Abstract: To address the challenges of identifying and matching high-value military targets in remote sensing images under open-world conditions, this paper proposes a general recognition framework that integrates a dynamic attention mechanism with zero-shot and few shot learning strategies. First, an improved dynamic attention method is introduced, which adaptively adjusts the size of attention windows to more effectively capture multi-scale target features. Then, targeting the need to recognize high-value targets such as new military equipment in surveillance scenarios, a retrieval-based zero-shot learning framework and a prototype-based few-shot learning framework are respectively designed to significantly enhance matching performance in open domains. Experimental results demonstrate that the proposed method achieves a detection accuracy of 90.54% and an average F1 score of 47.97% in remote sensing target recognition. In open-domain target matching, the Rank-1 accuracy reaches 55.18% under zero-shot conditions and 66.22% accuracy under few-shot conditions, verifying the effectiveness and practicality of the approach.