CN101510310A - Method for segmentation of high resolution remote sensing image based on veins clustering constrain - Google Patents

Abstract

Translated fromChinese

一种遥感技术领域的基于纹理聚类约束的高分辨率遥感图像分割方法，具体步骤为：第一步，计算图像中所有区域的Gabor能量纹理，并按照纹理距离使用FCM对所有区域进行聚类，根据聚类结果给每个区域设定纹理聚类标签；第二步，使用光谱、纹理和形状等特征建立综合距离空间模型，并加入纹理聚类距离对区域合并进行约束，使合并可以沿着纹理同质的方向进行；第三步，根据综合距离建立RAG和NNG模型，并按照全局最优对区域对进行合并。在合并过程中通过纹理聚类和最优合并序列的相互作用得到区域的真实边界。本发明能够较好地分割出高分辨率图像中纹理区域，提高图像的总体分割精度。A high-resolution remote sensing image segmentation method based on texture clustering constraints in the field of remote sensing technology, the specific steps are: first, calculate the Gabor energy texture of all regions in the image, and use FCM to cluster all regions according to the texture distance , set a texture clustering label for each region according to the clustering results; in the second step, a comprehensive distance space model is established using features such as spectrum, texture, and shape, and the texture clustering distance is added to constrain the region merging, so that the merging can be along the In the third step, the RAG and NNG models are established according to the comprehensive distance, and the region pairs are merged according to the global optimum. In the merging process, the real boundary of the region is obtained through the interaction of texture clustering and optimal merging sequence. The invention can better segment the texture area in the high-resolution image and improve the overall segmentation accuracy of the image.

Description

Translated fromChinese

基于纹理聚类约束的高分辨率遥感图像分割方法Segmentation Method of High Resolution Remote Sensing Image Based on Texture Clustering Constraint

技术领域technical field

本发明涉及的是一种遥感技术领域的图像分割方法，具体是一种基于纹理聚类约束的高分辨率遥感图像分割方法。The invention relates to an image segmentation method in the technical field of remote sensing, in particular to a high-resolution remote sensing image segmentation method based on texture clustering constraints.

背景技术Background technique

随着卫星空间分辨率的不断提高，面向对象的图像分析方法开始在遥感图像处理领域广泛应用。相对于基于像素的分类方法，面向对象的图像分析方法可以更大程度的降低噪声的影响，提取更多的特征，且易于与GIS(GeographicalInformation System，地理信息系统)相结合。但面向对象图像分析方法中的关键部分——图像分割——却始终没有得到很好的解决。With the continuous improvement of satellite spatial resolution, object-oriented image analysis methods have been widely used in the field of remote sensing image processing. Compared with the pixel-based classification method, the object-oriented image analysis method can reduce the influence of noise to a greater extent, extract more features, and is easy to combine with GIS (Geographical Information System, geographic information system). But the key part of the object-oriented image analysis method - image segmentation - has not been well resolved.

按照特征空间，图像分割方法可以大体分为三类：基于光谱的分割方法、基于纹理的分割方法和基于光谱与纹理将结合的分割方法，其中基于光谱分割的方法是现阶段最常用的方法。德国的Definiens Developer Software公司提出了一种基于光谱的遥感图像分割方法，即寻找遥感图像中光谱同质性区域，具体为：根据合并后区域的光谱方差变化判断待合并区域间的光谱相似性程度，然后按照光谱相似性程度，依次把具有较高相似度的区域合并，当所有区域间的相似度都大于阈值时终止合并，得到分割结果。这种方法在该公司的软件Definiens中得到了应用。According to the feature space, image segmentation methods can be roughly divided into three categories: spectrum-based segmentation methods, texture-based segmentation methods, and segmentation methods based on the combination of spectrum and texture. Among them, the method based on spectral segmentation is the most commonly used method at this stage. Definiens Developer Software Company of Germany proposed a remote sensing image segmentation method based on spectrum, which is to find the spectral homogeneity area in the remote sensing image, specifically: judge the spectral similarity between the areas to be merged according to the spectral variance change of the merged area , and then according to the degree of spectral similarity, the regions with higher similarity are merged sequentially, and when the similarity between all regions is greater than the threshold, the merger is terminated, and the segmentation result is obtained. This approach is used in the company's software, Definiens.

经对现有技术文献的检索发现，U.Benz等人在《ISPRS Journal ofPhotoGrammetry and Remote Sensing》(ISPRS摄影测量和遥感)，Volume.58，page.239-258上发表的“Multi-resolution，object-oriented fuzzy analysisof remote sensing data for GIS-ready information”(使用遥感图像提取GIS所需信息的多分辨率面向对象模糊分析方法)一文中给出了详细的介绍。但在高分辨率遥感图像中，更多的区域是纹理区域，其在光谱上并不同质。因此，使用Definiens软件中的光谱分割方法对高分辨率遥感图像进行分割时，很多的纹理区域被分为破碎的小块，不能得到完整的区域，分割精度较低。After searching the prior art documents, it was found that "Multi-resolution, object -oriented fuzzy analysis of remote sensing data for GIS-ready information" (a multi-resolution object-oriented fuzzy analysis method using remote sensing images to extract information required by GIS) gives a detailed introduction. But in high-resolution remote sensing images, more regions are texture regions, which are not spectrally heterogeneous. Therefore, when using the spectral segmentation method in the Definiens software to segment high-resolution remote sensing images, many texture areas are divided into broken small blocks, and complete areas cannot be obtained, and the segmentation accuracy is low.

