In modern manned-unmanned teaming systems, a fundamental challenge lies in drones’ target recognition accuracy - the ability to correctly identify operator-designated targets while discriminating among similar candidates in complex operational environments. While current object detection modules excel at locating objects via bounding boxes, they lack intra-class discrimination capabilities. Recent vision-language pre-training (VLP) models and multimodal large models (MMLMs) have improved the alignment between text and visual information, but they still have limitations in precise target recognition, often producing scattered attention maps rather than a single focused result. This paper achieves precise alignment between aerial images and operator instructions by enhancing and integrating the VLP model with the object detection module based on a large language model (DeepSeek). Key innovations include: (1) Designing a cross-modal attention mechanism tailored for the drone domain to address scale changes and perspective tilting in bird's-eye views; (2) Proposing a hierarchical feature fusion strategy that enhances the ability to distinguish color attributes, spatial relationships and dependency relationships through text-guided visual attention and visual-guided text attention modules; (3) Experiments on public datasets (Flickr30k, COCO, Visual Genome) and drone-specific datasets (AUG) demonstrate that this method significantly improves target discrimination capabilities in complex scenarios.