CN101976347A

CN101976347A - Method for recognizing overwater bridge in remote sensing image on basis of Mean Shift segmentation

Info

Publication number: CN101976347A
Application number: CN 201010517146
Authority: CN
Inventors: 张艳宁; 李映; 魏巍; 赵静; 马瑜; 孙瑾秋; 郭哲
Original assignee: Northwestern Polytechnical University
Current assignee: Northwestern Polytechnical University
Priority date: 2010-10-21
Filing date: 2010-10-21
Publication date: 2011-02-16

Abstract

本发明公开了一种基于Mean Shift分割的遥感图像中水上桥梁识别方法，用于解决现有的水上桥梁目标识别方法识别率低的技术问题。技术方案是采用Mean Shift方法，利用颜色特性进行图像分割，提取河流区域，并利用相似度确定河流域。结合膨胀、腐蚀连通河流区域，根据桥梁的特征，提取桥梁区域。通过细化得到连通的河流的骨架线，寻找与桥梁区域的交点来找到候选桥梁，最后利用桥梁的形状纹理特征识别桥梁，提高了水上桥梁目标识别的识别率。The invention discloses a water bridge recognition method in a remote sensing image based on Mean Shift segmentation, which is used to solve the technical problem of low recognition rate of the existing water bridge target recognition method. The technical solution is to use the Mean Shift method to segment the image using color features, extract the river area, and use the similarity to determine the river basin. Combined with expansion and erosion connected river areas, the bridge area is extracted according to the characteristics of the bridge. By thinning the skeleton line of the connected river, finding the intersection point with the bridge area to find the candidate bridge, and finally using the shape and texture features of the bridge to identify the bridge, which improves the recognition rate of the water bridge target recognition.

Description

Water Bridge Recognition Method in Remote Sensing Image Based on Mean Shift Segmentation

技术领域technical field

本发明涉及一种遥感图像识别方法，特别是一种基于Mean Shift分割的遥感图像中水上桥梁识别方法。 The invention relates to a remote sensing image recognition method, in particular to a method for recognizing bridges over water in remote sensing images based on Mean Shift segmentation. the

背景技术Background technique

文献“基于知识的航空图像中大型水上桥梁目标识别，武汉理工大学学报，2005，Vol.29(2)，p230-233”公开了一种航空图像中大型水上桥梁目标识别方法，该方法针对高空航拍图像提出了一种基于知识的桥梁目标识别方法。根据水域与陆地域、桥梁域之间的强对比关系，对航空图像进行二值分割，将水域与其它区域分割开来，根据水域对桥梁进行初步定位。然后使用种子点生长法来精确地标记出桥梁域，认为桥梁的宽度相对于长度而言很小，所以经HOUGH变换处理得到的轴向直线，根据轴线上两边的像素点得到桥梁的宽度。但是，因为桥梁在图像中所占的比例比较小，背景复杂，灰度反差比较小，很难在数据驱动下通过图像分割，提取目标特征进而来判断识别目标。而且认为桥梁的首要特征是存在两条平行的长直线，但实际拍摄中，由于传感器的视角、取像距离的原因并非存在平行的长直线。所以文献所述方法处理遥感图像时具有局限性，识别率较低。 The document "Knowledge-based Target Recognition of Large Water Bridges in Aerial Images, Journal of Wuhan University of Technology, 2005, Vol.29(2), p230-233" discloses a method for target recognition of large water bridges in aerial images. Aerial imagery proposes a knowledge-based method for bridge object recognition. According to the strong contrast relationship between the water area, the land area and the bridge area, binary segmentation is performed on the aerial image, the water area is separated from other areas, and the bridge is initially positioned according to the water area. Then use the seed point growth method to accurately mark the bridge domain. It is considered that the width of the bridge is small compared to the length, so the axial straight line obtained by the HOUGH transformation is obtained according to the pixel points on both sides of the axis to obtain the width of the bridge. However, because the proportion of bridges in the image is relatively small, the background is complex, and the grayscale contrast is relatively small, it is difficult to use data-driven image segmentation to extract target features to judge and identify targets. And it is believed that the primary feature of a bridge is the existence of two parallel long straight lines, but in actual shooting, due to the sensor's viewing angle and imaging distance, there are no parallel long straight lines. Therefore, the method described in the literature has limitations when dealing with remote sensing images, and the recognition rate is low. the