使用光谱和纹理相结合的方法是提高高分辨率遥感图像分割精度的方法之一。Y.Deng等人在《IEEE Trans.Pattern Anal.Mach.Intell.》(IEEE模式识别与机器职能)，Volume.14，2001，page.800-810上发表的“Unsupervisedsegmentation of spectral-texture regions in images and video”(基于图像和视频中光谱纹理区域的非监督分割方法)一文中提出一种光谱和纹理相结合的JSEG方法，具体为：将光谱量化，并根据局部量化光谱的频数得到空间纹理J图像，然后对J图像使用区域生长方法分割得到最后的结果。该方法在对场景图像分割时取得了较好的结果，但高分辨率遥感图像内容复杂，对光谱量化会降低图像信息量，得到的纹理描述不准确，也较难准确得到地物的真实边界。并且JSEG方法将纹理和光谱分开处理，没有实现光谱和纹理语义上的结合。如何使用多种特征分割高分辨率遥感图像，并且得到图像中纹理区域的准确边界仍然是现阶段没有解决的问题之一。The method of combining spectrum and texture is one of the methods to improve the segmentation accuracy of high-resolution remote sensing images. "Unsupervised segmentation of spectral-texture regions in images published by Y.Deng et al. in "IEEE Trans.Pattern Anal.Mach.Intell." (IEEE Pattern Recognition and Machine Functions), Volume.14, 2001, page.800-810 and video" (Unsupervised Segmentation Method Based on Spectral Texture Regions in Images and Videos) proposes a JSEG method combining spectrum and texture, specifically: quantize the spectrum, and obtain the spatial texture according to the frequency of the local quantized spectrum J image, and then use the region growing method to segment the J image to get the final result. This method has achieved good results when segmenting scene images, but the content of high-resolution remote sensing images is complex, quantifying the spectrum will reduce the amount of image information, the obtained texture description is inaccurate, and it is difficult to accurately obtain the real boundaries of ground objects. . Moreover, the JSEG method treats texture and spectrum separately, and does not realize the semantic combination of spectrum and texture. How to use multiple features to segment high-resolution remote sensing images and obtain accurate boundaries of texture regions in the image is still one of the unsolved problems at this stage.

发明内容Contents of the invention

本发明目的在于克服现有技术中高分辨率遥感图像分割中纹理信息利用不足的缺陷，提出一种基于纹理聚类约束的高分辨率遥感图像分割方法，通过纹理聚类预处理影响分割时区域合并的顺序，使区域按照纹理同质的方向进行合并，并且在合并过程中使用最优合并序列和纹理聚类的相互作用找到了区域的准确边界，提高了分割的精度。The purpose of the present invention is to overcome the defect of insufficient use of texture information in the high-resolution remote sensing image segmentation in the prior art, and propose a high-resolution remote sensing image segmentation method based on texture clustering constraints, which can affect the region merging during segmentation through texture clustering preprocessing The order of the region is merged according to the direction of texture homogeneity, and the interaction of the optimal merge sequence and texture clustering is used to find the accurate boundary of the region during the merging process, which improves the accuracy of the segmentation.

本发明是通过以下技术方案实现的，本发明包括步骤如下：The present invention is realized through the following technical solutions, and the present invention comprises steps as follows:

第一步，获取区域纹理聚类标签：计算图像中所有区域的Gabor能量纹理(高波能量纹理)，并按照纹理距离使用FCM(模糊C均值聚类算法)对所有区域进行聚类，根据聚类结果给每个区域设定纹理聚类标签；The first step is to obtain the regional texture clustering label: calculate the Gabor energy texture (high wave energy texture) of all regions in the image, and use FCM (fuzzy C-means clustering algorithm) to cluster all regions according to the texture distance, according to the clustering As a result, a texture clustering label is set for each region;

第二步，建立语义一致的距离空间模型：使用光谱、纹理和形状等特征建立综合距离空间模型，并加入纹理聚类距离对区域合并进行约束，使合并可以沿着纹理同质的方向进行；The second step is to establish a semantically consistent distance space model: use features such as spectrum, texture, and shape to establish a comprehensive distance space model, and add texture clustering distance to constrain the region merging, so that the merging can be carried out along the direction of texture homogeneity;

第三步，使用图模型算法对区域进行快速合并：根据综合距离建立RAG(区域连接图)和NNG(最近邻接图)模型，并按照全局最优对区域对进行合并，得到最终分割结果。在合并过程中通过纹理聚类和最优合并序列的相互作用得到区域的真实边界。The third step is to use the graph model algorithm to quickly merge the regions: establish the RAG (Region Connection Graph) and NNG (Nearest Neighbor Graph) models according to the comprehensive distance, and merge the region pairs according to the global optimum to obtain the final segmentation result. In the merging process, the real boundary of the region is obtained through the interaction of texture clustering and optimal merging sequence.

