【作者】 江平；

【导师】 倪巍伟； 陈松贵；

【作者基本信息】 东南大学 ， 软件工程（专业学位）， 2019， 硕士

【摘要】 近年来,随着遥感技术的飞速发展,各种成像方式、不同空间分辨率的遥感平台不断地涌现,产生了大量的遥感图像。飞机是众多遥感图像目标中,人们关注的重要目标,遥感图像飞机目标的识别研究已经成为业界关注的焦点。传统飞机目标检测算法多数通过提取Hu矩、仿射不变矩等特征,筛选出多个特征参数,利用SVM分类器实现飞机目标识别。特征的直接组合,存在识别抗噪能力差、鲁棒性弱、耗时较长等不足。针对传统飞机检测方法的上述不足,研究基于卷积神经网络模型实现的飞机识别系统,为用户提供更加精准、快捷的飞机识别服务,这对提升飞机识别的效果具有重要的意义和实用价值。论文的主要工作如下:(1)针对常见的图像识别系统不能同时执行多个飞机识别任务,无法发挥各个模块的协同作用问题,设计基于Golang实现的多任务管理方式。将用户对指定机场的飞机识别需求看作任务派发给系统执行,这样可以实现多个识别任务并发执行,提高识别系统的效率,用户不用等待当前任务执行完毕就可以创建下一个任务。(2)针对遥感图像的尺寸较大,分辨率较高,由用户直接从系统界面输入遥感图像,耗时较长而且容易受网络传输速率的影响导致任务出错问题,设计基于Sikuli实现的图像采集方式。在服务端采用Sikuli图形化编程技术模拟用户行为从91卫图助手数据源获取遥感图像。(3)针对遥感图像飞机识别问题,提出基于卷积神经网络模型实现的图像识别算法。选用快速、准确的RetinaNet检测器作为识别模型,并与focal loss损失函数配合,采用数据增强技术扩充样本,提高了检测器的识别效率。实验结果表明,所设计实现的基于CNN的遥感图像飞机目标识别系统,具有较好的性能。更多还原

【Abstract】 In recent years,with the rapid development of remote sensing technology,various imaging methods and remote sensing platforms with different spatial resolutions have emerged continuously,resulting in a large number of remote sensing images.Aircraft is an important target in many remote sensing image targets.The identification of aircraft targets in remote sensing images has become the focus of attention in the industry.Traditional aircraft target detection algorithms mostly extract multiple feature parameters by extracting Hu moments and affine invariant moments,and use SVM classifiers to achieve aircraft target recognition.The direct combination of features has the disadvantages of poor recognition of noise immunity,weak robustness and long time consuming.Aiming at the above-mentioned shortcomings of traditional aircraft detection methods,the aircraft identification system based on convolutional neural network model is studied to provide users with more accurate and fast aircraft identification service,which has important significance and practical value for improving the recognition effect of aircraft.The main work of the thesis is as follows:(1)Aiming at problems of the system without performing multiple aircraft identification tasks at the same time that the synergy of each module can not be played,the task management method based on Golang is designed.The user identification request of the designated airport is regarded as a task dispatched to the system for execution,so that multiple identification tasks can be executed concurrently,and the efficiency of the identification system is improved,and the user can create the next task without waiting for the current task to be executed.(2)Aiming at problems of remote sensing images that the size is large and the resolution is high,The remote sensing image is input directly from the system interface by the user,which takes a long time and is easily affected by the network transmission rate,resulting in task error,the way of image acquisition based on Sikuli is designed.On the server side,Sikuli that is one of graphical programming technology is used to simulate user behavior to obtain remote sensing images from the data source of 91 Satellite Assistant.(3)For the problem of aircraft recognition of remote sensing image,an image recognition algorithm based on convolutional neural network model is proposed.The fast and accurate RetinaNet detector is selected as the recognition model with focal loss function,and the data enhancement technology is used to expand the sample,in order to enhance the recognition efficiency of the detector.The experimental results show that the designed CNN-based remote sensing image aircraft target recognition system has better performance.更多还原