基于支持张量机的遥感图像目标分级识别研究
发布时间:2018-03-06 04:02
本文选题:遥感图像 切入点:目标分级识别 出处:《哈尔滨工业大学》2017年硕士论文 论文类型:学位论文
【摘要】:遥感图像的目标识别是空间遥感技术的主要应用之一。近年来,随着遥感图像分辨率的提高,越来越多的目标信息能够从遥感图像中挖掘出来,为描述目标提供了重要的支撑。因此,在遥感图像应用领域,对目标进行进一步地细致的研究已经受到了越来越多的重视。但是,遥感图像分辨率的提升也为遥感图像的处理带来了许多问题,通常来说,传统的方法一般只停留在对目标自身的关注,很少更深入地对目标的细节部位进行识别,同时,在处理高数据量的高分辨率图像时,传统利用向量来描述图像中的特征的方法效率不高。本文主要针对这些问题提出了基于支持张量机模型的遥感图像目标分级识别方法,通过对关注的目标进行由整体到局部的分级识别,更有效地把我目标的细节信息,并利用特征张量模型描述图像中的目标,由于张量能够保持数据的空间结构,因而能够取得较好的结果。本文首先研究了目标的局部空间特征提取与特征张量表达模型的构建。遥感图像自身就可以看作是张量的数据形式,因此利用张量可以更直接地描述图像数据的空间坐标与光谱信息,同时结合对图像局部空间特征的提取得到目标图像相关特征信息,进行目标图像的特征张量描述。为了更好地保持张量的空间特性,本课题采用了局部空间不变特征与加速鲁棒特征两种局部空间特征来建立特征张量模型,并研究分析了它们的优缺点与适用情况。之后,本文对支持张量机学习分类模型的学习训练算法与过程进行了阐述。利用张量来实现目标的分类识别能够更好地利用图像自身的空间结构信息,并在一定程度上减少向量模型中维数灾难的发生。本文通过一般的支持向量机算法引出支持张量机的学习模型,本课题主要利用基于梯度下降算法的学习模型对支持张量机进行了训练,通过对训练样本特征张量模型与对应分类值的反复迭代,求得最优分类超平面,对关注的目标和背景进行区分。在建立好支持张量机模型之后,通过不同的训练样本训练得到不同的分类器,进行目标的分级识别实验研究。本文通过对飞机、舰船两类目标以及它们的细节部位等进行实验测试,验证了支持张量机模型在目标分级识别应用中的准确性与可靠性。实验结果表明,利用支持张量机模型能够比较有效地实现遥感图像中目标的分级识别,方法具备一定的实际意义与应用价值。
[Abstract]:Object recognition of remote sensing image is one of the main applications of space remote sensing technology. In recent years, with the improvement of remote sensing image resolution, more and more target information can be extracted from remote sensing image. Therefore, in the field of remote sensing image application, more and more attention has been paid to the study of target. The improvement of remote sensing image resolution also brings many problems for remote sensing image processing. Generally speaking, the traditional methods only focus on the target itself, and rarely identify the target's detailed parts more deeply, at the same time, the traditional methods only focus on the target itself, at the same time, When dealing with high resolution images with high data volume, the traditional method of using vectors to describe the features of images is not efficient. In this paper, a classification recognition method for remote sensing images based on Zhang Liang model is proposed in order to solve these problems. By classifying the objects of concern from the whole to the local level, we can more effectively describe the details of our targets, and use the feature Zhang Liang model to describe the targets in the image, because Zhang Liang can maintain the spatial structure of the data. Therefore, better results can be obtained. Firstly, this paper studies the local spatial feature extraction of the target and the construction of the Zhang Liang expression model. The remote sensing image itself can be regarded as the data form of Zhang Liang. Therefore, Zhang Liang can more directly describe the spatial coordinate and spectral information of the image data, and combine with the extraction of the local spatial features of the image to obtain the relevant feature information of the target image. In order to better maintain the spatial characteristics of Zhang Liang, two local spatial features, local spatial invariant feature and accelerated robust feature, are adopted to establish the feature Zhang Liang model. After studying and analyzing their advantages and disadvantages and their application. In this paper, the algorithm and process of learning and training supporting Zhang Liang's machine learning classification model are expounded. The spatial structure information of image itself can be better utilized by using Zhang Liang to realize target classification and recognition. And to some extent reduce the occurrence of dimensionality disaster in the vector model. In this paper, the general support vector machine algorithm is used to derive the learning model of support Zhang Liang machine. In this paper, we mainly use the learning model based on gradient descent algorithm to train Zhang Liang machine, and get the optimal classification hyperplane by iterating the training sample feature Zhang Liang model and the corresponding classification value. After establishing the model of supporting Zhang Liang machine, different classifiers are obtained by training different training samples, and the classification recognition experiment of target is carried out. The accuracy and reliability of supporting Zhang Liang machine model in target classification recognition are verified by experimental tests on two kinds of ship targets and their detailed positions. The experimental results show that, The classification recognition of objects in remote sensing images can be realized effectively by using the support Zhang Liang machine model. The method has certain practical significance and application value.
【学位授予单位】:哈尔滨工业大学
【学位级别】:硕士
【学位授予年份】:2017
【分类号】:TP751
【参考文献】
相关硕士学位论文 前2条
1 周艳果;高分辨率光学遥感数据海上船舶提取[D];大连海事大学;2016年
2 周蓉;支持张量机的在线学习算法研究[D];华南理工大学;2014年
,本文编号:1573212
本文链接:https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/1573212.html