融合多层级特征的遥感图像地面弱小目标检测  

Multi-level feature fusion based dim small ground target detection in remote sensing images

作　　者：闫钧华[1,2] 张琨 施天俊 朱桂熠 刘勇 张寅 Yan Junhua;Zhang Kun;Shi Tianjun;Zhu Gniyi;Liu Yong;Zhang Yin(Key hiboratory of Space Photoelectric Detection and Perception(Nanjing University of Aeronautics and Astronautics),Ministry of Industry and Information Technology.Nanjing 211106,China;College oj Astronautics,JSanjing University of Aeronautics and Astronautics,Nanjing 211106,China;Research Center for Space Optical Engineering,Harbin Institute of Technology,Harbin 150001,China;National Innovation Institute of Defense Technology,Academy of Military Sciences,Beijing 100071,China)

机构地区：[1]南京航空航天大学空间光电探测与感知工业和信息化部重点实验室,南京211106 [2]南京航空航天大学航天学院,南京211106 [3]哈尔滨工业大学空间光学工程中心,哈尔滨150001 [4]军事科学院国防科技创新研究院,北京100071

出　　处：《仪器仪表学报》2022年第3期221-229,共9页Chinese Journal of Scientific Instrument

基　　金：中央高校基本科研业务费资助(NJ2020021);中央高校基本科研业务费资助(NT2020022);国家自然科学基金(61901504)项目资助。

摘　　要：为解决遥感图像地面弱小目标检测中弱小目标信息量少、信息真假混杂的难题,本文提出一种融合多层级特征的遥感图像地面弱小目标检测算法CC-YOLO。该算法首先利用深度卷积神经网络逐级对目标图像进行特征提取,得到高低层特征空间金字塔图;然后,对空间金字塔图进行跨层级通道特征融合,结合新增的位置注意力机制CA,分别沿两个空间方向聚合特征,保留弱小目标精确的位置信息;最后,在聚合后生成的双支路特征图上进行端到端的目标检测,联合多通道检测信息输出检测结果。为解决算法实验中图像数据匮乏的问题,构建了遥感图像地面弱小目标数据集GDSTD。实验结果表明,算法AP;达到42.3%,AP;达到94.6%,检测速率FPS达到58.8帧/s,具有一定的鲁棒性和实时性。The detection of dim small ground targets in remote sensing images has problems of less target information and mixed information.To address these issues,a detection algorithm based on the multi-level feature fusion is proposed in this article,which is named as CC-YOLO.Firstly,the deep convolution neural network is used to extract features of the target image step by step,and the high-level and low-level feature spatial pyramid is obtained.Then,the cross-level channel feature fusion is implemented on the spatial pyramid,and features are aggregated along two spatial directions.The newly added CA is combined to retain the accurate location information of dim small targets.Finally,the end-to-end target detection method is implemented on the dual feature map generated after aggregation.And the detection results are output by combining multi-channel detection information.To solve the problem of lacking image data in algorithm experiment,this article establishes the ground-based dim small target dataset(GDSTD)of remote sensing image.Experimental results show that the proposed algorithm achieves 42.3%at AP;and 94.6%at AP;,and the detection rate FPS reaches 58.8 frames/s,which has certain robustness and real-time performance.

关 键 词：弱小目标检测 多层级特征 融合 位置注意力机制 遥感图像 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置] TP753[自动化与计算机技术—控制科学与工程] TH701[机械工程—仪器科学与技术]