확산 모델 기반 배경 재구성 합성 데이터를 활용한 군용 차량 객체 탐지 연구 = Research on Military Vehicle Object Detection using Diffusion Model-based Background Reconstruction Synthetic Data

다국어 초록 (Multilingual Abstract)

In military vehicle detection, security constraints severely restrict access to training imagery. We present a synthetic data generation pipeline that couples Stable Diffusion–based inpainting with ControlNet conditioning, following a principle of preserving object shape while reconstructing the background to enhance both realism and diversity. We further introduce a staged mixing strategy that progressively integrates real and synthetic data to strengthen domain generalization. On YOLOv8, our approach achieves an mAP (0.5:0.95) improvement from 58.4% to 65.6% (+7.2 percentage points), while the generated images achieve the lowest FID (66.263) among all compared methods. Consistent mAP gains are observed across multiple model architectures, and our method outperforms GPT-based generative approaches in zero-shot detection. These results indicate that the proposed technique can meaningfully mitigate data scarcity and domain shift, with applicability to a broad range of computer vision tasks