Domain-Adapted Diffusion for Synthetic Data Augmentation in UAV Small Object Detection

Recently, diffusion-based generative models offer new avenues for synthetic data augmentation, yet VA V imagery still suffers from significant domain shift compared to general image datasets used to pre-train these models: small objects, high altitudes, oblique views and cluttered backgrounds all diverge from web-scale image distributions. To address this, we introduce a novel augmentation pipeline combining three key innovations. 1) Fine-tuning a pretrained text-to-image diffusion model by weighting latent loss in small-object regions, thereby adapting the model toward the UAV domain. 2) Deriving diverse scene prompts automatically by applying UMAP dimensionality reduction and HDBSCAN clustering on the original UAV training set, then using a large language model to generate scene-level prompts that when combined with class labels guide synthesis. 3) Filtering generated image-label pairs by cropping detected boxes and computing Contrastive Language-Image Pre-training (CLIP, a model for measuring cross-modal similarity between images and texts) similarity between each crop and its class description, discarding mismatched pairs. Experiments on a UAV small-object detection dataset demonstrate that our augmented data significantly improves detection performance.