对遥感图像进行目标检测，具有广阔的应用前景，在许多领域有着十分关键的作用。针对小样本背景下遥感图像目标检测任务存在特征提取不足、定位困难、分类易混淆的问题，提出了一种基于协同注意力模块和对比学习分支的小样本目标检测算法。首先，对训练样本进行数据增强操作，以扩充数据集规模。其次，提出了一种协同注意力模块，包括设计的背景衰减注意力和空间感知注意力，利用遥感图像丰富的背景与目标特征信息，指导网络关注与目标定位相关的重点信息，从而便于 RPN 网络生成更好的区域建议框，减少遗漏目标的概率，提升模型对小样本类别的定位性能。然后，设计了一种对比学习分支。基于设计的对比损失函数，通过联合训练策略，在训练时从特征学习逐步过渡到分类器学习，提高了分类的准确率。最后，设计出一种基于微调的迁移学习范式的小样本目标检测模型，分为基础训练阶段和微调阶段，在基础训练阶段借助充足的基类样本训练模型学习类无关的参数，在微调阶段使用制作的小样本数据集帮助目标检测模型适应特定类别目标，提升其检测性能。本文以TFA为基准，通过在遥感数据集NWPU VHR-10和DIOR上验证了算法的有效性，结果显示本文算法在NWPU VHR-10和DIOR数据集上与基准算法比较，平均精度(mAP)均有大幅提升。此外，本文算法均优于现有的多种先进算法。

Object detection in remote sensing images has broad application prospects and plays a crucial role in many fields. A small sample object detection algorithm based on collaborative attention module and contrastive learning branch is proposed to address the problems of insufficient feature extraction, difficult localization, and easy classification confusion in remote sensing image object detection tasks under small sample backgrounds. Firstly, perform data augmentation on the training samples to expand the number of datasets. Secondly, a collaborative attention module is proposed, which includes designed background attenuation attention and spatial perception attention. By utilizing the rich background and target feature information of remote sensing images, the network is guided to focus on key information related to target localization, which facilitates the RPN network to generate better region suggestion boxes, reduce the probability of missing targets, and improve the model""s positioning performance for small sample categories. Then, a contrastive learning branch was designed. Based on the designed contrast loss function, through the joint training strategy, the training gradually transits from feature learning to classifier learning, which improves the accuracy of classification. Finally, a small sample target detection model based on the fine-tuning transfer learning paradigm is designed, which is divided into the basic training stage and the fine-tuning stage. In the basic training stage, sufficient base class samples are used to train the model to learn class independent parameters, and in the fine-tuning stage, the produced small sample dataset is used to help the target detection model adapt to specific class targets and improve its detection performance. This article uses TFA as the benchmark and verifies the effectiveness of the algorithm on the remote sensing datasets NWPU VHR-10 and DIOR. The results show that compared with the benchmark algorithm on the NWPU VHR-10 and DIOR datasets, the average accuracy (mAP) of the algorithm is significantly improved. In addition, the algorithms in this article are superior to various existing advanced algorithms.