hard-to-detect classes. These results confirm that our augmentation strategy significantly improves detection accuracy and recall for small objects in aerial scenes, overcoming the limitations of sparse, dense, and imbalanced datasets.