多光谱遥感图像的特征提取与比较

III
compare the spectrum features of multi-spectral images. Using this algorithm, the computation is greatly reduced and the arguments turn to be dynamic. The procedure of cluster splitting and merging is based on the result of primary classification. By changing absolute values into ratio values, dynamic parameters are realized to normalize the required parameters in the iteration. Without setting the number of the iteration, it is completed until a balance is reached. (2) Texture feature is one of the attributes of image, which describes the space distribution of gray levels of image pixels. An image contains texture if the objects in the image have a distinct but not simple hue change. Texture feature extraction in this thesis is based on Least Squares method and region segmentation. The contributions of this algorithm: the coefficient vectors achieved by Least Squares method properly express the texture information of the multi-spectral images, and the concept of texture in single-band image is developed to that of multi-spectral images. The shrinking-expanding method is proposed to regulate the coefficient vectors because of the anomaly of complicated texture. In processing of region segmentation, a method of transforming the open region into close region and normalizing the closed region is proposed. (3) Shape feature is also called contour feature, which describes the edge characteristics of image or part of image. Shape feature extraction in this thesis is based on band grouping and moment invariants. A method of dividing the bands of multi-spectral images into groups is proposed and realized based on the attributes
本文采用的光谱特征提取方法采用基于改进 ISODATA(Iterative
Self-Organizing Data Analysis Techniques A, 迭代自组织数据分析技术 A)算法 的聚类分析方法 与原算法相比 改进算法的优点如下 在保留原算法初始
聚类的成果的基础上 以类自身的状态作为合并与分裂是否进行的判定标准 极大的降低了计算量 参数动态化 将绝对性质的参数转变为比值使得原算法循环里面的
该算法具有以下创新
相关程度将多光谱图像的波段分组的方法 合并形状特征相似的区域的方法
(4) 特征比较 在三种特征提取的基础上 本文提出了四条矢量特征比较 的标准 用于比较两幅多光谱图像特征提取完成后的比较 通过比较可以反 纹理 形状特征上的相似程度 提出了新的提取多光谱图像
映出两幅多光谱图像在光谱
本文综合多光谱图像特征提取的常规方法 光谱 纹理 形状特征的方法
Baidu NhomakorabeaIV
of spectrograph or the correlation degree between the bands. And a method of merging the regions of similar shape feature is proposed and realized based on moment invariants. (4) Feature comparison: Based on the three feature extraction methods, four rules are designed to compare the features of the multi-spectral images. The comparison shows the similarity or the differences in spectrum, texture and shape features of two multi-spectral images. In this thesis, the usual methods on feature extraction of multi-spectral images are introduced, three new methods are designed to extract the spectrum, texture and shape features of multi-spectral images, and the comparison of the feature vectors of two multi-spectral images is proposed. Finally, the simulation of the three methods using MATLAB achieves good results.
上海交通大学 硕士学位论文 多光谱遥感图像的特征提取与比较 姓名：刘磊 申请学位级别：硕士 专业：控制理论与控制工程 指导教师：敬忠良 20050101
多光谱遥感图像的特征提取与比较
摘 要
基于内容的图像检索方式 CBIR(Content-Based Image Retrieval) 就是根据 给定的图像特征 从存储在数据库中的大量图像中进行检索 找出与给定图 基于内容的图像检索主要涉及到四项关键技术 特征提取与匹配技术 快速检索技术 图像
像特征相似的图像来 数据库技术
内容描述技术
本文的研究内容着重于多光谱遥感图像的特征提取与比较上 从光谱特 征 如下 (1) 光谱特征 通过原始波段的点运算获得的图像中目标物的颜色及灰度 或者波段间亮度的比较 构无关 光谱特征对应于每个像素 与像元的排列等空间结 纹理特征 形状特征三个方面进行研究 理论与方法部分的创新和成果
很好的表达了多光谱图像的纹理信息 缩放法针对较复杂纹理的不 提出了对系数矢量进行调整的方法 在区域分割的过程中 提出了
将开区域转化为闭区域和将闭区域规则化的方法 (3) 形状特征 也称为轮廓特征 是指整个图像或者图像中子对象的边缘 特征和区域特征 特征 本文采用基于波段分组和不变矩的聚类分析方法提取形状 提出并实现了基于传感器成像特性或者波段间 提出并实现了基于不变矩矢量来
原算法叠代结束的条件是由叠代次数人为控制的 改进算法是
I
以类自身达到一种内部平衡作为叠代结束的判定标准的
合理性更强
(2) 纹理特征 一种反映图像像素灰度级空间分布的属性 如果物体内部 的灰度级变化明显又不是简单的色调变化 那么该物体就有纹理 本文采用
的纹理特征提取方法采用基于最小二乘和区域分割技术的聚类分析方法 该 算法具有以下创新 的发展 规则性 通过最小二乘法拟合的系数矢量是对单幅图像纹理表达
KEY WORDS: multi-spectral remote sensing images, spectrum feature, texture feature, shape feature
V
上海交通大学 学位论文原创性声明
本人郑重声明 所呈交的学位论文 是本人在导师的指导下 独 立进行研究工作所取得的成果 除文中已经注明引用的内容外 本论 文不包含任何其他个人或集体已经发表或撰写过的作品成果 对本文 的研究做出重要贡献的个人和集体 均已在文中以明确方式标明 本 人完全意识到本声明的法律结果由本人承担
并相应提出了比较两幅多光谱图像的特征矢
量比较方法
最后给出了 MATLAB 仿真实现结果
关键字
多光谱遥感图像
光谱特征
纹理特征
形状特征
II
FEATURE EXTRACTION AND COMPARISON OF MULTI-SPECTRAL REMOTE SENSING IMAGES
ABSTRACT
Content-Based Image Retrieval (CBIR) is used to find out the target image from the image database according to the given image features. The image features can be extracted from the sample images provided or inputted by customers. CBIR mainly contains four key techniques, which are image database, content description, feature extraction and matching and fast searching. This thesis deals with the feature extraction and comparison of multi-spectral remote sensing images. It contains three aspects, which are spectrum feature extraction, texture feature extraction and shape feature extraction. The main achievements and contributions about methods and algorithms are described as follows: (1) Spectrum feature is defined as the color or gray value of target in images, or comparison of Intensity between bands of the images. It corresponds to every pixel, and is independent to space structure. Improved ISODATA (Iterative Self-Organizing Data Analysis Techniques A) algorithm is provided to extract and