发明内容Contents of the invention

为了克服现有的水上桥梁目标识别方法识别率低的不足，本发明提供一种基于Mean Shift分割的遥感图像中水上桥梁识别方法。该方法采用Mean Shift方法，利用颜色特性进行图像分割，提取河流区域，并利用相似度确定河流域。结合膨胀、腐蚀连通河流区域，根据桥梁的特征，提取桥梁区域。通过细化得到连通的河流的骨架线，寻找与桥梁区域的交点来找到候选桥梁，最后利用桥梁的形状纹理特征识别桥梁，可以提高水上桥梁目标识别的识别率。 In order to overcome the deficiency of low recognition rate of existing bridge target recognition methods on water, the present invention provides a bridge recognition method on water in remote sensing images based on Mean Shift segmentation. This method adopts the Mean Shift method, uses the color feature to segment the image, extracts the river area, and uses the similarity to determine the river basin. Combined with expansion and erosion connected river areas, the bridge area is extracted according to the characteristics of the bridge. By thinning the skeleton lines of the connected rivers, finding the intersection points with the bridge area to find candidate bridges, and finally using the shape and texture features of bridges to identify bridges, the recognition rate of water bridge target recognition can be improved. the

本发明解决其技术问题所采用的技术方案：一种基于Mean Shift分割的遥感图像中水上桥梁识别方法，其特点是包括以下步骤： The technical solution adopted by the present invention to solve its technical problems: a method for identifying bridges over water in remote sensing images based on Mean Shift segmentation, which is characterized in that it comprises the following steps:

(a)将遥感图像格式转化到LUV空间，对遥感图像进行Mean Shift分割和区域合并； (a) Convert the remote sensing image format to LUV space, and perform Mean Shift segmentation and region merging on the remote sensing image;

令区域合并后联合域中原始像素点{x_i}i＝1，2...，n；联合域中滤波后像素点{z_i}i＝1，2...，n；分割后遥感图像中第i个像素标记为L_i，i＝1，2...，n。 Make the original pixel point {xi _} i=1, 2..., n in the joint domain after the region is merged; the filtered pixel point {z _i }i=1, 2..., n in the joint domain; The i-th pixel in the image is marked as L _i , i=1, 2..., n.

利用高斯核函数k(x)估计特征密度空间，对遥感图像中的任一点用高斯核函数k(x)进行漂移： Use the Gaussian kernel function k(x) to estimate the feature density space, and use the Gaussian kernel function k(x) to drift any point in the remote sensing image:

$k k ((x x)) = = {((22 Π Π))}^{- - d d / / 22} exp exp ((- - \frac{11}{22} {| | | | x x | | | |}^{22})) - - - - - - ((11))$

利用k(x)得出具有收敛性的递推公式 Use k(x) to get a convergent recursive formula

${y the y}_{i i + + 11} = = \frac{{Σ Σ}_{i i = = 11}^{n no} {x x}_{i i} g g (({| | | | \frac{{x x - - x x}_{i i}}{h h} | | | |}^{22}))}{{Σ Σ}_{i i = = 11}^{n no} g g (({| | | | \frac{{x x - - x x}_{i i}}{h h} | | | |}^{22}))} - - - - - - ((22))$

和Mean Shift向量 and the Mean Shift vector

${m m}_{h h} ((x x)) = = \frac{{Σ Σ}_{i i = = 11}^{n no} {x x}_{i i} g g (({| | | | \frac{{x x - - x x}_{i i}}{h h} | | | |}^{22}))}{{Σ Σ}_{i i = = 11}^{n no} g g (({| | | | \frac{{x x - - x x}_{i i}}{h h} | | | |}^{22}))} - - x x = = {y the y}_{i i + + 11} - - y the y - - - - - - ((33))$

进行迭代卷积，直到满足停止准则，即移动距离小于设定数||m_h(x)-x ||＜ε或者漂移次数达到最大值；式中，h是带宽参数，h＝(hs，hv)，hs是空域颜色特征带宽，hv是空间带宽，{c_j}是核函数k(x)的剖面函数的负导数； Carry out iterative convolution until the _stop criterion is satisfied, that is, the moving distance is less than the set number || hv), hs is the spatial color feature bandwidth, hv is the spatial bandwidth, {c _j } is the negative derivative of the profile function of the kernel function k(x);

对遥感图像进行Mean Shift滤波并把所有关于收敛点的信息都保存在z_i中； Carry out Mean Shift filtering on the remote sensing image and save all the information about the convergence point in _zi ;

