Fine-Grained Feature Refinement and Spatial Alignment for Few-Shot Object Detection With Minimal Data

Few-shot object detection (FSOD) aims to detect objects from novel classes with minimal annotated examples. While prevailing meta-learning approaches extract class prototypes from the support branch and integrate them into the query branch for prediction, they often suffer from insufficient interaction between the two branches before final fusion, resulting in coarse prototype representations and less discriminative query features. This limitation becomes particularly critical in extremely data-scarce scenarios. To address this, we propose a novel framework that promotes fine-grained collaborative interaction between the query and support branches. Our approach introduces two key modules: the Fine-Grained Feature Coupling (FC) module, which aligns query features with class prototypes in the spatial feature domain via a similarity matrix, and further employs a foreground attention mechanism to suppress background interference; and the Prototype Fusion (PF) module, which adaptively integrates multiple feature interactions through learnable weights, thereby simultaneously enhancing intra-class compactness and inter-class separability. Experiments on PASCAL VOC and MS COCO datasets demonstrate that our method achieves state-of-the-art performance under various few-shot settings, validating the effectiveness of the proposed modules in improving feature discriminability and model robustness.