CN103294792B - Based on the semantic information and the Decomposition Classification Based polarization sar - Google Patents

Based on the semantic information and the Decomposition Classification Based polarization sar Download PDF

Info

Publication number
CN103294792B
CN103294792B CN201310192057.1A CN201310192057A CN103294792B CN 103294792 B CN103294792 B CN 103294792B CN 201310192057 A CN201310192057 A CN 201310192057A CN 103294792 B CN103294792 B CN 103294792B
Authority
CN
China
Prior art keywords
segment
region
based
line
classification
Prior art date
Application number
CN201310192057.1A
Other languages
Chinese (zh)
Other versions
CN103294792A (en
Inventor
刘芳
石俊飞
李玲玲
焦李成
戚玉涛
郝红侠
武杰
张向荣
马晶晶
尚荣华
于昕
Original Assignee
西安电子科技大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 西安电子科技大学 filed Critical 西安电子科技大学
Priority to CN201310192057.1A priority Critical patent/CN103294792B/en
Publication of CN103294792A publication Critical patent/CN103294792A/en
Application granted granted Critical
Publication of CN103294792B publication Critical patent/CN103294792B/en

Links

Abstract

本发明公开了一种基于语义信息和极化分解的极化SAR地物分类方法。 The present invention discloses a polarimetric SAR feature classification method based on the semantic information and polarization decomposition. 其实现包括:对span图进行均值漂移,提取span图的边脊草图,并在边脊草图中用基于语义信息的区域提取技术提取线段聚集区域;基于线段聚集区域并采用临界区域众数投票合并策略和基于极化特征合并策略对span图均值漂移过分割区域进行合并,得到图像分割结果;融合基于语义信息的图像分割结果和基于MRF的H/α-Wishart分类结果,得到最终分类结果。 Which implement comprising: a span of mean shift diagrams, sketches span FIG ridge edge extraction, and the extraction area based on region extracting line segment aggregates semantic information in the sketch by side ridge; accumulation region based on the line number and uses all voting combined critical region the combined policies and strategies based on polarization characteristics mean shift through the FIG span divided regions are merged to obtain an image segmentation result; image segmentation based on the result of the semantic information and the MRF H / α-Wishart classification results to obtain a final classification result based fusion. 本发明将语义信息、图像处理技术和极化散射特性相结合,主要解决了现有基于极化分解的分类技术对具有聚集特性地物(如森林、建筑群等)的分类结果区域一致性较差的问题,提高了具有聚集特性地物的分类结果的区域一致性和边界保持性。 The present invention is semantic information, and image processing technology combined polarization scattering properties, mainly solves the area classification result of the consistency of a feature (e.g., forest, buildings, etc.) having characteristics based on the Decomposition aggregate classification techniques than the problem of poor and improve the consistency and border areas have classified the results gathered retention characteristics of surface features.

Description

基于语义信息和极化分解的极化SAR地物分类方法 Based on the semantic information and the Decomposition polarimetric SAR Classification Based

技术领域 FIELD

[0001] 本发明属于图像处理和遥感技术领域,涉及极化SAR图像的地物分类,具体是一种基于语义信息和极化分解的极化SAR地物分类方法,可用于含有具有聚集特性地物的低分辨极化SAR图像的地物分类。 [0001] The present invention belongs to the field of remote sensing and image processing, feature classification polarimetric SAR images relates, in particular polarimetric SAR is a feature and semantic information classification methods based on decomposition of polarization, it may be used to contain aggregate having the characteristics low-resolution polarimetric SAR images of the feature classifier.

背景技术 Background technique

[0002] 极化合成孔径雷达(Polarimetric Synthetic Aperture Radar,P0LSAR)图像处理是国防建设和经济发展的重要学科,受到越来越多人的关注和研究。 [0002] polarimetric synthetic aperture radar (Polarimetric Synthetic Aperture Radar, P0LSAR) image processing is an important subject of national defense building and economic development, attention and study more and more people. 与普通的单极化合成孔径雷达(Synthetic Aperture Radar,SAR)相比,极化SAR进行的是全极化测量,能获得目标更丰富的地物信息,为更加深入地研究目标的散射特性提供了重要的依据。 Compared with ordinary unipolar SAR (Synthetic Aperture Radar, SAR), carried out a full polarimetric SAR polarization measurements can be obtained target richer feature information, provide more in-depth study of the scattering characteristics of targets an important basis. 极化SAR 地物分类是极化SAR图像处理的重要任务之一,是极化SAR图像解译的前提。 Terrain Classification of polarimetric SAR is one of the important tasks polarimetric SAR image processing is a prerequisite for polarimetric SAR image interpretation. 极化SAR分割或地物分类的关键和难点在于同一地物的区域一致性和不同地物之间的边界保持性。 Key and difficult polarization SAR feature divided or classified in that a boundary region between the retention feature of the same consistency and different feature.

[0003] 极化SAR地物分类的方法有很多,主要分为三种:1)基于统计模型的分类方法;2) 基于电磁波散射机理的分类方法;3)基于图像处理技术的分类方法。 [0003] polarimetric SAR have many feature classification method, divided into three types: 1) based on the statistical classification model; 2) classification based on the electromagnetic wave scattering mechanism; 3) classification based on image processing techniques. 基于统计模型的方法主要有:Lee et al.根据极化协方差矩阵满足复wishart分布,提出了有监督的极化SAR 分类。 Based on statistical models are mainly:. Lee et al polarimetric covariance matrix distribution to meet the complex wishart proposed supervised polarimetric SAR classification. 但在实际的应用中,关于SAR图像类别的先验知识非常少。 However, in practical applications, prior knowledge about SAR image category is very small. 基于电磁波散射机理的方法有很多,1997年,Cloude等人首先提出了Η/α分类方法,通过分解得到了地物散射熵H和表征地物散射机理的角度α,实现了无监督的极化SAR图像分类。 There are many ways based on the mechanism of electromagnetic wave, in 1997, Cloude, who first proposed Η / α classification, obtained by decomposing the feature characterizing feature scattering entropy H and the angle [alpha] of the scattering mechanism, to achieve a polarization unsupervised SAR image classification. 1999年,Lee等人在Η/α分类方法的基础上结合统计分布引入了Wishart分类器,通过对Η/α分类方法的结果进行Wishart迭代提高了分类的精度。 In 1999, Lee et al., In conjunction with the introduction of a statistical distribution Wishart classifier based on Η / α classification methods, conduct Wishart iteration improves the accuracy of the classification by the results of Η / α classification methods. 2004年,Lee等人又提出了一种保持极化散射特性的分类方法,该方法利用Freeman分解得到的3种极化散射机理成分的功率进行初始分类, 并通过Wishart迭代进行合并与类别修正,达到了更好的分类效果。 In 2004, Lee et al also proposed a classification method of maintaining the polarization scattering characteristics, the method utilizes three kinds of power-polarized component scattering mechanism Freeman initial classification was subjected to decomposition, and combined with the category corrected by iterative Wishart, achieve better classification results.