所述计算区域Gabor能量纹理，并使用FCM对区域聚类，得到区域纹理聚类标签，具体为：使用3尺度、8方向Gabor滤波对图像进行卷积，取对应的对称和反对称Gabor滤波结果计算算数平方和的根得到Gabor能量，Gabor能量为24维向量，使用该向量的欧式距离进行聚类得到区域纹理聚类标签。The calculation area Gabor energy texture, and use FCM to cluster the area to obtain the regional texture clustering label, specifically: use 3-scale, 8-direction Gabor filter to convolve the image, and take the corresponding symmetric and anti-symmetric Gabor filter results The root of the arithmetic sum of squares is calculated to obtain the Gabor energy, which is a 24-dimensional vector, and the Euclidean distance of the vector is used for clustering to obtain the regional texture cluster label.

所述语义一致距离空间，具体为：使用纹理聚类标签对区域合并顺序进行约束使区域合并按照纹理区域同质方向进行。然后综合使用光谱、纹理和形状等信息，并加入纹理聚类距离，建立使用多特征的距离空间模型。在这个距离空间模型中，区域间纹理同质时，光谱距离较小；光谱同质时，纹理距离较小。The semantic consistent distance space specifically includes: using texture clustering labels to constrain the order of region merging so that region merging is performed in a homogeneous direction of texture regions. Then comprehensively use information such as spectrum, texture and shape, and add texture clustering distance to establish a distance space model using multiple features. In this distance-space model, the spectral distance is smaller when the textures between regions are homogeneous, and the texture distance is smaller when the spectra are homogeneous.

所述基于RAG和NNG图的快速合并方法，具体为：使用RAG图描述区域之间邻接关系，并记录区域间的合并距离；使用NNG图描述每个区域对应的最优合并区域，并记录最优合并距离。在NNG图中，互为最优合并区域的区域对为图像内的局部最优合并区域对，将所有的局部最优区域对排序得到全局最优合并序列，按照该序列合并即可得到分割结果。在合并过程中，合并后的区域纹理聚类标签发生改变，进而影响序列的合并顺序，通过这种相互作用可以得到地物的真实边界。The fast merging method based on the RAG and NNG graphs is specifically: use the RAG graph to describe the adjacency relationship between the regions, and record the merged distance between the regions; use the NNG graph to describe the optimal merged region corresponding to each region, and record the most Optimal merge distance. In the NNG graph, the region pairs that are mutually optimal merged regions are the local optimal merged region pairs in the image, and all the local optimal region pairs are sorted to obtain the global optimal merged sequence, and the segmentation result can be obtained by merging according to this sequence . During the merging process, the texture clustering labels of the merged regions change, which in turn affects the merging order of the sequences, and the real boundaries of the ground objects can be obtained through this interaction.

这个图方法大大加快了合并的效率，降低了分割算法的时间复杂度。This graph method greatly speeds up the efficiency of merging and reduces the time complexity of the segmentation algorithm.

本发明能够较好地得到高分辨率遥感图像的分割结果，结果中纹理区域较为完整，边界比较准确，分割精度较高。The invention can better obtain the segmentation result of the high-resolution remote sensing image, and in the result, the texture area is relatively complete, the boundary is relatively accurate, and the segmentation precision is high.

本发明将纹理聚类引入到图像分割算法中，着重解决了如何使用多种特征准确分割高分辨率遥感图像的问题。在分割算法中，充分考虑了光谱同质区域与纹理同质区域之间的关系，使图像中的纹理区域可以准确地分割出来。此外，算法通过纹理聚类和最优区域合并的相互作用得到了区域的准确边界。这个算法避免了传统遥感图像分割方法中的缺陷：使用单纯光谱特征较难分割出正确纹理区域。因此，与传统方法相比，本发明具有更高的分割精度。本发明提出的基于纹理聚类约束的分割算法，将纹理信息和光谱等信息综合使用，大幅提高了高分辨率遥感图像的分割精度，着力解决遥感影像处理领域使用多特征分割的问题，对学科的发展具有推动作用。The invention introduces texture clustering into the image segmentation algorithm, and focuses on solving the problem of how to accurately segment high-resolution remote sensing images using multiple features. In the segmentation algorithm, the relationship between the spectral homogeneous area and the texture homogeneous area is fully considered, so that the texture area in the image can be accurately segmented. Furthermore, the algorithm obtains accurate boundaries of regions through the interplay of texture clustering and optimal region merging. This algorithm avoids the defect in the traditional remote sensing image segmentation method: it is difficult to segment the correct texture area by using pure spectral features. Therefore, compared with traditional methods, the present invention has higher segmentation accuracy. The segmentation algorithm based on texture clustering constraints proposed by the present invention comprehensively uses texture information and spectrum information, greatly improves the segmentation accuracy of high-resolution remote sensing images, and strives to solve the problem of using multi-feature segmentation in the field of remote sensing image processing. development has a stimulating effect.

