针对遥感目标检测中，目标尺寸较小、相似地物易混淆和背景复杂干扰大等问题，提出了一种基于稠密连接递归特征金字塔的遥感目标检测算法。首先，为充分利用遥感图像特征，对特征融合模式进行了改进，利用典型相关分析(Canonical Correlation Analysis,CCA)，取代较为简单的逐像素相加融合模式，增强特征融合有效性；其次，为加强对小尺度目标的特征提取，加入多感受野机制，利用不同尺寸的空洞卷积提取并融合不同感受野的特征，增强网络感知力；接着，为解决遥感多尺度目标泛化问题，对特征递归形式进行改进，引入稠密连接结构，增强特征融合密度，充分利用骨干网络与高低层特征信息；最后，在前文基础上，构建稠密连接递归特征金字塔模型(Densely-connected Recursive Feature Pyramid Network, DR-FPN)，并利用融合信息实现对遥感目标的精准定位。实验结果表明，在通用数据集MS-COCO2017上，使用本文金字塔模型平均精度可以提升9.9％；在遥感数据集NWPU VHR上，使用本文算法平均精度可以提升1.1％，超过其他特征金字塔模型和检测算法，实现了对遥感目标的高精度检测。

An object detection algorithm based on densely-connected recursive feature pyramid is proposed for the problems of small targets, similar object interference and complex background in remote sensing images. First, in order to make full use of remote sensing image features, the mode of feature fusion is improved. Canonical Correlation Analysis(CCA) is used to replace the simple pixel-by-pixel addition ,which enhance the effectiveness of feature fusion; second, to stregthen the feature extraction of small targets, multiple receptive field mechanism is bulit. The features of different receptive fields are extracted and fused by dilated convolution of different sizes; third, the feature recursive structure is imporved and densely connection construction is built. It make full use of backbone network and multi-layer features; finally, Densely-connected Recursive Feature Pyramid Network(DR-FPN) is proposed on the basis of above. It can achieve precise detection of remote sensing targets. The results show: on general dataset MS-COCO2017, our algorithm can improve Average Precision(AP) by 9.9％ compared to original algorithm. On remote sensing dataset NWPU VHR, our algorithm can bring AP by 1.1％, which is superior to other algorithm and achieve high precision detection of remote sensing objects.