[0004] 上述方法很好的利用了极化SAR数据的散射特性和极化信息进行分类,但这种基于像素的分类方法并没有考虑极化SAR图像的视觉特性,没有结合计算机视觉的方法和图像处理的方法进行分类。 [0004] The method takes advantage of a good scattering characteristics polarized polarization SAR data and classified information, but based on this classification of pixels does not consider the visual characteristic polarization SAR image without binding methods and computer vision the image processing method of classifying. 因此,包括上述方法的传统的极化SAR地物分类的方法存在很多缺陷:(1)同一地物的区域一致性不好,产生椒盐噪声式的分类结果图;(2)基于传统图像处理方法的极化SAR地物分类方法,对于具有明暗相间灰度变化的地物,如传统的基于像素点和超像素合并的分类方法都很难将这类地物分为一类;(3)对于复杂地物,如建筑群, 由于地物本身含有房屋、道路等,因此,地物散射特性并不单一,具有明暗相间的地物散射特性,很难很好的分为一个完整的区域,即使提取各种底层特征,使用各种区域合并的方法都很难将这些区域分在一起,但对于低分辨极化SAR图像地物分类,从人类视觉和图像理解的角度上应该将其分为一类。 Accordingly, there is a conventional method of classifying polarimetric SAR feature many drawbacks of the above method comprises: (1) poor consistency to the same object region, generating impulse noise categorization result FIG formula; (2) based on the conventional image processing method polarimetric SAR classification feature, the feature with respect to changes in light and dark gray, such as a conventional pixel and super pixel are combined classification based on such a feature is difficult to be divided into a class; (3) complex surface features, such as buildings, because the feature itself contains houses, roads, etc. Therefore, scattering characteristics are not a single feature, the feature has scattering properties of light and dark, good difficult area into a complete, even if extracting various underlying features, various methods are combined area of ​​these regions is difficult to be grouped together, but for low-resolution polarimetric SAR image classification, from the perspective of human vision and image understanding should be divided into a class. 因此,底层特征的提取已经很难将这类地物很好的分在一起,基于地物特性的高级特征需要进一步挖掘来进行分类。 Therefore, it is difficult to extract the underlying characteristics of this type of feature has good points together, based on surface features characteristic of advanced features require further excavation to classify.

[0005] 综上所述,上述几种极化SAR地物分类方法的像素分类精细,但仍存在一些缺陷, 尤其对具有聚集特性的地物(如建筑群、森林等),由于其本身地物散射不单一,具有明暗相间的地物散射特性,分类区域一致性较差,且边界易受噪声影响,容易产生椒盐式的分类结果。 [0005] In summary, the above several polarimetric SAR pixel classification Classification Based fine, but there are still some drawbacks, especially for the feature (e.g., buildings, forest, etc.) having an aggregating property, due to its own was not a single scattering, the scattering surface features having a characteristic light and dark, poor consistency classification region, and the boundary is susceptible to noise, easy to produce salt and pepper type classification results.

发明内容 SUMMARY

[0006] 本发明的目的在于克服上述已有方法的不足,提出了一种基于语义信息和极化分解的极化SAR地物分类方法,该方法对具有聚集特性的地物具有区域一致性好且边界精准的分类结果,提高了极化SAR地物分类的效果。 [0006] The object of the present invention to overcome the deficiencies of the above prior methods, a new classification polarimetric SAR semantic feature information and based on the decomposition of polarization, the method having the feature region having consistency properties aggregation good and the precise boundaries of the classification results, to improve the effect of terrain classification polarization SAR.

[0007] 本发明是一种基于语义信息和极化分解的极化SAR地物分类方法,针对事先获取的低分辨极化SAR图像进行无监督的地物分类,分类过程包括如下步骤: [0007] The present invention is a polarimetric SAR Classification Based on semantic information and polarization Decomposition performs unsupervised classification feature for the low-resolution polarimetric SAR images acquired beforehand, the classification process comprising the steps of:

[0008] 步骤1.输入待分类的极化SAR图像的数据,对该极化SAR数据进行处理,得到极化SAR数据三个通道的幅度值,融合三个通道幅度值得到极化SAR图像的后向散射总功率图,即span图,使用均值漂移得到span图的过分割结果图;并根据初始草图(prime sketch)稀疏表示模型提取span图由线段组成的边脊草图,即SketchMap。 [0008] Step 1. polarimetric SAR image input data to be classified, the polarimetric SAR data is processed to obtain an amplitude value of the three polarization SAR data channel, the channel amplitude is worth to three fusion polarimetric SAR image FIG total backscattered power, i.e. FIG span, used in FIG mean shift segmentation results obtained in FIG span; sparse representation model extraction and edge ridges span the sketch of FIG line segments, i.e., the initial sketch SketchMap (prime sketch).

[0009] 步骤2.对Sketch Map中的线段进行语义信息分析,根据线段聚集特性的统计分布,对线段赋予语义信息即两侧聚集、单侧聚集和孤立线段。 [0009] Step 2. Sketch Map segment information in semantic analysis, according to the statistical characteristics of the aggregate distribution line, i.e. on both sides of the line segment to impart semantic information aggregated, aggregation and unilateral isolation segments.

[0010] 步骤3.在aetch Map中,根据对线段赋予的语义信息,采用线段集合求解算法提取若干个不相交的聚集线段集合,并对每个聚集线段集合采用区域提取方法得到线段聚集区域R。 [0010] Step 3. aetch Map in accordance with the semantic information given segment, using segment collection algorithm for extracting a plurality of sets of disjoint segments aggregates, and each aggregate segment region extraction method using a set of line segments obtained aggregation region R .

[0011] 步骤4.对过分割结果进行区域合并:对步骤1中得到span图的过分割结果图,将线段聚集区域R对应的过分割区域采用临界区域众数投票合并策略;提取孤立线段所在过分割区域,采用不合并策略;对于其他区域,即剩余的区域,采用基于极化特征的区域合并策略,得到基于语义信息的极化SAR图像分割结果。 [0011] Step 4. The results of over-segmented regions merge: the over segmentation result obtained in step 1 in FIG. FIG span, corresponding to the line accumulation region R using the number of divided regions too critical region merge all voting strategy; extracts isolated segment is located through the divided region, not merging strategy employed; for other regions, i.e., the remaining region using region merging strategy based on the polarization characteristics, obtained based on polarimetric SAR image segmentation result semantic information.

[0012] 步骤5.利用极化分解对极化SAR数据进行H/ a -Wishart分类,并用马尔可夫随机场(Markov Random Field,MRF)对H/ a -Wishart分类结果进行邻域优化。 [0012] Step 5. Using the Decomposition of polarimetric SAR data H / a -Wishart classification, and for H / a -Wishart neighbor classification results used to optimize MRF (Markov Random Field, MRF).