在联合域中生成聚类{c_j}j＝1，...，m，把所有在空域距离小于hv并且在色度域距离小于hs的z_i组合在一起； Generate clusters {c _j } j=1,...,m in the joint domain, combine all z _i whose distance in the space domain is less than hv and the distance in the chromaticity domain is less than hs;

对于任意i＝1，2...，n，令Li＝{j|z_i∈c_j}，设定最小区域M，剔除小于M的空间区域； For any i=1, 2..., n, let Li={j|z _i ∈ c _j }, set the minimum area M, and eliminate the spatial area smaller than M;

(b)计算所有区域之间的相似值，各区域的向量作内积进行相似性判断； (b) Calculate the similarity value between all regions, and make the inner product of the vectors of each region to judge the similarity;

或者，将各区域的三分量转化为灰度值，求各区域局部区域方差，进行相似性判断；并将最小局部区域方差，作为第一块河流； Or, convert the three components of each region into gray values, find the local area variance of each area, and perform similarity judgment; and use the minimum local area variance as the first river;

或者，计算剩余各区域与首块河流区域之间的均方差判断相似性； Or, calculate the mean square error between the remaining areas and the first river area to judge the similarity;

依次选取满足以上相似性标准的区域作为河流区域； Sequentially select the areas that meet the above similarity criteria as river areas;

(c)对遥感图像进行二值化处理，利用连通区域标记法标记潜在河流区域，并去除噪声河流区域；对河流区域执行膨胀操作，连通河流；然后用最大连通水域与原未连通二值化河流区域做差，利用两个河流区域之间是桥梁域的特点，提取桥梁区域，并细化提取河流中心线； (c) Binarize the remote sensing image, use the connected area marking method to mark the potential river area, and remove the noisy river area; perform an expansion operation on the river area to connect the river; then use the maximum connected water area to binarize the original unconnected area The difference between the river areas, using the characteristics of the bridge domain between the two river areas, extracts the bridge area, and refines the extraction of the river centerline;

(d)寻找河流域中心线与未连通河流图像的交点，并且落在桥梁区域的点，提取候选桥梁；提取桥梁轮廓线，利用形状纹理特征来识别桥梁。 (d) Find the intersection of the centerline of the river basin and the unconnected river image, and the point that falls in the bridge area, and extract the candidate bridge; extract the bridge outline, and use the shape and texture features to identify the bridge. the

所述带宽参数h＝(hs，hv)中，空域颜色特征带宽hs最佳值是9，空间带宽hv最佳值是8.5。 In the bandwidth parameter h=(hs, hv), the optimal value of the spatial color feature bandwidth hs is 9, and the optimal value of the spatial bandwidth hv is 8.5. the

所述最小区域M是500。 The minimum area M is 500. the

本发明的有益效果是：由于采用Mean Shift方法，利用颜色特性进行图像分割，提取河流区域，并利用相似度确定河流域。结合膨胀、腐蚀连通河流区域，根据桥梁的特征，提取桥梁区域。通过细化得到连通的河流的骨架线，寻找与桥梁区域的交点来找到候选桥梁，最后利用桥梁的形状纹理特征识别桥梁，提高了水上桥梁目标识别的识别率。 The beneficial effects of the present invention are: because the Mean Shift method is adopted, the image is segmented by using the color characteristics, the river area is extracted, and the river basin is determined by the similarity. Combined with expansion and erosion connected river areas, the bridge area is extracted according to the characteristics of the bridge. By thinning the skeleton line of the connected river, finding the intersection with the bridge area to find the candidate bridge, and finally using the shape and texture features of the bridge to identify the bridge, which improves the recognition rate of the water bridge target recognition. the

下面结合具体实施方式对本发明作详细说明。 The present invention will be described in detail below in combination with specific embodiments. the

具体实施方式Detailed ways

1.首先把输入的彩色的遥感图像格式转化到LUV空间，对遥感图像进行MeanShift分割。 1. First, convert the input color remote sensing image format to LUV space, and perform MeanShift segmentation on the remote sensing image. the

1)进行Mean Shift滤波： 1) Perform Mean Shift filtering:

利用高斯核函数k(x)估计特征密度空间，对遥感图像中的任一点(不重复)用高斯核函数k(x)进行漂移。 Use the Gaussian kernel function k(x) to estimate the feature density space, and use the Gaussian kernel function k(x) to drift any point in the remote sensing image (not repeated). the

然后利用k(x)得出具有收敛性的递推公式(2)和Mean Shift向量(3)，进行迭代卷积，直到满足停止准则，移动距离小于设定数||m_h(x)-x||＜ε或者漂移次数达到最大值，Mean Shift向量是归一化的概率密度梯度，不断地沿着概率密度的梯度方向移动，总是指向概率密度增加的最大方向。 Then use k(x) to get the convergent recursive formula (2) and Mean Shift vector (3), and perform iterative convolution until the stopping criterion is met, and the moving distance is less than the set number||m _h (x)- x||<ε or the number of drifts reaches the maximum value, and the Mean Shift vector is the normalized probability density gradient, constantly moving along the gradient direction of the probability density, always pointing to the maximum direction of probability density increase.

