基于级联特征融合的遥感图像目标检测技术研究  

Research on Remote Sensing Image Object Detection Technology Based on Cascade Feature Fusion

作　　者：王留洋 芮挺 Wang Liuyang;Rui Ting(Army Engineering University of PLA,Nanjing 201007,China)

机构地区：[1]陆军工程大学,江苏南京201007

出　　处：《长江信息通信》2021年第8期15-18,23,共5页Changjiang Information & Communications

摘　　要：针对高分辨率遥感图像中目标尺度不一且排列分布密集导致检测漏检率高精度低的问题,提出一种基于级联特征融合的检测算法,用于遥感图像目标检测。首先,提出级联并行扩张卷积结构,并结合特征金字塔应用在主干网络的76×76、19×19尺度特征层,丰富网络浅层深层图像细节特征,使网络能够在把握特征图全局重要特征信息的同时,提高模型对小目标的感知能力;其次,利用k-means++聚类算法对遥感图像数据集进行分析,生成适合遥感图像目标的锚框,减少锚框配准时间,提高模型的检测效率;最后,后处理机制利用Soft-nms算法来抑制冗余检测框,减少密集分布目标的漏检率。通过在RSOD数据集上测试,在检测效率达到实时性的同时,mAP达到了95.20%,比YOLOv4提高了2.02%,证明了所提算法的有效性。Aiming at the problem of low-precision miss detection rate caused by the different scales of objects in high-resolution remote sensing images and the densely arranged and distributed objects,a detection algorithm based on cascaded feature fusion is proposed for object detection in remote sensing images.First,the cascade parallel expansion convolution structure is proposed,and the feature pyramid is applied to the 76×76 and 19×19 scale feature layers of the backbone network to enrich the detailed features of the shallow and deep images of the network,so that the network can grasp the global important features of the feature map.At the same time,it improves the model's ability to perceive small objects.Secondly,the k-means++clustering algorithm is used to analyze the remote sensing image data set to generate an anchor frame suitable for the remote sensing image object,reducing the anchor frame registration time and improving the detection of the model Efficiency;Finally,the post-processing mechanism uses the Soft-nms algorithm to suppress redundant detection frames and reduce the missed detection rate of densely distributed objects.Through testing on the RSOD data set,while the detection efficiency reaches real-time,mAP reaches 95.20%,which is 2.02%higher than YOLOv4,which proves the effectiveness of the proposed algorithm.

关 键 词：目标检测 遥感图像 级联特征 扩张卷积 Soft-nms 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]