Few-shot object detection in remote sensing imagery faces significant challenges, including limited labeled samples, complex scene backgrounds, and subtle inter-class differences. To tackle these issues, we design a novel detection framework that effectively transfers supervision from a few annotated support examples to the query domain. We introduce a feature enhancement mechanism that injects fine-grained support cues into the query representation, helping the model focus on relevant regions and suppress background noise. This allows the model to generate more accurate proposals and perform robust classification, especially for visually confusing or small objects. Additionally, our method enhances feature interaction between support and query images through a nonlinear combination strategy, which captures both semantic similarity and discriminative differences. The proposed framework is fully end-to-end and jointly optimizes the feature fusion and detection processes. Experiments on three challenging benchmarks, NWPU VHR-10, iSAID and DIOR, demonstrate that our method consistently achieves state-of-the-art results under different few-shot settings and category splits. Compared with other advanced methods, it yields superior performance, highlighting its strong generalization ability in low-data remote sensing scenarios.

In the realm of remote sensing image object detection compression distillation, establishing an efficient method for instance feature knowledge transfer between teacher and student models holds paramount importance. To this end, this paper introduces an innovative deep structured instance graph distillation method that endeavors to delve into the underlying information between instance features, thereby optimizing detection performance. Specifically, our proposed method incorporates feature instances and their relations into a graph-based structure (SIG). In this graph, feature instances serve as the nodes, while the relations between them serve as the edges. This structure enables us to capture both the individual significance of each feature instance and their collective influence within the context. Furthermore, in the experiment, we found that the index of some dense and small-target objects did not improve much because the edge assembly generated by a large number of background feature nodes in the SIG module inhibited the loss. To address the perennial imbalance between foreground and background features, we introduce an adaptive background feature mining strategy. Through carefully calibrated weights, this strategy effectively extracts and integrates background information, thereby minimizing noise interference in detection results and augmenting the expressive capacity of foreground features. We achieved state-of-the-art results on both the challenging DIOR and DOTA datasets, with the two-stage Oriented RCNN-based student Resnet18 model achieving a 73.23 mAP on the DOTA benchmark, close to the teacher Resnet101’s 76.16. In addition, on the DIOR dataset, the student Resnet18 based on the two-stage Faster RCNN achieved 70.13 mAP, higher than the baseline 66.31, and the student Resnet50 achieved 72.28, higher than the teacher’s 72.25.