基于视觉显著性和卷积神经网络的机场目标快速检测  

Rapid Detection of Airport Targets Based on Visual Saliency and Convolutional Neural Network

作　　者：张一民 韩现伟 张世超 高伟[1] ZHANG Yimin;HAN Xianwei;ZHANG Shichao;GAO Wei(School of Physics and Electronics,Henan University,Kaifeng 475000,China)

机构地区：[1]河南大学物理与电子学院,开封475000

出　　处：《航天返回与遥感》2021年第3期117-127,共11页Spacecraft Recovery & Remote Sensing

摘　　要：针对现有的机场目标检测算法用于大幅面遥感图像时检测速度慢、准确率低的问题,文章提出了一种基于视觉显著性和卷积神经网络相结合的高效、精确的机场目标检测方法。首先,根据机场形状特征,采用基于直线分布特征的视觉显著性检测方法提取候选区域,对机场的可能位置进行粗定位;然后,设计了一种改进的卷积神经网络分类模型判断候选区域是否为机场;最后,使用非极大值抑制的方法去除冗余的预测框,获得最终的检测结果。利用从谷歌地球收集的图像数据集对该神经网络模型进行训练和测试,结果表明其在精准率和召回率上均具有较大优势。此外,文章所提算法在来自不同卫星平台的大量大幅面遥感图像上进行了机场目标检测,结果显示其适应性强且检测效率有大幅度提升。To address the problems of slow detection speed and low accuracy when the existing airport target detection algorithms are applied to large-scale remote sensing images,this paper proposes an efficient and accurate airport target detection method based on visual saliency detection and convolutional neural network.Firstly,according to the shape characteristics of the airports,a visual saliency detection method based on the line segments distribution characteristics is used to extract the candidate regions,and the possible positions of the airport are roughly located.Then,an improved classification model based on a convolutional neural network is designed to determine whether the candidate region is an airport.Finally,the non-maximum suppression method is utilized to remove the redundant prediction bounding boxes.In the experiments,the data-set collected from Google Earth is used to train and test the network model in this paper,and the results show that the model has great advantages in both precision and recall rate.In addition,the presented algorithm in this paper is used to detect airport targets on a large number of large-scale remote sensing images from different satellite platforms.The experimental results indicate that the algorithm has strong adaptability and the detection efficiency has been greatly improved.

关 键 词：遥感图像 机场目标 视觉显著性 卷积神经网络 迁移学习 遥感应用 

分 类 号：P407.8[天文地球—大气科学及气象学]