具体实施方式Detailed ways

下面对本发明的实施例作详细说明：本实施例在以本发明技术方案为前提下进行实施，给出了详细的实施方式和具体的操作过程，但本发明的保护范围不限于下述的实施例。The embodiments of the present invention are described in detail below: the present embodiment is implemented under the premise of the technical solution of the present invention, and detailed implementation and specific operation process are provided, but the protection scope of the present invention is not limited to the following implementation example.

本实施例根据纹理特征使用FCM算法对区域进行聚类，并用聚类结果对区域赋予标签。然后，提出语义一致的多特征距离模型。距离模型综合使用光谱、纹理和形状等信息度量区域间的同质性距离，并加入区域间纹理聚类距离使区域合并可以按照纹理同质方向进行。最后，将多特征距离模型应用到RAG和NNG图模型中，通过快速全局最优区域合并得到分割结果。在合并过程中，纹理聚类和最优特征合并之间的相互作用保证了纹理区域边界的准确性。In this embodiment, the FCM algorithm is used to cluster the regions according to the texture features, and the clustering results are used to assign labels to the regions. Then, a semantically consistent multi-feature distance model is proposed. The distance model comprehensively uses the information of spectrum, texture and shape to measure the homogeneity distance between regions, and adds the texture clustering distance between regions so that the region can be merged according to the direction of texture homogeneity. Finally, the multi-feature distance model is applied to the RAG and NNG graph models, and the segmentation results are obtained by fast global optimal region merging. During the merging process, the interplay between texture clustering and optimal feature merging guarantees the accuracy of texture region boundaries.

1、获取区域Gabor能量纹理特征，具体实现如下：1. Obtain regional Gabor energy texture features, the specific implementation is as follows:

一个2维的Gabor函数可以写为如下形式：A 2-dimensional Gabor function can be written as follows:

x′＝xcosθ+ysinθx'=xcosθ+ysinθ

y′＝-xsinθ+ycosθy'=-xsinθ+ycosθ

其中λ为波长，1/λ可以用来表示频域内目标的高频和低频中心；θ为Gabor的条带方向；r为长宽比，决定了Gabor的椭球率；δ/λ为Gabor滤波器的频域带宽。

是Gabor滤波器的相位参数，当相位为0和π时，滤波器为中心对称；当相位为1/2π和-1/2π时，滤波器为反对称函数。Where λ is the wavelength, and 1/λ can be used to represent the high-frequency and low-frequency center of the target in the frequency domain; θ is the strip direction of Gabor; r is the aspect ratio, which determines the ellipsoidal rate of Gabor; δ/λ is Gabor filtering The frequency domain bandwidth of the device.

is the phase parameter of the Gabor filter. When the phase is 0 and π, the filter is centrosymmetric; when the phase is 1/2π and -1/2π, the filter is an antisymmetric function.

Gabor能量函数可以定义为如下形式：The Gabor energy function can be defined as follows:

${\mathrm{<span><span>e</span>e</span>}}_{\mathrm{<span><span>\&lambda;</span>\&lambda;</span>}<span><span>,</span>,</span>\mathrm{<span><span>\&theta;</span>\&theta;</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\sqrt{{{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>2</span>2</span>}}}_{\mathrm{<span><span>\&lambda;</span>\&lambda;</span>}<span><span>,</span>,</span>\mathrm{<span><span>\&theta;</span>\&theta;</span>}<span><span>,</span>,</span>\mathrm{<span><span>0</span>0</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span><span><span>+</span>+</span>{{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>2</span>2</span>}}}_{\mathrm{<span><span>\&lambda;</span>\&lambda;</span>}<span><span>,</span>,</span>\mathrm{<span><span>\&theta;</span>\&theta;</span>}<span><span>,</span>,</span>}<span><span>-</span>-</span>\frac{\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>2</span>2</span>}}\mathrm{<span><span>\&pi;</span>\&pi;</span>}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span>}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>2</span>2</span>}<span><span>)</span>)</span>$

r_λ，θ，0(x，y)和

分别是使用

和

对图像卷积后的结果。r_{λ, θ, 0} (x, y) and

are used respectively

and

The result of convolving an image.

2、建立语义一致的距离空间模型，具体实现为：2. Establish a semantically consistent distance space model, specifically implemented as:

使用光谱、纹理和形状等信息的距离空间模型为：The distance space model using information such as spectrum, texture and shape is:

d＝d_t·d_c·d_s                (3)d=d_t d_c d_s (3)

其中，d_t为纹理距离，d_c为颜色距离，d_s为形状距离。Among them, d_t is the texture distance, d_c is the color distance, and d_s is the shape distance.

