CPRNet: Class Prototype Reweight Neural Network for Few-Shot Object Detection

Few-Shot Object Detection (FSOD) addresses the challenge of detecting and localizing target objects in a test set when only a limited number of training samples are available. In few-shot learning, the support set provides reference samples to guide the model, while the query set evaluates its performance on unseen data. While the many FSOD models struggle to fully capture contextual information and fine-grained details during feature extraction, which can lead to overfitting due to limited training diversity. To address these limitations, we propose a framework, named Class Prototype Reweight Network (CPRNet), that introduces a dual-strategy improvement based on DEVIT module. First, the FE-FSOD module enhances feature extraction through a spatial attention mechanism and an adaptive feature enhancement algorithm. Second, the PR-FSOD module refines class prototype construction using a dual-feature aggregation layer and a class prototype optimization algorithm. Experimental results on COCO datasets show that our proposed CPRNet achieves significant performance gains, demonstrating its effectiveness in few-shot object detection tasks.