[0013] 步骤6.通过众数投票(majority vote)策略将基于语义信息的极化SAR图像分割结果和基于MRF的H/ a -Wishart分类结果进行融合,得到待分类的极化SAR图像地物分类的最终分类结果。 [0013] 6. Step through all the votes (majority vote) policy polarimetric SAR image segmentation based on the results of the semantic information and polarization SAR image feature fusion based on MRF H / a -Wishart classification results obtained to be classified the final classification result of the classification.

[0014] 实现本发明的关键技术在于:针对在具有聚集性的地物(如建筑群、森林等)分类的区域一致性较差的问题,分析可知,低分辨极化SAR图像一般包括农田、城区、森林、山脉、桥梁等,根据人类的先验知识可知建筑群的结构线段应该很聚集且呈球形分布,桥梁的结构线段是线形分布等,将这些认知作为先验知识,对线段所含语义信息进行分析,赋予线段语义信息。 [0014] The key technology of the present invention is: in the area for the feature (e.g., buildings, forest, etc.) having aggregation classification problem of poor consistency, analysis, low-resolution polarimetric SAR images typically comprises fields, city, forest, mountains, bridges, construction of buildings segment should be gathered based on a priori knowledge of mankind and found spherical distribution, the structure of the bridge segments are linear distribution, as these cognitive prior knowledge of line segments containing semantic information analysis, semantic information given segment. 通过对线段语义信息分析,可以提取线段聚集区域,线段聚集区域对应于图像中建筑群、森林等地物,通过提取线段聚集区域得到了这些地物的一致区域,根据线段的语义信息分析,可以将过分割图像划分为线段聚集区域、孤立线段所在区域和无线段区域,线段聚集区域对应于建筑群等地物,孤立线段所在区域对应于线目标等,无线段区域一般对应于海洋、农田等地物,本发明对不同区域采用不同的合并策略,针对不同类型的地物采用更有针对性的合并策略,使各种地物都能够得到较好的合并,最后将分割和分类结果融合, 将语义信息和极化信息有机结合,得到区域一致性好且边界精准的分类结果,解决了具有聚集性地物的分类区域一致性较差的问题。 By analyzing the information of semantic segments may be extracted aggregated segment region corresponding to the line segment in the image region aggregation buildings, forest and other surface features, such a feature has been unanimously region by line segment extraction region aggregation, analysis semantic information segment to be dividing the image into line segments through the aggregation area, and the isolated segments area wireless segment region corresponding to the region segment aggregate feature, the isolated segment area buildings, etc. and the like corresponding to the target line, wireless segment generally corresponds to the ocean region, farmland feature, the present invention uses different for different regions of the consolidation strategy, more targeted policies for different types of merge feature, the variety of surface features are able to get a better merger, the final results of the segmentation and classification integration, the semantic information and polarization information combine to give a good consistency and a boundary region precise classification results, having solved the problem of the poor classification regions aggregation feature consistency.

[0015] 本发明与现有技术相比具有如下优点: [0015] The present invention and the prior art has the following advantages:

[0016] 1.从语义信息的分析上,本发明利用Primal Sketch稀疏表示模型得到span图的Sketch Map,根据Sketch Map,对线段包含的语义信息进行分析,提出了基于线段语义信息分析的区域划分技术,在Sketch Map上有效了提取了线段聚集区域。 [0016] 1. From the analysis of the semantic information, the present invention utilizes a sparse representation model obtained Primal Sketch span FIG Sketch Map, according Sketch Map, the information of line segments contained in the semantic analysis, the semantic information segment divided based on the area analysis technology to effectively extract the segment of the area gathered on the Sketch Map. 这些线段聚集区域对应于极化SAR图像中的城区、森林等地物。 These segments correspond to the collecting area polarimetric SAR images in urban areas, forests and other figures. 这些地物由于存在明暗相间的灰度变化而经常被分为多类,本发明很好的克服了这个缺点,有效提高了线段聚集区域分类的区域一致性。 These surface features due to the presence of light and dark gradation change is frequently divided into multiple categories, according to the present invention well overcomes this shortcoming, the region effectively improve the consistency of the line aggregation area classification.

[0017] 2.从图像处理技术上,在均值漂移过分割区域进行合并时,本发明对不同类型的地物区域采用不同的合并策略,区域合并更有针对性,保证了不同地物区域都能得到较好的合并,得到了基于语义信息的分割结果。 When [0017] 2. from the image processing technology, the mean shift through merging segmented regions, the present invention is combined with different strategies for different types of feature region, merge more targeted region, to ensure that the regions of different objects We can get a better combined to obtain the segmentation result based on semantic information.

[0018] 3.从极化分解上,本发明使用Η/α -Wishart分类,并用MRF进行邻域优化,得到像素级的分类结果,最后融合分割和分类结果,使用分割区域指导分类的区域一致性,同时分类结果也帮助分割区域的进一步合并,分割和分类相互作用得到更好的分类结果。 [0018] 3. from the Decomposition, consistent use of the present invention Η / α -Wishart classified and optimized with MRF neighborhood, pixel level classification result obtained, the final convergence result of segmentation and classification, region segmentation using a region classification guide , while also helping to further classification result of merging the divided areas, segmentation and classification interact better classification results. 本发明结合了图像处理的技术和基于电磁波散射机理的技术,融合了语义信息和极化信息,将语义信息、图像处理技术和极化散射特性相结合,提高了极化SAR地物分类结果的区域一致性和边界保持性。 The present invention combines the technology and image processing techniques based on the mechanism of electromagnetic wave scattering, polarization and integration of information semantic information, semantic information combining image processing techniques and polarization scattering characteristics, to improve the classification result polarimetric SAR feature regional border consistency and retention.

附图说明 BRIEF DESCRIPTION

[0019] 图1是本发明对极化SAR数据地物分类的流程图; [0019] FIG. 1 is a flowchart of a polarimetric SAR data classification feature of the present invention;

[0020] 图2是本发明使用的NASA/JPL AIRSAR L波段的全极化San Francisco数据的span 图; [0020] FIG. 2 is used in the present invention NASA / span FIG Polarimetric JPL AIRSAR L-band data of San Francisco;

[0021] 图3是本发明中均值漂移得到的过分割结果图; [0021] FIG. 3 is a result of over-segmentation according to the present invention, FIG mean shift obtained;

[0022] 图4是采用本发明得到的边脊草图,即Sketch Map [0022] Figure 4 is a sketch of the present invention obtained sides of the ridge, i.e. Sketch Map

[0023] 图5是本发明中线段的语义信息树型结构图; [0023] FIG 5 is a configuration diagram of a tree of semantic information segments in the present invention;

[0024] 图6是采用本发明得到的赋予语义信息的边脊草图; [0024] FIG. 6 is a side ridge sketch using impart semantic information obtained in the present invention;

[0025] 图7是本发明中的线段聚集区域提取过程示意图; [0025] FIG. 7 is a line segment region in the present invention aggregation schematic extraction process;