纹理距离d_t定义为：The texture distance_dt is defined as:

d_t＝d_tr+d_tc                 (4)d_t =d_tr +d_tc (4)

其中d_tr为区域间Gabor纹理距离，其定义为：where d_tr is the Gabor texture distance between regions, which is defined as:

${\mathrm{<span><span>d</span>d</span>}}_{\mathrm{<span><span>tr</span>tr</span>}}<span><span>=</span>=</span>\underset{\mathrm{<span><span>b</span>b</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\sqrt{\underset{\mathrm{<span><span>\&lambda;</span>\&lambda;</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\underset{\mathrm{<span><span>\&theta;</span>\&theta;</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}{<span><span>(</span>(</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>1</span>1</span>}}<span><span>-</span>-</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>2</span>2</span>}}<span><span>)</span>)</span>}^{\mathrm{<span><span>2</span>2</span>}}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>5</span>5</span>}<span><span>)</span>)</span>$

b为波段数；u₁和u₂分别为区域1和区域2的Gabor平均能量。b is the number of bands; u₁ and u₂ are the Gabor average energy in region 1 and region 2, respectively.

d_tc为区域间纹理聚类距离，其定义为：d_tc is the texture clustering distance between regions, which is defined as:

${\mathrm{<span><span>d</span>d</span>}}_{\mathrm{<span><span>tc</span>tc</span>}}<span><span>=</span>=</span>\left\{\begin{array}{c}{\mathrm{<span><span>d</span>d</span>}}_{\mathrm{<span><span>ij</span>ij</span>}}<span><span>=</span>=</span>\underset{\mathrm{<span><span>b</span>b</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\sqrt{\underset{\mathrm{<span><span>\&lambda;</span>\&lambda;</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\underset{\mathrm{<span><span>\&theta;</span>\&theta;</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}{<span><span>(</span>(</span>{\mathrm{<span><span>e</span>e</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>-</span>-</span>{\mathrm{<span><span>e</span>e</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>)</span>)</span>}^{\mathrm{<span><span>2</span>2</span>}}}<span><span>,</span>,</span>\mathrm{<span><span>i</span>i</span>}<span><span>\&NotEqual;</span>\&NotEqual;</span>\mathrm{<span><span>j</span>j</span>}\\ {\mathrm{<span><span>d</span>d</span>}}_{\mathrm{<span><span>ii</span>i</span>}}<span><span>=</span>=</span>\mathrm{<span><span>min</span>min</span>}<span><span>(</span>(</span>{\mathrm{<span><span>d</span>d</span>}}_{\mathrm{<span><span>ij</span>ij</span>}}<span><span>)</span>)</span><span><span>/</span>/</span>\mathrm{<span><span>2</span>2</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}<span><span>,</span>,</span>\mathrm{<span><span>i</span>i</span>}<span><span>\&NotEqual;</span>\&NotEqual;</span>\mathrm{<span><span>j</span>j</span>}\end{array}\right.<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>6</span>6</span>}<span><span>)</span>)</span>$

e_i和e_j分别为区域1和区域2所属聚类类别i和j的中心Gabor能量。e_i and e_j are the central Gabor energies of the cluster categories i and j to which region 1 and region 2 belong, respectively.

颜色距离d_c的定义：Definition of color distance d_c :

${\mathrm{<span><span>d</span>d</span>}}_{\mathrm{<span><span>c</span>c</span>}}<span><span>=</span>=</span>\underset{\mathrm{<span><span>b</span>b</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>(</span>(</span>{\mathrm{<span><span>\&sigma;</span>\&sigma;</span>}}_{\mathrm{<span><span>m</span>m</span>}}<span><span>-</span>-</span><span><span>(</span>(</span>{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>1</span>1</span>}}<span><span>\&CenterDot;</span>\&Center\; Dot;</span>{\mathrm{<span><span>\&sigma;</span>\&sigma;</span>}}_{\mathrm{<span><span>1</span>1</span>}}<span><span>+</span>+</span>{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>2</span>2</span>}}<span><span>\&CenterDot;</span>\&Center\; Dot;</span>{\mathrm{<span><span>\&sigma;</span>\&sigma;</span>}}_{\mathrm{<span><span>2</span>2</span>}}<span><span>)</span>)</span><span><span>/</span>/</span>{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>m</span>m</span>}}<span><span>)</span>)</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>7</span>7</span>}<span><span>)</span>)</span>$

b为波段数；w_i为图像每个波段的权重；n_m是区域1和区域2合并后的像素数，n₁，n₂分别为区域1和区域2的像素数；σ为对象在i波段下的均方差，其下标分别表示合并后区域、区域1和区域2。b is the number of bands; w_i is the weight of each band of the image; n_m is the number of pixels after the combination of area 1 and area 2, n₁ and n₂ are the number of pixels in area 1 and area 2 respectively; σ is the object at i The mean square error under the bands, whose subscripts denote the merged area, area 1, and area 2, respectively.

