<p>Few-shot object detection aims to achieve accurate detection of novel objects with only a limited number of training samples, which poses a significant challenge. Although many existing methods commonly employ Faster R-CNN for transfer learning, two critical issues are often overlooked. First, the scarcity of foreground object labels in few-shot scenarios causes the standard classifier to overemphasize interfering objects, leading to classification bias. Second, existing detectors fail to adequately capture the feature distributions of novel classes and neglect the spatial misalignment of semantically similar pixels, resulting in numerous low-quality confidence scores from the classification branch. To address these limitations, we propose a decoupled classifier and a confidence score refinement network for few-shot object detection. The former replaces the conventional classification head with two independent components: one designed to process clean positive samples and the other to handle noisy negative samples, mitigating classification bias. The latter is based on metric learning for few-shot classification, incorporating an attention-based feature enhancement block and a feature matching module to compute more accurate similarity scores between image pairs, further improving classification accuracy. Extensive experiments conducted on the PASCAL VOC and MS COCO benchmarks demonstrate that the proposed method significantly outperforms most state-of-the-art approaches under various few-shot settings, confirming its effectiveness and robustness.</p>