[0026] 图8是本发明中基于语义信息分析的线段聚集区域提取结果图; [0026] FIG. 8 is a view of the invention results in the extraction line accumulation region based analysis of semantic information;

[0027] 图9是本发明中线段聚集区域对应的过分割区域合并结果图; [0027] FIG 9 is a line segment corresponding to the collecting area through the divided region in the present invention, FIG combined results;

[0028] 图10是本发明中基于语义信息的图像分割结果图; [0028] FIG. FIG. 10 is an image segmentation result semantic information based on the present invention;

[0029] 图11是本发明中对分割和分类结果融合过程的示意图; [0029] FIG. 11 is a schematic diagram of the present invention, segmentation and classification result of the fusion process;

[0030] 图12是本发明使用的NASA/JPL AIRSAR L波段的全极化San Francisco数据的span 图; [0030] FIG. 12 is a NASA / JPL AIRSAR L-band used in the present invention, FIG fully polarized span data of San Francisco;

[0031] 图13是本发明中基于MRF的H/ a -Wishart分类结果图; [0031] FIG. 13 is the MRF H / a -Wishart classification result based on the present invention in FIG;

[0032] 图14是本发明的分类结果图。 [0032] FIG. 14 is a classification result of the present invention of FIG.

具体实施方式 Detailed ways

[0033] 实施例1 [0033] Example 1

[0034] 本发明是基于语义信息和极化分解的极化SAR地物分类方法,针对事先获取的低分辨极化SAR图像进行无监督的地物分类,参照图1,本发明的分类过程实现步骤包括: [0034] The present invention is semantic information and polarization decomposition polarimetric SAR classification based on the feature, perform unsupervised classification feature for the low-resolution polarimetric SAR images acquired beforehand, referring to FIG. 1, the classification process of the present invention is achieved steps include:

[0035] 步骤1,输入待分类的极化SAR图像的数据,对该极化SAR数据进行处理,得到极化SAR数据三个通道的幅度值,融合三个通道幅度值得到极化SAR图像的后向散射总功率图,如图2所示,即NASA/JPLAIRSAR L波段的全极化San Francisco数据的span图。 [0035] Step 1, the input polarization SAR image data to be classified, the polarimetric SAR data is processed to obtain an amplitude value of the three polarization SAR data channel, the channel amplitude is worth to three fusion polarimetric SAR image after the total backscatter power as shown in FIG 2, i.e. span FIG NASA / JPLAIRSAR L full band polarization data of San Francisco. 对span图使用均值漂移得到span图的过分割结果图;并根据prime sketch稀疏表示模型提取span图由线段组成的边脊草图,即Sketch Map。 FIG using mean and span drift over the span of FIG. FIG segmentation results; and said ridge side model extraction sketch of FIG span line segments, i.e. Sketch Map The prime sketch sparse.

[0036] 首先对极化SAR数据进行处理得到协方差矩阵,根据协方差矩阵对角线元素的三个值得到三个通道的幅度值,融合三个通道幅度值得到极化SAR图像的span图。 [0036] First polarimetric SAR data obtained by processing the covariance matrix, the amplitude value is worth three channels according to three of the diagonal elements of the covariance matrix, fusion to span three channels worth amplitude polarimetric SAR image of FIG. . 在span 图上进行第一个操作是使用均值漂移得到span图的过分割结果图,如图3所示。 The first operation performed on the span is too FIG FIG using mean-shift segmentation result obtained in FIG span, as shown in FIG.

[0037] 第二个操作是采用边-脊检测稀疏编码方法提取Sketch Map,其提取步骤包括: [0037] The second operation is the use of edge - detecting sparse coding method for extracting a ridge Sketch Map, which extracting step comprises:

[0038] 首先,构造N个尺度和M个方向的高斯一阶导滤波器和高斯二阶导滤波器,形成滤波器组。 [0038] First, the configuration of the N and M directions scale Gaussian filter and the first derivative of the Gaussian second order derivative filter, form a filter bank. 其中N取值为3,且M取值为18。 Wherein the value of N is 3, and M value is 18. 如图2所示为span图像,将span图像与滤波器组进行卷积,得到每个像素的联合响应,提取联合响应的最大值作为该像素的边/脊强度,且将最大响应滤波器的方向作为该像素的局部方向。 As shown in FIG 2 will span to span the image and image convolution filter bank, to obtain the combined response of each pixel, extracts the maximum value as a response to joint edge / ridge intensity of the pixel, and the maximum response of the filter as the local orientation direction of the pixel. 对边/脊强度图进行非极大抑制处理, 得到建议草图S=,根据建议草图最大联合响应的位置,把建议草图S=中与该位置连通的点连接成线段,生成一个边/脊原始模型S sk,。 On the edge / ridge suppress non-maximum intensity diagrams to give S = recommended draft, draft recommendation based on the position of the maximum joint response, S = the suggestions in the sketch in communication with the position of the connecting line segment, creates an edge / ridge original model S sk ,. ;

[0039] 其次,在边脊模型中添加新线段,评价图像的编码长度增益AL,若AL< ε,ε是阈值,取值为10,则拒绝接受该线段,否则接受,并搜索,将建议草图S:;中该新线段末端与其余像素在平均拟合误差内的分割线作为下一个新建议线段,若存在新建议线段,则重新计算添加该新建议线段后的图像编码长度增益AL,若AL< ε则拒绝接受该新建议线段,否则接受该新建议线段,迭代地添加新线段,直到不存在新建议线段即得到了边脊草图,如图4所不,即为Sketch Map。 [0039] Next, to add a new line on the side ridges model, the code length gain AL evaluation image, if AL <ε, ε is a threshold value of 10, then reject the segment, or to accept, and search for the recommended S :; sketch of the new line segment end pixels and dividing the remaining line in the average fitting error in a new recommendation as the next segment, if the segment new proposals exist, add the new recalculated recommendation after image encoding segment length AL gain, If AL <ε then rejected the new proposal segment, or segment accept the new proposal, a new segment is added iteratively, until there are no new proposals edge ridge line segment to obtain a sketch, not shown in Figure 4, namely sketch Map.

[0040] 步骤2,对Sketch Map中的线段进行语义信息分析,根据线段聚集特性的统计分布,对线段赋予语义信息即两侧聚集、单侧聚集和孤立线段。 [0040] In step 2, the line segment Sketch Map of semantic information analysis, according to the statistical distribution of the aggregate characteristic line, i.e. on both sides of the line segment to impart semantic information aggregated, aggregation and unilateral isolation segments.