形状距离d_s的定义为：The shape distance d_s is defined as:

${\mathrm{<span><span>d</span>d</span>}}_{\mathrm{<span><span>s</span>the\; s</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>l</span>l</span>}}_{\mathrm{<span><span>m</span>m</span>}}<span><span>\&CenterDot;</span>\&Center\; Dot;</span>{\mathrm{<span><span>l</span>l</span>}}_{\mathrm{<span><span>m</span>m</span>}}}{{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>c</span>c</span>}}}<span><span>\&CenterDot;</span>\&Center\; Dot;</span>\mathrm{<span><span>exp</span>exp</span>}<span><span>(</span>(</span><span><span>-</span>-</span>\frac{\mathrm{<span><span>5</span>5</span>}}{{\mathrm{<span><span>d</span>d</span>}}_{\mathrm{<span><span>c</span>c</span>}}}<span><span>)</span>)</span><span><span>\&CenterDot;</span>\&Center\; Dot;</span>{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>m</span>m</span>}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>8</span>8</span>}<span><span>)</span>)</span>$

其中l_m是对像的边长，n_c是区域间相邻的像素数，d_c为区域间光谱距离，n_m为合并后的区域像素数。Among them, l_m is the side length of the object, n_c is the number of adjacent pixels between regions, d_c is the spectral distance between regions, and n_m is the number of pixels in the merged region.

3、建立RAG和NNG图模型，具体实现为：3. Establish RAG and NNG graph models, the specific implementation is:

RAG表示区域结点的邻接关系，其描述如下：一幅K个区域的图像可以表示为简单图G＝(V，E)，V为图像的区域，在图结构中为图的顶点集V＝{1，2…，K}；E为图像的区域相邻连接关系，在图结构中为边集$E=\{e=\mathrm{Edge}(v_i,v_j)|i,j\&Element;(\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;k),i\&NotEqual;j\},E\&Subset;V\&times;V$。当区域相邻时，区域结点间拥有边e，e表示区域间的合并代价，可以通过公式(3)得出；当区域不相邻时，区域顶点之间没有边存在，即合并代价为无穷大。RAG represents the adjacency relationship of regional nodes, and its description is as follows: an image of K regions can be expressed as a simple graph G=(V, E), V is the region of the image, and in the graph structure, it is the vertex set of the graph V= {1, 2..., K}; E is the adjacent connection relationship of the image area, which is an edge set in the graph structure$\mathrm{E.}=\{e=\mathrm{Edge}(v_i,v_j)|i,j\&Element;(\mathrm{1,2}\&CenterDot;\&Center\; Dot;\&Center\; Dot;k),i\&NotEqual;j\},\mathrm{E.}\&Subset;V\&times;V$ . When the regions are adjacent, there is an edge e between the region nodes, and e represents the merge cost between regions, which can be obtained by formula (3); when the regions are not adjacent, there is no edge between the vertices of the region, that is, the merge cost is gigantic.

NNG记录区域的合并顺序，其描述为：设一幅K个区域的图像，其RAG图为G＝(V，E)，则G有子图G_m＝(V_m，E_m)，其中V_m＝V；E_m＝{e_i＝mine_i，j|i，j∈(1，2…k)}，即子图的边集E_m为和顶点v_i相关联的最小权值边e_i，j的集合。对于无向图G_m，令其边集E_m中每个元素增加一个方向得到弧集A_m＝{a_i＝Arc(v_i，v_j)|e_i，j∈E_m}，则无向图G_m生成其定向图G_m＝(V_m，A_m)，即NNG图。每个顶点v_i的入度不一定为1，但出度都为1，并且以v_i为尾的弧的头即是其最同质的区域。当A_m中存在环子集A_c＝{a_i，a_j|a_i∈A_m，a_j∈A_m}时，v_i与v_j互为最优，也是局部最优区域对。The merging order of the NNG recording area is described as: set an image of K areas, its RAG map is G=(V, E), then G has sub-graph G_m =(V_m , E_m ), where V_m = V; E_m = {e_i = mine_{i, j} |i, j∈(1, 2...k)}, that is, the edge set E_m of the subgraph is the minimum weight edge e associated with the vertex v_{ii, the collection of j} . For an undirected graph G_m , add a direction to each element in its edge set E_m to obtain an arc set A_m = {a_i =Arc(v_i , v_j )|e_{i, j} ∈ E_m }, then there is no Generate its directed graph G_m = (V_m , A_m ) to the graph G_m , that is, the NNG graph. The in-degree of each vertex v_i is not necessarily 1, but the out-degree is 1, and the head of the arc ending with v_i is its most homogeneous area. When there is a ring subset A_c = {a_i , a_j |a_i ∈ A_m , a_j ∈ A_m } in_A m , v_i and v_j are mutually optimal, and they are also a local optimal region pair.

算法只搜索环集合来决定全局合并序列。对于一幅图像，NNG最少有一个环子集，最多有

个环子集，所以NNG大大降低了搜索的时间。使用一个辅助序列记录环子集节点的位置。对序列按区域间合并代价从小到大排序得到的就是全局最优合并序列。The algorithm searches only the set of rings to determine the global merge sequence. For an image, NNG has at least one ring subset and at most

ring subset, so NNG greatly reduces the search time. Use an auxiliary sequence to record the positions of the ring subset nodes. Sorting the sequences according to the inter-regional merging cost from small to large is the globally optimal merging sequence.