式中，h是带宽参数，h＝(hs，hv)，hs是空域颜色特征带宽，hv是空间带宽，应用中选用经验值。{c_j}是核函数k(x)的剖面函数的负导数。 In the formula, h is the bandwidth parameter, h=(hs, hv), hs is the color characteristic bandwidth of the space domain, hv is the space bandwidth, and the empirical value is used in the application. {c _j } is the negative derivative of the profile function of the kernel function k(x).

Mean Shift分割算法融合了图像空间域与颜色域信息，带宽参数决定着滤波时的分辨率。这两个参数对于Mean Shift滤波效果具有重要的影响。实验发现hs设置9，hv设置为8.5比较合适。 The Mean Shift segmentation algorithm combines image space domain and color domain information, and the bandwidth parameter determines the resolution of filtering. These two parameters have an important impact on the Mean Shift filtering effect. The experiment found that setting hs to 9 and hv to 8.5 is more appropriate. the

2)区域合并： 2) Regional merger:

实现对遥感图像的区域合并，把相邻并且像素值差别小的区域合并起来，得到一个较大的区域。具体步骤为：令{x_i}i＝1，2...，n和{z_i}i＝1，2...，n分别表示联合域中的原始和滤波后像素点，令分割后图像中第i个像素的标签用L_i，i＝1，2...，n表示。 Realize the region merging of remote sensing images, and merge adjacent regions with small pixel value differences to obtain a larger region. The specific steps are: let {xi _} i=1, 2..., n and {zi _} i=1, 2..., n denote the original and filtered pixels in the joint domain respectively, let the divided The label of the i-th pixel in the image is denoted by L _i , i=1, 2..., n.

a)对遥感图像进行Mean Shift滤波并把所有关于收敛点的信息都保存在z_i中； a) Carry out Mean Shift filtering on the remote sensing image and save all the information about the convergence point in _zi ;

b)在联合域中生成聚类{c_j}j＝1，...，m，把所有在空域距离小于hv并且在色度域距离小于hs的z_i组合在一起； b) Generate clusters {c _j }j=1,...,m in the joint domain, combining all z _i whose distance in the space domain is less than hv and distance in the chromaticity domain is less than hs;

c)对于任意i＝1，2...，n，令Li＝{j|z_i∈c_j}，设定最小区域M，本实施例取M＝500，剔除那些小于M空间区域。 c) For any i=1, 2...,n, set Li={j|z _i ∈c _j }, set the minimum area M, and take M=500 in this embodiment, and remove those space areas smaller than M.

2.将LUV三分量再转回到RGB空间。对Mean Shift各区域按大小进行排序，删除小区域。对所有区域进行相似性判断，提取河流区域： 2. Convert the three components of LUV back to RGB space. Sort each area of Mean Shift by size and delete small areas. Perform similarity judgment on all regions and extract river regions:

1)计算所有区域之间的相似值，各区域的向量作内积进行相似性判断。 1) Calculate the similarity value between all regions, and make the inner product of the vectors of each region to judge the similarity. the

2)将各区域的三分量转化为灰度值，求各区域局部区域方差，进行相似性判断。并将最小局部区域方差，作为第一块河流。 2) Convert the three components of each region into gray values, calculate the local variance of each region, and perform similarity judgment. And take the minimum local area variance as the first river. the

3)计算剩余各区域与首河流区域之间的均方差判断相似性。 3) Calculate the mean square error between the remaining regions and the first river region to judge the similarity. the

最终采用的是上述三种方法的综合进行相似性判断，判断准则为经验值。选取满足以上相似性标准的区域作为河流区域。 Finally, a combination of the above three methods is used to judge the similarity, and the judgment criterion is the empirical value. Select the area that meets the above similarity criteria as the river area. the

