多尺度分割原理与应用

遥感分类前所进行的图像分割

Hierarchical HSeg SegmentationReading Report

遥感分类前所进行的图像分割

CONTENTS

PART ONE

About the Introduction.

PART TWO

The Hseg Segmentation.

PART THREE

Application in ENVI.

遥感分类前所进行的图像分割

1

PART ONEAbout the Introduction.

遥感分类前所进行的图像分割

AbstractRecent advances in spectral–spatial classification of hyperspectral images are presented in this paper. Several techniques are investigated for combining both spatial and spectral information. Highlight the importance of spectral–spatial strategies for the accurate classification of hyperspectral images and validate the proposed methods.

遥感分类前所进行的图像分割

IntroductionAdvantages of hyperspectral remote sensor technology: The detailed spectral information increases the possibility of more accurately discriminating materials of interest. The fine spatial resolution of the sensors enables the analysis of small spatial structures in the image. Many operational imaging systems are currently available providing a large amount of images for various thematic applications. But,it also brings some problem: the Curse of Dimensionality and the need for specific spectral–spatial classifiers.

PS: 维数灾难(Curse of Dimensionality):通常是指在涉及到向量的计算的问题中，随着维数的增加，计算 量呈指数倍增长的一种现象。维数灾难在很多学科中都可以碰到，比如动态规划，模式识别和影像识别 等。

遥感分类前所进行的图像分割

IntroductionThe Curse of Dimensionality of hyperspectral remote sensor technology : In high-dimensional spaces, normally distributed data have a tendency to concentrate in the tails, which seems to be contradictory with its bell-shaped density function. the rate of convergence of the statistical estimation decreases when the dimension grows while conjointly the number of parameters to estimate increases, making the estimation of the model parameters very difficult. with a limited training set, beyond a certain limit, the classification accuracy actually decreases as the number of features increases.

遥感分类前所进行的图像分割

IntroductionHow to build accurate classifiers for hyperspectral images? SVMs perform a nonlinear pixel-wise classification based on the full spectral information which is robust to the spectral dimension of hyperspectral images. Iterative statistical classifier based on Markov random field (MRF) modeling. Note that recently adaptive MRF have been introduced in remote sensing. Use advanced morphological filters as an alternative way of performing joint classification.

PS: 鲁棒性(Robust):即系统的健壮性，是在异常和危险情况下系统生存的关键。

遥感分类前所进行的图像分割

Introduction

遥感分类前所进行的图像分割

2

PART TWOThe Hseg Segmentation.

遥感分类前所进行的图像分割

The Hseg SegmentationThe HSeg algorithm is a segmentation technique combining region growing, using the hierarchical stepwise optimization (HSWO) method , which produces spatially connected regions, with unsupervised classification, that groups together similar spatially disjoint regions. The algorithm can be summarized as follows.

Initialization: Initializ

e the segmentation by assigning each pixel a region label. If apresegmentation is provided, label each pixel accordingly. Otherwise, label each pixel as a separate region.

遥感分类前所进行的图像分割

The Hseg Segmentation

Calculate the dissimilarity criterion value between all pairs of spatially adjacent regions

III III IV

Merge spatially adjacent regions

Merge spatially nonadjacent regions

Stop if convergence is achieved. Otherwise, return to step I

遥感分类前所进行的图像分割

The Hseg Segmentation

I

计算每一对空间相邻区域的相异准则值(dissimilarity criterion value),例如计算向量 模或者区域平均向量间的光谱角填图SAM。 SAM是基于物理的一种光谱分类，利用n维角度来匹配两组像元光谱，将光谱看成是维数 与波段数相等的空间里的向量，计算光谱间的角度的算法，决定了两个光谱之间的相似性， 也就是通过计算两向量之间的广义角来确定它们的相似性(角度越小越相似)。 这种方法充分利用了光谱维的信息。

遥感分类前所进行的图像分割

The Hseg Segmentation

II

找到最小的相异准则值并设为阈值，如果两相邻区域的相异准则值等于阈值则合并这两个 区域。 即dissim_val=thresh_val时，合并两相邻区域。

遥感分类前所进行的图像分割

The Hseg Segmentation

III

如果Swght>0.0,则将dissim_val <Swght*thresh_val的不相邻区域合并( Swght是衡量 基于光谱信息的聚类相对于区域增长的相对重要性的选择性参数，当Swght=0.0时，仅空 间上相邻的区域才允许合并， 当0<Swght<1 时，则空间相邻区域相对于不相邻区域有一 1/Swght的系数更可能被合并)

遥感分类前所进行的图像分割

The Hseg Segmentation

IV

如果合并完成则结束，否则回到步骤I。 通过迭代运算，最终形成从细到粗的多尺度、多层次的分割结果。

遥感分类前所进行的图像分割

The Hseg Segmentation因为要对空间上非相邻的区域进行合并，运算量变得非常巨大。为解决运算量问题，发展 出了迭代分而治之的多尺度分割近似计算及其相应的有效实现方式。 多尺度分割能输出一个从初始化分割直至一个区域的分割的多尺度序列，在这个序列中， 一个特定的对象既可以表示成几个区域，从而具有较好的细节信息，也可以与其他对象一 起被一个区域吸收。在实际应用中，可以通过HSegViewer程序交互式的选择适当的水平 或者选择其他为应用而定制的自动化方法。