4、区域合并得到分割结果，具体实现为：4. Regions are merged to obtain segmentation results, which are specifically implemented as follows:

按照全局最优合并序列进行合并，每次合并后使用平均Gabor能量纹理作为新区域的Gabor能量，并根据公式(5)计算新区域到各个纹理聚类中心间的距离，将距离最小的纹理聚类类别赋予新区域。然后根据公式(1)重新计算新区域和邻接区域之间的合并距离，更新RAG和NNG，寻找新区域相应环子集，并更新最优合并序列，继续合并。当全局最优区域对的合并距离大于阈值时，停止合并得到最终分割结果。According to the global optimal merging sequence, the average Gabor energy texture is used as the Gabor energy of the new area after each merging, and the distance between the new area and each texture cluster center is calculated according to formula (5), and the texture with the smallest distance is clustered. Class categories are assigned to new regions. Then recalculate the merging distance between the new region and adjacent regions according to formula (1), update RAG and NNG, find the corresponding ring subset of the new region, update the optimal merging sequence, and continue merging. When the merging distance of the globally optimal region pair is greater than the threshold, the merging is stopped to obtain the final segmentation result.

在合并过程中，合并后区域的聚类类别发生了改变，同时，合并后的区域聚类类别又影响着最优序列的合并顺序。通过纹理聚类和最优序列合并的相互作用，纹理区域的边界可以准确得到。During the merging process, the clustering category of the merged region has changed, and at the same time, the clustering category of the merged region affects the merging order of the optimal sequence. Through the interaction of texture clustering and optimal sequence merging, the boundaries of texture regions can be accurately obtained.

以下进一步说明本实施例应用情况：The application of this embodiment is further described below:

使用了两幅图像数据测试分割算法的准确性，一幅来自于上海市的2.5mSPOT5真彩色遥感影像，一幅为上海市崇明岛25cm的航空影像。SPOT5图像是2005年9月30日获取，截取大小为512×512个像元的数据用做试验。所取样区中主要包括了5类地物：居民地、工业用地、水域、耕地和鱼塘。其中，居民地光谱不均匀，但明显与其它地物不同；工业用地光谱不均匀，但呈现条状纹理；耕地间光谱有差异，但纹理较相似；水域和鱼塘光谱同质，纹理细密。航空图像截取大小为512×512的数据进行试验。在航空图像中，主要包括了林地、河流、居民地、道路、大棚菜地、灌溉水田和旱地等土地类别。其中，林地、大棚菜地、灌溉水田和旱地等表现出明显不同的纹理特征。Two image data were used to test the accuracy of the segmentation algorithm, one was a 2.5mSPOT5 true-color remote sensing image from Shanghai, and the other was a 25cm aerial image of Chongming Island in Shanghai. The SPOT5 image was acquired on September 30, 2005, and the data with a size of 512×512 pixels was intercepted for experimentation. The sampling area mainly includes five types of land features: residential areas, industrial land, water areas, cultivated land and fish ponds. Among them, the spectrum of residential land is not uniform, but it is obviously different from other land objects; the spectrum of industrial land is not uniform, but it has a strip texture; Aerial image interception data with a size of 512×512 is used for experimentation. In aerial images, land types such as forest land, rivers, residential areas, roads, greenhouse vegetable fields, irrigated paddy fields, and dry land are mainly included. Among them, forest land, greenhouse vegetable field, irrigated paddy field and dry land show obvious different texture characteristics.

以人工分割的结果作为参考结果，对比两种分割方法的准确性，一种为基于光谱同质性的Definiens分割方法，另一种为本实施例方法。试验结果如下：(1)SPOT5图像分割准确性比较(见表1)；(2)航空图像分割准确性比较(见表2)。Taking the result of manual segmentation as a reference result, the accuracy of two segmentation methods is compared, one is the Definiens segmentation method based on spectral homogeneity, and the other is the method of this embodiment. The test results are as follows: (1) Comparison of SPOT5 image segmentation accuracy (see Table 1); (2) Comparison of aerial image segmentation accuracy (see Table 2).

从四个方面对变化检测结果进行评价：(1)主要对象的分割精度：占图像主要部分的地物，如耕地、林地等；(2)微小对象的分割精度：较小面积的地物但在图像解译时不可缺少，如孤立的房子、特殊标志等；(3)线性对象的分割精度：河流、道路等地物；(4)总体分割精度。Evaluate the change detection results from four aspects: (1) Segmentation accuracy of main objects: ground objects that occupy the main part of the image, such as cultivated land, woodland, etc.; (2) Segmentation accuracy of tiny objects: Smaller area ground objects but Indispensable in image interpretation, such as isolated houses, special signs, etc.; (3) Segmentation accuracy of linear objects: rivers, roads and other ground objects; (4) Overall segmentation accuracy.