3.对图像进行二值化处理，利用连通区域标记法标记潜在河流区域，并去除噪声河流区域。对河流区域执行膨胀操作，连通河流。然后用最大连通水域与原未连通二值化河流区域做差，利用两个河流区域之间是桥梁域的特点，提取桥梁区域。并细化提取河流中心线。 3. Binarize the image, use the connected region labeling method to mark potential river regions, and remove noisy river regions. Perform dilation on the river region to connect the river. Then, the difference between the maximum connected water area and the original unconnected binarized river area is used, and the bridge area is extracted by using the characteristics of the bridge area between the two river areas. And refine and extract the centerline of the river. the

4.寻找河流域中线与原未连通二值化河流区域相交并且落在备候选区域的交点，得到桥梁候选区域。根据河流区域的总数确定桥梁个数，并提取桥梁轮廓线，利用利用形状纹理特征来识别桥梁。 4. Find the intersection point where the center line of the river basin intersects the original unconnected binarized river area and falls in the candidate area to obtain the bridge candidate area. Determine the number of bridges according to the total number of river areas, extract the outline of the bridge, and use the shape and texture features to identify the bridge. the

Claims

1. on-water bridge recognition methods in the remote sensing images of cutting apart based on Mean Shift is characterized in that comprising and has write step:

(a) the remote sensing images format conversion is arrived the LUV space, remote sensing images are carried out Mean Shift cut apart and the zone merging;

The order zone merges original pixel { x in the associating territory, back _iI=1,2..., n; Pixel { z after the filtering in the associating territory _iI=1,2..., n; Cut apart the back remote sensing images in i pixel be labeled as L _i, i=1,2..., n.

Utilize gaussian kernel function k (x) to estimate the characteristic density space, any point in the remote sensing images drifted about with gaussian kernel function k (x):

k (x) = {(2 Π)}^{- d / 2} \exp (- \frac{1}{2} {| | x | |}^{2}) - - - (1)

Utilize k (x) to draw and have constringent recursion formula

y_{i + 1} = \frac{Σ_{i = 1}^{n} x_{i} g ({| | \frac{{x - x}_{i}}{h} | |}^{2})}{Σ_{i = 1}^{n} g ({| | \frac{{x - x}_{i}}{h} | |}^{2})} - - - (2)

With Mean Shift vector

m_{h} (x) = \frac{Σ_{i = 1}^{n} x_{i} g ({| | \frac{{x - x}_{i}}{h} | |}^{2})}{Σ_{i = 1}^{n} g ({| | \frac{{x - x}_{i}}{h} | |}^{2})} - x = y_{i + 1} - y - - - (3)

Carry out the iteration convolution, up to satisfying stopping criterion, promptly displacement is less than setting number || m _h(x)-x||＜ε or the drift number of times reach maximal value; In the formula, h is a bandwidth parameter, and h=(hs, hv), hs is a spatial domain color characteristic bandwidth, hv is a spatial bandwidth, { c _jIt is the negative derivative of the section function of kernel function k (x);

Remote sensing images are carried out Mean Shift filtering and the information all about convergence point all is kept at z _iIn;

In the associating territory, generate cluster { c _iJ=1 ..., m, all in spatial domain distance less than hv and at the z of colourity territory distance less than hs _iCombine;

For any i=1,2..., n makes Li={j|z _i∈ c _j, set Minimum Area M, reject area of space less than M;

(b) similar value between the calculating All Ranges, the vector in each zone are carried out similarity as inner product and are judged;

Perhaps, the three-component that each is regional is converted into gray-scale value, asks each regional regional area variance, carries out similarity and judges; And with minimum regional area variance, as first plot of river;

Perhaps, the mean square deviation of calculating between each zone of residue and the first river region is judged similarity;

Choose satisfy above similarity standard the zone as river region;

(c) remote sensing images are carried out binary conversion treatment, utilize the potential river region of connected component labeling method mark, and remove the noise river region; River region is carried out expansive working, be communicated with the river; Do poorly with largest connected waters with the former binaryzation river region that is not communicated with then, utilizing between two river region is the characteristics in bridge territory, extracts the bridge zone, and river axis is extracted in refinement;

(d) seek river valley center line and the intersection point that is not communicated with the river image, and drop on the point in bridge zone, extract candidate's bridge; Extract the bridge outline line, utilize the shape textural characteristics to discern bridge.

2. method according to claim 1 is characterized in that: described bandwidth parameter h=(hs, hv) in, spatial domain color characteristic bandwidth hs optimum value is 9, spatial bandwidth hv optimum value is 8.5.

3. method according to claim 1 is characterized in that: described Minimum Area M is 500.