Training-Free Open-World Fine-Grained UAV Recognition via Structure-Semantic Prototype Augmentation

Abstract

Open-world fine-grained unmanned aerial vehicle (UAV) recognition is challenging due to the simultaneous presence of subtle inter-class variations, large intra-class diversity, and previously unseen categories encountered during deployment. Existing approaches typically rely on supervised fine-tuning with extensive annotated data, limiting their scalability and adaptability. To address these limitations, we propose a trainingfree UAV recognition approach based on Structure-Semantic Prototype Augmentation (SPAR). The proposed method leverages complementary structural and semantic knowledge from pretrained vision foundation models to construct discriminative class prototypes without parameter optimization. Specifically, structural representations extracted from self-supervised visual features are combined with semantic embeddings derived from vision-language models to form unified prototypes for known UAV categories. To improve generalization to unseen classes, the prototype space is further augmented through structure-semantic neighborhood expansion, which propagates discriminative priors from semantically related categories while preserving structural consistency. During inference, open-world recognition is achieved through similarity-based prototype matching together with unknown-category rejection. Extensive experiments on finegrained UAV benchmarks demonstrate that the proposed method achieves strong recognition performance without any additional training. In particular, the proposed framework improves closedset classification accuracy from $\mathbf{6 8. 8 \%}$ to $\mathbf{7 1. 2 \%}$ and increases open-world recognition F1-score from 55.2% to 58.4% compared with existing training-free baselines, while also reducing unknown-class confusion. Furthermore, the method consistently outperforms conventional CLIP-based and prototype-based approaches under varying openness settings, demonstrating superior robustness and generalization capability for real-world UAV recognition scenarios.