[0041] 2. 1针对低分辨率极化SAR图像,对于具有聚集特性的地物,以建筑群为例,其线段是由亮的建筑物和暗的地面形成的,这样的结构反复出现,则形成了建筑群,其对应的sketch线段特点通常是分布密集,且线段方向大多成近似水平和垂直。 [0041] 2.1 for the low-resolution polarimetric SAR images, with respect to the aggregate feature characteristic to buildings, for example, which is a bright line and a dark building ground form, this configuration repeated, a complex is formed, which corresponds to the characteristic line sketch is usually densely distributed, and most of the line segment to be approximately horizontal and vertical direction. 对于森林地物,其sketch线段也分布密集,但线段方向杂乱无章。 For the forest surface features, its sketch segment also densely distributed, but the line direction disorganized. 对于桥梁,其sketch线段成流形分布等。 For the bridge, which sketch line segment distribution manifolds. 因此,线段的分布结构都含有一定的语义信息,根据不同地物类型对应的sketch线段的分布不同,得出线段主要对应于三种地物信息:线目标、球形聚集分布的地物和不同地物之间的边界。 Thus, the structure of the distribution segment contains some semantic information, corresponding to the distribution depending on the different types of feature segments sketch, corresponding to the three major segments derived feature information: feature and the target line differently, spherical aggregates distributed the boundary between the objects.

[0042] 2. 2两个线段之间的距离定义为线段中点的欧式距离,用线段K近邻的平均距离表示线段的聚集程度;根据线段的聚集性的统计分布,将线段赋予语义信息:聚集线段和孤立线段;根据聚集线段的拓扑结构可以分为两侧聚集和单侧聚集。 [0042] 2.2 defines the distance between two line segments is the midpoint of the Euclidean distance, the average distance of nearest neighbor line K indicates the degree of aggregation of segments; according to the statistical distribution of the aggregated segment, the semantic information given segment: aggregation and isolated line segment; topology aggregation according to the line segment can be divided into one side and both sides of the aggregation aggregates.

[0043] 2. 3根据线段聚集性的统计分布,将线段的语义信息以树型结构表示,如图5所示,即线段的语义信息树型结构示意图。 [0043] 2.3 The statistical distribution of the aggregation of segments, the segments semantic information represented in a tree structure, shown in Figure 5, a schematic diagram of the semantic tree structure is information segment. 两侧聚集对应于森林、建筑群等地物;单侧聚集对应于一边有森林或建筑群等地物的边界;孤立线段对应于线目标、桥梁等流形地物或两种不同地物的边界。 Both correspond to feature aggregation forest, buildings and the like; unilateral gathered corresponding to one side of the boundary have a feature or group of buildings and other forest; isolated manifold segment corresponding to the target feature lines, bridges, etc., or of two different feature boundary. 图6显示了赋予线段语义信息的Sketch Map,其中,灰色线段为聚集线段,黑色线段为孤立线段。 Figure 6 shows a segment given semantic information Sketch Map, wherein an aggregated segment gray line, an isolated black line segment.

[0044] 本发明根据线段聚集特性的统计分布,对线段赋予语义信息,赋予的语义信息有两侧聚集、单侧聚集和孤立线段,线段的语义信息分析是线段聚集区域提取的前提,为后面的地物区域划分提供依据。 [0044] According to the present invention, the statistical characteristics of the aggregate distribution segment, a line segment for imparting semantic information, semantic information are given to both aggregated, aggregation and unilateral isolation segment, segment analysis semantic information is aggregated segment region extraction prerequisite for later the feature provides a basis for zoning.

[0045] 步骤3,在Sketch Map中,根据对线段赋予的语义信息,采用线段集合求解算法提取若干个不相交的聚集线段集合,并对每个聚集线段集合采用区域提取方法得到线段聚集区域R。 [0045] Step 3, in Sketch Map in accordance with the semantic information given segment, using segment collection algorithm for extracting a plurality of sets of disjoint segments aggregates, and each aggregate segment region extraction method using a set of line segments obtained aggregation region R .

[0046] 3. 1符号定义:sketch线段集合为S ;空间约束阈值δ 1;线段生长阈值δ 2;满足空间约束线段集合U ;聚集线段集合 [0046] 3.1 Symbols defined: sketch segment collection is S; spatial restriction threshold δ 1; line growth threshold δ 2; segment meet space constraints set U; aggregated segment collection

Figure CN103294792BD00091

线段聚集区域R = Ir1, r2, ···, rn}; Aggregated segment region R = Ir1, r2, ···, rn};

[0047] 3. 2首先采用线段集合求解算法,本算法类似于区域生长的方法,不过本发明是以线段为基元进行生长的,得到聚集的线段集合,利于线段聚集区域的提取,具体步骤如下: [0047] Firstly, 3.2 segment collection algorithm, this algorithm is similar to the method of region growing, but the present invention is a cell line grown to obtain aggregated segment collection, aggregation facilitate the extraction of the line area, the specific steps as follows:

[0048] 3. 2. 1首先得到sketch集合S,依据森林、建筑群等线段聚集区域的线段具有聚集性,对每条线段的k近邻进行统计,计算每条线段的k近邻平均距离,从k近邻平均距离的直方图统计看出图像线段是否具有聚集性,如果具有某种聚集性,说明有存在这样的地物, 根据直方图统计,得到空间约束阈值S 1和线段生长的阈值S 2。 [0048] 3. 2.1 was first sketch set S, the line segment having aggregation based aggregation forest areas, buildings and the like, for each segment k nearest neighbor statistics, the average nearest neighbor distance k is calculated for each line segment, from k-nearest neighbor histogram of the average distance to see whether the line segment image aggregation, if having some aggregation, indicating the presence of such a feature, according to the histogram, a threshold value S to give space constraints. 1 and the line segment growth threshold S 2 .

[0049] 3. 2. 2初始设1\为空集;根据种子线段的阈值得到初始种子线段 [0049] 3.2.2 Initial set 1 \ is the empty set; worth the initial seed to the seed line segment based on the threshold

Figure CN103294792BD00092

随机选取种子线段仍进行生长,此时,?:· = {奶·};生长的准则为,如果线段的某个近邻%满足线段生长阈值S 2,则生长为聚集线段集合=他,%},遍历其k近邻直到没有可生长的线段,假设此日T Randomly selected seed segment still be growing, at this time,:? * = {*} Milk; guidelines for growth, if a neighbor segments satisfy the segment% growth threshold S 2, the aggregate growth for the segment collection = he,%} traversing the line until its k-nearest neighbor of no growth, assuming that this day T

Figure CN103294792BD00093

对此时T1中没有遍历过的线段,依次作为种子线段进行生长,这样迭代生长直到所有生长进来的线段不能再生长为止,此时得到一个聚集线段集合T1。 No T1 of segment traversed this time, sequentially line grown as a seed, so that the growth of the iteration until all segments come regrowth growth, at which time segment to obtain a set of aggregated T1.

[0050] 3. 2. 3若初始种子线段集合U中还有线段未进行生长,则选一条线段为种子线段继续生长,这样迭代生长,直到所有的初始种子线段都得到生长。 [0050] 3.2.3 If the initial seed set U line segment is not grown there, it is a segment selected from segments continue to grow as the seed, the growth of this iteration, the initial seed until all segments have been grown. 最后得到若干个不相交的线段集合T k。 Finally, to obtain a plurality of disjoint segment collection T k.