表1 SPOT5图像分割结果准确性比较Table 1 Accuracy comparison of SPOT5 image segmentation results

表2 航空图像分割结果准确性比较Table 2 Accuracy comparison of aerial image segmentation results

从表中可以看到，对于图像中的主要对象，本实施例取得了68.2％和87.5％的分割精度，远高于Definiens软件54.5％和62.5％的分割精度。这主要是由于大面积对象纹理清晰，本实施例通过纹理聚类限制更容易得到准确地纹理区域，而Definiens方法则较难通过光谱得到纹理区域的正确分割结果。而对于微小地物和线状地物等，本实施例也取得了比Definiens高的分割精度。从总体精度上看，本实施例相对于Definiens软件提高了10％-20％的分割精度。It can be seen from the table that for the main object in the image, this embodiment achieves segmentation accuracies of 68.2% and 87.5%, which are much higher than the segmentation accuracies of 54.5% and 62.5% of the Definiens software. This is mainly due to the clear texture of large-area objects. In this embodiment, it is easier to obtain accurate texture regions through texture clustering restrictions, while it is difficult for the Definiens method to obtain correct segmentation results of texture regions through spectra. For tiny ground objects and linear ground objects, etc., this embodiment also achieves higher segmentation accuracy than Definiens. In terms of overall accuracy, this embodiment improves the segmentation accuracy by 10%-20% compared with the Definiens software.

在视觉上，本实施例对于具有规则纹理的地物，如工厂厂房、林地等，取得很好的分割效果。这些地物分割后结构完整，边界准确，远远好于Definiens的分割结果。Visually, this embodiment achieves a good segmentation effect on ground objects with regular textures, such as factory buildings and forest land. After segmentation, these features have complete structures and accurate boundaries, which are far better than the segmentation results of Definiens.

总的来说，本实施例对于高分辨率遥感图像可以取得比传统方法更高的精度，特别是对于图像中的规则纹理区域。Generally speaking, this embodiment can achieve higher accuracy than traditional methods for high-resolution remote sensing images, especially for regular texture regions in images.

Claims (10)

1, a kind of method for segmentation of high resolution remote sensing image based on veins clustering constrain is characterized in that, comprises the steps:

The first step, obtain zone-texture cluster label: the Gabor energy textures of All Ranges in the computed image, and use FCM that All Ranges is carried out cluster according to texture, set veins clustering label for each zone according to cluster result;

Second goes on foot, and sets up the metric space model of semantic congruence: use features such as spectrum, texture and shape to set up the comprehensive distance spatial model, and add veins clustering apart from the zone is merged into row constraint;

In the 3rd step, use the graph model algorithm that the zone is merged fast: set up RAG and NNG according to comprehensive distance, and according to global optimum to the zone to merging, obtain net result.

2, the method for segmentation of high resolution remote sensing image based on veins clustering constrain according to claim 1, it is characterized in that, set the veins clustering label, be specially: use 3 yardsticks, 8 direction Gabor filtering that image is carried out convolution, get the count root of quadratic sum of corresponding symmetry and antisymmetry Gabor filtering result and obtain the Gabor energy, the Gabor energy is 24 dimensional vectors, uses this vectorial Euclidean distance to carry out cluster and obtains zone-texture cluster label.

3, the method for segmentation of high resolution remote sensing image based on veins clustering constrain according to claim 1 is characterized in that, described constraint is specially: use the veins clustering label that regional merge order is retrained.

4, the method for segmentation of high resolution remote sensing image based on veins clustering constrain according to claim 1, it is characterized in that, described comprehensive distance spatial model, be specially: comprehensively use spectrum, texture and shape information, and add the veins clustering distance, set up the metric space model that uses many features.

5, the method for segmentation of high resolution remote sensing image based on veins clustering constrain according to claim 1, it is characterized in that, described the zone is merged fast, be specially: use RAG figure to describe syntople between the zone, and the combined distance between posting field; Use NNG figure to describe the corresponding optimum in each zone and merge the zone, and write down optimum combined distance; In NNG figure, the optimum each other zone that merges the zone obtains global optimum collating sequence with all local optimum zones to ordering to being the local optimum in the image, merges according to this sequence to obtain segmentation result.

6, the method for segmentation of high resolution remote sensing image based on veins clustering constrain according to claim 1 is characterized in that, described merging is meant the real border that obtains the zone in merging process by the interaction of veins clustering and optimum collating sequence.

7, the method for segmentation of high resolution remote sensing image based on veins clustering constrain according to claim 1, it is characterized in that, described global optimum is meant: cut apart the order that time domain merges by veins clustering pre-service influence, the zone is merged according to the direction of texture homogeneity.

8, the method for segmentation of high resolution remote sensing image based on veins clustering constrain according to claim 5, it is characterized in that, the described optimum zone that merges, be meant: in RAG figure, when the combined distance of regional A and certain adjacent area B during, then claim adjacent area B to merge the zone for the optimum in this zone less than the combined distance of regional A and other adjacent area.

9, the method for segmentation of high resolution remote sensing image based on veins clustering constrain according to claim 5 is characterized in that, described optimum combined distance is meant: in NNG figure, the optimum each other zone that merges the zone between combined distance.

10, according to claim 5 or 6 described method for segmentation of high resolution remote sensing image based on veins clustering constrain, it is characterized in that, described optimum collating sequence, be meant: in NNG figure, the optimum each other zone that merges the zone is regional right to be that local optimum in the image merges, the collating sequence that all local optimum zones are obtained ordering.