高分辨率遥感图像的目标检测  

Object detection methods for high resolution remote sensing images

作　　者：梁海翔 唐艳慧 王宇庆[1] 张德浩 LIANG Haixiang;TANG Yanhui;WANG Yuqing;ZHANG Dehao(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [2]中国科学院大学,北京100049

出　　处：《液晶与显示》2024年第10期1350-1360,共11页Chinese Journal of Liquid Crystals and Displays

摘　　要：卫星遥感图像的分辨率高且目标在图像内的相对尺寸小,因此难以同时确保检测准确率和运行速度。为解决高像素遥感图像的目标检测问题,本文提出了一种结合滑动窗口分割和小目标检测器的检测方法。首先使用滑动窗口法将图像分割成多个子图,滑动步长略小于窗口的大小以使每个子图之间具有一定的重叠部分,并采用较大的分割窗口以降低子图数量。之后对子图进行压缩,使用目标检测算法处理压缩后的图片,降低算法运行时间。最后合并检测结果并采用非极大化抑制策略以去除在重叠部分重复检测的目标。在检测算法方面,本文以YOLOv8n为基础,使用SPD卷积核对网络结构进行改进,基于NWD方法调整正负样本匹配策略,并改进特征金字塔结构以提升算法对小目标的检测性能,从而使算法能够适应在更大尺寸下压缩的子图以减少图像分割数量,提升检测速度。实验证明,在图像平均分辨率为4000×4000的车辆检测数据集上,该方法对目标检测的平均准确率为55.7%,平均每张图片的计算时间约为47.5 ms,准确率比YOLOv8n提升16%,比YOLOv5s提升15%,比YOLOv6s提升7.6%。本文方法的运行效率满足实时化要求,能够以更高精度实时检测卫星遥感图像中的目标。The high resolution of satellite remote sensing images and the small relative size of the target within the image make it difficult to ensure both detection accuracy and operation speed.In order to solve the problem of target detection in high pixel remote sensing images,this paper proposes a detection method that combines sliding window segmentation and a small target detector.Firstly,the image is segmented into multiple subgraphs using the sliding window method,the sliding step is slightly smaller than the size of the window to make each subgraph have a certain overlap between them,and a larger segmentation window is used to reduce the number of subgraphs segmented.After that,the subgraphs are compressed and the compressed images are processed using a target detection algorithm to reduce the running time of the algorithm.Finally,the detection results are merged and a non-maximization suppression strategy is used to remove the targets that are repeatedly detected in the overlapping parts.In terms of detection algorithm,based on YOLOv8n,this paper uses SPD convolutional kernel and NWD to improve the network structure,and adjusts the feature pyramid structure to improve the algorithm’s performance in detecting small targets,which enables the algorithm to adapt to compressed subgraphs at larger sizes in order to reduce the number of image segmentation and improve the detection speed.The experiment proves that on the vehicle detection dataset with an average image resolution of 4000×4000,the average accuracy of the method for target detection is 55.7%,and the average computation time per image is 47.5 ms.The accuracy is improved by 17%compared to YOLOv8n,15%compared to YOLOv5s and 7%compared to YOLOv6s.The operation efficiency of the proposed method meets the real-time requirement,which is capable of detecting targets in satellite remote sensing images in real time with higher accuracy.

关 键 词：目标检测 遥感图像 YOLOv8算法 小目标检测 滑动窗口方法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]