[0051] 3. 3对每个聚集线段集合采用区域提取方法:在线段集合的基础上,以圆形的基元进行区域提取得到聚集线段集合所在的区域。 [0051] 3.3 region extraction method using a set of aggregation for each line: set on the basis of the segment, a circular region extracting primitives obtain aggregated segment collection area located.

[0052] 3. 3. 1圆形基元构造:取线段生长阈值δ 2为圆的半径构造圆盘。 [0052] 3.3.1 circular configuration primitives: Take line growth threshold δ 2 is configured as a disk radius of the circle. 采用圆形是为了保持区域边界的平滑特性,半径取S2是为了保证填充线段间的最大间隙。 Using a circular region in order to maintain smooth boundary characteristics, taking the radius of S2 is to ensure that the maximum clearance between the fill line. 因为同一线段聚集区域其线段间隔应该是相近的,而生长阈值S 2代表了生长出的线段集合的最大线段间隔,因此,这里取S2作为圆盘半径。 Because the same segment aggregate area which should be close to the line spacing, S 2 grown threshold line represents the maximum interval of the set of segments grown, therefore, there is taken as S2 radius of the disk.

[0053] 3. 3. 2闭操作:使用结构元素B对集合A的闭操作,表示为A · Β,定义为 [0053] 3.3.2 closing operations: using the structure element B and closing operation of the set A, expressed as A · Β, defined as

Figure CN103294792BD00094

[0055] 其中,X Φ 5表示B对A进行膨胀操作,JQ 5表示B对A进行腐蚀操作。 [0055] wherein, X Φ 5 A to B represents an expansion operation, JQ 5 A to B represents the etching operation.

[0056] 这个公式说明,使用结构元素B对A的闭操作,就是用B对A进行膨胀,然后用B 对结果进行腐蚀。 [0056] This formula is described, using the structure element B of the closing operation of the A, B to A is carried out by expansion, the results can be etched using B. 图7为本发明中的线段聚集区域提取过程示意图,在图7(a)中,结构元素B为上面构造的圆形基元,集合A是由线段构成的集合。 Aggregation line area in FIG. 7 is a schematic view of the invention during the extraction, in FIG. 7 (A), a circular structuring element B cell configuration above, set A is a set of line segments. 对集合A进行膨胀是指使用结构B在图像A中线段上的每一点移动,所有位移的集合即为膨胀后的结果。 A set of expansion means for using the result of structure B moves each point on the line segment of image A, is the set of all displacement expander. 膨胀操作如图7(b)所示,膨胀结果如图7(c)所示。 Dilation FIG 7 (b), the expansion results are shown in 7 (c) shown in FIG. 膨胀之后进行腐蚀操作,腐蚀操作如图7(d)所示,最终的闭操作结果如图7(e)所示。 After expansion etching operation, an etching operation in FIG. 7 (d), the final result of the closing operation of FIG 7 (e) in FIG. 从图中可以看出,闭操作得到了线段集合A所在的区域, 消除了狭长的细缝,得到了一致的连通区域。 As can be seen from the figure, closed area obtained in a segment where the set A, eliminating the narrow slit, to obtain a consistent communication area. 对每个聚集线段集合都进行区域提取,得到线段聚集区域R,图8显示了线段聚集区域提取的结果。 Aggregation are set for each segment region extraction, to obtain aggregated segment region R, FIG. 8 shows the results of line segment extraction collecting area.

[0057] 步骤4,对过分割结果进行区域合并:在线段聚集区域R对应的过分割区域采用临界区域众数投票合并策略;提取孤立线段所在过分割区域,采用不合并策略;对于其他区域采用基于极化特征的区域合并策略,得到极化SAR图像分割结果。 [0057] Step 4, to the over-segmentation result region merging: line segment aggregate area R corresponding to the over-segmentation region using several critical region public voting combined policies; extracts isolated segment located over the divided region, using no merging policy; for other regions using region merging strategy based on the polarization characteristics, the polarization SAR image segmentation results obtained.

[0058] 4. 1线段聚集区域对应的过分割区域采用临界区域众数投票合并策略:由于线段聚集区域的区域一致性好,但边界不精准,而过分割的边界精准,因此,对于线段聚集区域的边界和过分割区域边界不吻合情况,采用临界区域众数投票合并策略;对于线段聚集区域和过分割区域的重叠情况有两种:一是某些过分割区域被线段聚集区域全部覆盖;二是线段聚集区域的边缘区域和过分割区域部分重叠,这里将边缘部分重叠区域叫做临界区域。 [0058] 4.1 segment focusing region corresponding to the divided region had adopted several critical areas congregation voted to merge strategy: As good area consistency collecting area of ​​the line, but the border is not accurate, precise and over-segmentation of the border, therefore, for the gathering line border and over the border region divided areas do not match, the congregation voted to adopt several critical areas consolidation strategy; line for the collecting area and over the divided region of overlap, there are two: First, some areas are too divided to cover all segments focusing region; Second edge region segment and a collecting area through the divided regions partially overlap, where an edge portion of the overlapping area is called the critical region. 对于第一种情况,直接合并均值漂移过分割区域,对于第二种情况,根据众数投票策略, 如果线段聚集区域占过分割区域的50%以上,则将这个过分割区域全部合并为线段聚集区域,否则,将其划分为无线段区域;最后在过分割图中得到合并的线段聚集区域及,这就保证了这些很难合并的过分割区域得到很好的合并。 In the first case, a direct merger mean shift over the divided region, the second case, according to the number of public voting strategy, if the line had gathered region accounts for more than 50% of the divided region, this too will be divided region gathered all combined into segments area, otherwise, be divided into regional wireless segment; and finally get merged segment had gathered in the area and dividing the figure, which ensures that those hard to merge over the divided region is well combined. 图9显示了合并后的线段聚集区域的结果,可以看出,建筑群这种线段聚集区域得到了很好的合并。 Figure 9 shows the results of a line segment of the combined aggregate area, it can be seen that line segment buildings accumulation region has been well consolidated.

[0059] 4. 2对于孤立线段,提取其所在的过分割区域。 [0059] For isolation segments 4.2, in which it is over-extracted divided region. 对这些区域不进行合并。 These areas do not merge. 根据线段的语义信息分析,对于孤立线段对应于图像中的线目标或者两种地物的边界,在进行区域合并时,如果对孤立线段所在区域进行合并,则会使线目标消失,或者两个不同的区域合并。 The segment analysis semantic information, corresponding to the image for the isolated segment of the boundary line, or two kinds of surface features of the target, when performing the merging region, if isolated segment Area merge, the line will cause the target to disappear, or two different regions merge. 因此,本发明对孤立线段所在的区域不进行区域合并。 Accordingly, the present invention is an isolated segment of the region where merging region is not performed.

