This study introduces a synthetic data generation approach utilizing generative artificial intelligence to tackle data scarcity and class imbalance in satellite imagery for military object detection. Traditional data augmentation techniques and GAN/CycleGAN-based methods often fall short in performance due to shape distortion of small objects and inconsistencies in complex backgrounds. To address these challenges, we combine a diffusion model (ControlNet) with a large language model (LLM) to produce synthetic satellite images that maintain the geometric characteristics of MiG-series fighter aircraft while featuring varied backgrounds and situational contexts. The synthetic data generated was integrated with real-world datasets and tested on YOLOv10 and RT-DETR models. Evaluation on an independent test set revealed that YOLOv10 improved its mean Average Precision (mAP@0.5) and Recall when utilizing the combination of LLM and ControlNet. Similarly, RT-DETR showed enhancements in mAP@0.5 and Recall when trained with LLM-based synthetic data. In contrast, traditional augmentation methods like Dust/Fog and Noise/Blur had limited or detrimental effects on performance for both models. These findings demonstrate the effectiveness of generative synthetic data in improving model generalization for satellite image object detection tasks. Furthermore, the study emphasizes that the contributions of ControlNet and LLMs may differ based on the underlying model architecture.



Keywords : 위성영상, 객체 탐지, 합성데이터, 확산모델, 대형언어모델, satellite imagery, object detection, synthetic data, diffusion model, large language model, YOLOv10