<p>Traditional object detection methods have achieved a high level of performance. It is very dependent on a large amount of high-quality labeled datasets, which is often laborious to prepare. However, some scenarios are difficult to provide many images, which requires the few-shot object detection to accomplish the corresponding task. In this paper, we propose a new few-shot object detection method, Few-Shot object detection via Double Matching Branch (FSDMB), which can detect novel classes based on few instances. The core of our approach is to achieve efficient extraction of semantic information in positive support categories and to suppress the generation of similar categories by a double branching structure. Besides, we propose the Reweighting RPN module and Convolution and Fully connected Matching Head (CFMH) to explore the correlation between positive support instances and query images to detect novel categories. In order to verify the effectiveness of the method, we conduct sufficient experiments on PASCAL VOC and MS-COCO datasets. Experiments show that FSDMB can achieve competitive results at low shots compared to other state-of-the-art few-shot detectors. In PASCAL VOC, our proposed method gains +5.3AP over the current SOTA for shot=1, +2.5AP for shot=2, and +1.4AP for shot=3. For the MS-COCO dataset, with shot=10, our proposed methods gain +1.5AP and +1.2AP75 than the current SOTA. These improvements indicate that the model has a strong ability to identify potential new targets, reducing the model confusion of new classes. It is of great significance in the field of target detection with few shots.</p>