[0060] 4. 3对于其他区域,定义为无线段区域,采用基于极化特征的合并策略。 [0060] 4.3 for the other regions, defined as regions of a wireless section, combined using the strategy based on polarization characteristics. 首先将均值漂移得到的每个过分割区域看作超像素,统计超像素的极化特性,采用三通道灰度直方图统计作为特征,对于每个通道,将灰度值量化为16份,然后计算在这个特征空间的区域直方图。 First through each divided region obtained is regarded as the mean shift superpixel, the superpixel statistical polarization characteristic, a three-channel histogram as statistical characteristics, for each channel, the 16 gradation value is quantized parts, and in this calculation region histogram feature space. 三个通道共有16X3 = 48份。 A total of three channels 16X3 = 48 parts. 每个区域可以用一个48维的向量表示,如用Histp 表示区域P的归一化直方图特征。 Each area can be a 48-dimensional vector representation, such as represented by region P Histp normalized histogram feature.

[0061] 根据Bhattacharyya系数计算公式,计算两个区域P和Q的相似性P (P,Q),P (P, Q)定义如下: [0061] The Bhattacharyya coefficient calculation formula to calculate the P and Q of two regions of similarity P (P, Q), P (P, Q) is defined as follows:

Figure CN103294792BD00101

[0063] 其中,HiStp和HiSttj分别是R和Q的归一化直方图。 [0063] wherein, HiStp HiSttj and R and Q are normalized histograms. 上标u表示直方图的第u个分量。 Superscript u represents the u-th component of the histogram.

[0064] 设定合并阈值U,相似性大于阈值的相邻区域进行合并,合并后的区域再次计算直方图特征,迭代合并直到没有可合并的区域为止,得到基于语义信息的分割结果。 [0064] The combined set threshold value U, the similarity is larger than the threshold value adjacent regions are merged, the merged area calculated histogram feature again, the iterative merging until no far region can be combined to obtain semantic information based on the segmentation result. 图10为基于语义信息的分割结果。 10 is based on the segmentation result semantic information.

[0065] 本发明不仅提出了基于语义信息的线段聚集区域提取方法提取边脊草图上的线段聚集区域,还采用不同策略对过分割区域进行合并:对线段聚集区域,采用临界区域众数投票策略指导过分割块的区域合并;对于孤立线段所在的过分割区域,采用不合并策略; 剩下的区域为无线段区域,采用基于极化信息的区域合并策略。 [0065] The present invention is not proposed extraction line collecting area on the side ridge sketch extraction segments aggregate area semantic information based also uses different strategies over the divided regions merge: line segment aggregate area, using the mode critical region voting strategy guided the block merging divided regions; for isolated line over the divided regions is located, without using the merge policy; remaining area wireless segment region using region merging strategy based on the polarization information. 本发明结合了语义信息对线段聚集区域进行提取,对不同类型的区域采用不同的合并策略,很好的解决了线段聚集的区域分类难的问题。 The invention combines semantic information gathering area for line extraction, the combined use of different strategies for different types of areas, a good solution to the problem of aggregation of segments area classification difficult.

[0066] 步骤5,利用极化分解对极化SAR数据进行H/ a -Wi shart分类,并用MarkoV Random Field对H/a -Wishart分类结果进行邻域优化。 [0066] Step 5, using the polarization decomposition polarimetric SAR data H / a -Wi shart classification, and for H / a -Wishart neighbor classification results used to optimize MarkoV Random Field.

[0067] 5. 1使用H/ a -Wishart分类方法得到初始的分类结果 [0067] 5.1 using H / a -Wishart initial classification of the classification result obtained

Figure CN103294792BD00111

其中S是像素点的集合。 Where S is the set of pixels. Wishart距离采用的是Kersten等修正后的基于wishart分布的距离测度。 Wishart distance used is the distance Kersten et wishart distribution measure based on the corrected. 1[°]中每个像素标记 1 [°] of each pixel marked

Figure CN103294792BD00112

L为总的类别数。 L is the total number of categories. 这里L = 8。 Where L = 8.

[0068] 5.2给定一组观测值0= {Ts|se S},其中,!^是像素点s的极化相干矩阵。 [0068] 5.2 Given a set of observations 0 = {Ts | se S}, where s ^ is the pixel polarized coherent matrix!. 已知协方差矩阵服从复wishart分布。 Known covariance matrix obey complex wishart distribution. 根据初始分类结果,使用第i类的观测样才 The initial result of the classification, the observation using only the sample class i

Figure CN103294792BD00113

:来估计该类的分布参数σ,并计算LXL的类间距离矩阵D: : Parameters of the distribution of the class σ, and calculates the distance matrix between LXL Class D:

Figure CN103294792BD00114

[0070] 其中D1,表示第i类和第j类的距离,d表示平均相干矩阵的欧式距离。 [0070] where D1, represents the distance class i and class j, d represents the average Euclidean distance coherent matrix.

[0071] 5. 3基于MRF的框架,数据项是每个像素点的类似然值,平滑项是类间距离。 [0071] 5.3-based framework of the MRF, and then the data item is similar to the value of each pixel, smoothing term is the distance between classes. 最小化能量函数如下: Minimizing the following energy function:

Figure CN103294792BD00115

[0073] 其中,I气)是像素s处观测数据的类条件概率,Ns是像素s的邻域像素集合。 [0073] wherein, I gas) is observed at the class conditional pixel data s probability, Ns s is a pixel neighborhood set of pixels. 是正则化参数。 It is the regularization parameter. 式(3)中的总能量通过a-expansion算法来最小化。 The total energy of the formula (3) is minimized by a-expansion algorithm.

[0074] 步骤6,融合基于语义信息的分割结果和基于MRF的H/ a -Wishart分类结果。 [0074] Step 6, the fusion / a -Wishart classification result based on the semantic information and segmentation result based on the MRF H.

[0075] 本发明结合了分割结果的区域一致性和分类结果的像素级精准性的优点,得到更加好的分类结果。 Pixel level accuracy advantages [0075] The present invention combines the results of the segmentation and classification regions consistent results, more good classification results obtained. 这种融合策略组合了无监督分割和基于像素的分类结果,基于maiority vote策略来进行分类,得到待分类的极化SAR图像地物分类的最终分类结果。 This fusion strategy combines the final classification result based on unsupervised segmentation and classification result of the pixel-based classification strategy maiority vote, to obtain a polarization SAR Image Classification is to be classified. 图11为分割和分类结果融合过程示意图,其主要步骤包括: 11 is a schematic view of the classification result and dividing the fusion process, which mainly comprises the step of:

[0076] 6. 1分割:分割得到一致的区域,区域数要大于最终类别数,且稍稍高于分类数目;图11(a)为4个分割区域的分割示意图,其中用1~4来表示4个分割区域; [0076] 6.1 segmentation: obtained by dividing the same region, the number of regions to be larger than the final number of categories, and slightly higher than the number of classification; FIG. 11 (a) is a schematic view of dividing the four divided regions, wherein a represents 1 to 4 four divided regions;

[0077] 6. 2基于像素的分类:基于图像的散射特性进行像素级的分类,图11 (b)为基于像素点的分类示意图,其中用白色、黑色和灰色代表三类。 [0077] 6.2 pixel based classification: the classification is based on pixel-level image scattering characteristics, FIG. 11 (b) is a schematic view of the classification based on the pixel, wherein the white, black and gray three representatives.

