Dual-Prompting and Class-Aware Token Selection for UAV-Based Few-Shot Ship Classification

The rapid advancement of consumer-grade unmanned aerial vehicles (UAVs) has enabled low-cost and flexible maritime surveillance, yet the high cost of large-scale annotation and complex maritime backgrounds hinders the effectiveness of conventional deep learning. Few-shot learning offers a promising solution, but existing approaches often suffer from overfitting and feature redundancy in complex scenes. We propose a Dual-Prompting and Class-Aware Token Selection (DCTS) framework for few-shot ship classification in UAV-based maritime monitoring. Inspired by human visual memory, DCTS integrates Long-term Prompts to retain stable knowledge and Working Prompts to adapt to episode-specific features. A Quaternion Neural Network-based Prompt Generator (Q-Prompt Generator), trained with weakly supervised contrastive learning and a Patch Similarity Loss, enhances the Vision Transformer’s discrimination capability. The Class-Aware Token Selector (CATS) further suppresses redundant features and strengthens semantic relevance at the patch level. We also introduce the Vessel Classification Dataset (VCD), a large-scale UAV-perspective vessel dataset with 18,000 images in 30 categories. Experiments on VCD, FGSCR, and MiniImageNet demonstrate that DCTS achieves superior performance compared with existing methods, showing improved generalization and robustness in real-world UAV maritime applications. Our dataset is publicly available at https://github.com/GG-Lizen/VCD.git.