【作者】 王巍；

【导师】 袁春；

【作者基本信息】 中国地质大学(北京) ， 资源管理工程， 2012， 硕士

【摘要】 地物信息的自动提取一直是遥感影像处理、土地利用分类及动态监测等领域的研究热点，目前已经形成了一些比较成熟的地物提取方法，如监督和非监督分类、面向对象信息提取等，均可得到较好的效果；其中，BP神经网络和支持向量机则非常适合于高分辨率遥感影像上的小地物识别，但仍然存在算法效率低，分类器训练速度较慢、依赖经验性等问题。本文以影像上的汽车为例，在利用WorldView-2全色影像数据进行地物识别研究的基础上，提出了在Microsoft Visual C++平台下，借助OpenCV函数库，从影像上提取汽车样本图像的Haar特征，并采用Adaboost算法训练分类器，在高分辨率遥感影像上进行小地物的自动识别、提取的新方法，通过具体的实验效果探讨了这种方法的适用性与可行性，总结归纳出了完整的技术方法流程及相关程序的说明；在此基础之上，深入研究了影像分辨率改变对汽车识别效果的影响，通过实验论述了遥感影像应用方面的尺度问题。本文的主要研究内容和结论包括以下三个方面：（1）在总结前人研究的基础上，利用WorldView-2全色影像，以汽车为例，通过实验论证了提取目标地物的Haar特征，采用Adaboost算法训练分类器的目标识别方法可以应用于高分辨率遥感影像上的小地物识别，拓展了算法的应用领域；（2）归纳总结了完整的实验技术方法流程，并对相关识别程序代码进行了说明，为遥感影像上特定物体的识别提供参考；（3）研究了影像分辨率改变对汽车识别效果的影响，在一定程度上论证了遥感影像应用的尺度问题，指出要识别一定尺度的地物所需影像的分辨率必须满足一定的条件。更多还原

【Abstract】 Automatic feature extraction has always been a hot topic in the area ofremote sense image processing, land us classification and land use dynamicmonitoring. There are many well developed methods, such as classification withand without monitoring, object-oriented feature extraction, among which BPNeural Networks and Support Vector Machine methods are better for smallobject detection in the high resolution remote sense image. In general, theabove methods can produce good results, but most of them have the problems oflow efficiency, low classifier-training speed and experience dependent. Basedon the study of vehicle targets detection in WorldView-2full-color image, weproposed a new approach in platform of Microsoft Visual C++for automaticallyidentifying small targets in high-resolution remote sense image. This methoduses OpenCV Library to extract Haar features of samples and then trains theclassifier with Adaboost Algorithm. We discussed the applicability andfeasibility of this method through experimental results and presented thecomplete technical procedure and program scheme. Also we studied the effectof image resolution on the detection of vehicle targets and discussed thepossible scale problem in application.The main content of this article can be summarized as follow:(1) Based on the previous studies, a new method of object identification usingHaar features and Adaboost classification training algorithm is proposed, whichcan be used for small object detection in the high resolution remote sense image.Expand the applications of Adaboost Algorithm.(2) The complete technical procedure and program scheme are summarized forfurther application of special objects detection in high-resolution remote senseimage.(3) The effect of image resolution on the detection of vehicle targets is studied,as well as the necessary resolution for the detected targets of given scale.更多还原