Robust small object detection via multimodal pre-trained representations and lightweight task-specific adaptation

Small object detection (SOD) in military ground scenes remains a critical challenge due to limited pixel resolution, environmental interference, and the scarcity of mission-specific labeled data. Existing methods often fail to generalize in adversarial conditions or require extensive computational resources. To address these issues, this paper proposes a novel Foundation-Model-Driven SOD framework that leverages multimodal pre-trained representations to enhance feature discriminability in complex battlefields. The framework introduces a domain-adaptive Transformer layer to align general semantic priors with specific military targets, coupled with a prototype-based few-shot learning module to achieve high precision under extreme data-scarce regimes. To ensure real-time feasibility on edge devices, a hierarchical knowledge distillation strategy is implemented to compress the large-scale representation into a lightweight, high-performance detector. Furthermore, an automated cross-modal annotation tool is developed to construct a specialized military SOD dataset for validation. Experimental results demonstrate that the proposed method achieves a state-of-the-art mAP of 65.4% while maintaining real-time performance.