[0078] 6. 3融合分割和分类:采用majority vote策略,对于分割图中的每个区域,选择对应的分类结果中像素个数最多的类别作为这个区域的类别,将最终分类结果图的对应区域标记为该类别。 [0078] 6.3 Fusion segmentation and classification: The majority vote policy, for each divided region in the figure, corresponding to the selected classification result in the largest number of pixels as a category class of this region, corresponding to the final classification result of FIG. regional mark for the category. 这样使分类结果的区域一致性大大提高。 This makes the classification results of regional coherence greatly improved. 需要注意的是,在majority vote中,像素的邻域不是固定的邻域窗,而是分割属于同一个区域中的像素。 Note that, in the majority vote, the neighborhood of the pixel neighborhood window is not fixed, but belong to the same divided region of the pixel. 图11(c)为对分割图和基于像素点的分类结果进行融合的示意图,对该图中每个区域采用众数投票策略,得到图11(d)所示的分类结果。 FIG. 11 (c) is performed on the segmentation map and classification based on the result of the pixel schematic fusion, the mode of each region using the voting policy FIG afford FIG. 11 (d), the classification results. 经过融合分割和分类结果,得到待分类的极化SAR图像地物分类的最终分类结果图,如图14所示。 After fusion segmentation and classification results to obtain a final classification result of FIG polarization SAR Image Classification be classified, as shown in FIG.

[0079] 本发明利用Primal Sketch稀疏表示模型得到span图的Sketch Map,根据Sketch Map,对线段包含的语义信息进行分析,提出了基于线段语义信息分析的线段聚集区域提取技术,在Sketch Map上有效了提取了线段聚集区域。 [0079] The present invention utilizes Primal Sketch sparse representation model of the span of FIG Sketch Map, according Sketch Map, semantic information segments contained in the analysis, the extract based on the line aggregation region segment semantic information analysis technology, effective in Sketch Map the extracted segment focusing region. 这些线段聚集区域对应于极化SAR图像中的城区、森林等地物。 These segments correspond to the collecting area polarimetric SAR images in urban areas, forests and other figures. 这些地物由于存在明暗相间的灰度变化而经常被分为多类,本发明很好的克服了这个缺点,有效提高了线段聚集区域分类的区域一致性。 These surface features due to the presence of light and dark gradation change is frequently divided into multiple categories, according to the present invention well overcomes this shortcoming, the region effectively improve the consistency of the line aggregation area classification. 同时,为保持极化散射特性,对极化SAR数据进行H/ a -Wishart分类,并用MRF进行邻域优化。 Meanwhile, in order to maintain polarization scattering characteristics of polarimetric SAR data H / a -Wishart classification, and the neighborhood Optimization MRF. 基于极化分解的分类结果精细,但杂点较多,因此,本发明融合分割结果和基于MRF的H/ a -Wishart分类结果,得到待分类极化SAR图像的地物分类结果。 Based on the Decomposition fine classification results, but more hetero point, therefore, the present invention is based on fusion segmentation results and the MRF H / a -Wishart classification result, feature classification for classification results polarimetric SAR images. 将语义信息和极化分解有效的融合得到最终的分类结果。 The semantic information and Polar Decomposition effective integration result of classification obtained.

[0080] 实施例2 [0080] Example 2

[0081] 基于语义信息和极化分解的极化SAR地物分类方法同实施例1,仿真的数据和图像说明如下: [0081] Classification polarimetric SAR feature and semantic information based on polarization with decomposition 1, the simulation and the image data is illustrated by Examples as follows:

[0082] 1.仿真条件 [0082] 1. The simulation conditions

[0083] (1)选取NASA/JPL AIRSAR L 波段的全极化San Francisco 数据; [0083] (1) Select the NASA / JPL AIRSAR L full band polarization San Francisco transactions;

[0084] (2)仿真实验中,Primal Sketch稀疏表示模型中的参数N取值为3,M取值为18, 阈值ε取值为20 ; [0084] (2) Simulation experiment, Primal Sketch sparse representation model parameter value of N is 3, M value is 18, the threshold value ε is 20;

[0085] (3)仿真实验中,近邻数k取9 ; [0085] (3) In the simulation experiment, the number of neighbor k has 9;

[0086] (4)仿真实验中,种子线段阈值δ诹20 ;线段生长阈值δ 2取12 ; [0086] (4) In the simulation experiments, the threshold value [delta] Suwa seed line 20; Line Growth 12 take the threshold value δ 2;

[0087] (5)仿真实验中,区域合并阈值U取0· 7 ; [0087] (5) in the simulation experiment, takes region merging threshold U 0 · 7;

[0088] (6)仿真实验中,基于MRF的H/ a -Wishart分类中邻域窗选择为3*3。 [0088] (6) in the simulation experiment, MRF of H / a -Wishart classification is based on the window selecting neighborhood 3 * 3.

[0089] 2.仿真内容与结果 [0089] 2. Content and simulation results

[0090] 利用NASA/JPLAIRSAR L波段的全极化San Francisco数据,用本发明对其进行地物分类。 [0090] San Francisco using full polarimetric data NASA / JPLAIRSAR L band, with the feature of the present invention is subjected to classification. 图12为span图,与图2为同一幅图,为方便对分类结果进行评价,将图12、图13、 图14 一并显示,图14为本发明的分类结果图。 FIG 12 is a span, and Figure 2 is the same figure, the evaluation of the classification results for convenience, FIG 12, FIG 13, FIG 14 collectively shows the results of the classification of FIG. 14 of the present invention. 从图中可以看出,本发明分类结果的区域一致性较好且边界部分也比较精准,尤其对建筑群区域,能够得到一个大的一致区域,更符合人类视觉对图像的理解,对桥梁这种线目标,本发明的策略能够得到较好的分类结果,能够将桥梁很好的分出来。 As can be seen from the figure, the present invention is the consistency of the region classification result and the boundary portion is preferably also more accurate, especially for buildings area, it is possible to obtain a large uniform area, more consistent understanding of the human visual images, which bridge kind of line objectives, strategies present invention can get a better classification results can be very good points out of the bridge. 综上所述,由于语义信息的加入,本发明能够得到更适用于人类进行图像理解的分类结果,地物的区域一致性和边缘精准性都得到了提高。 In summary, the addition of semantic information, the present invention is possible to obtain an image more suitable for human understanding classification result, the consistency and the edge region of the feature precision are improved.

[0091] 实施例3 [0091] Example 3

[0092] 基于语义信息和极化分解的极化SAR地物分类方法同实施例1-2,其中基于MRF的H/a -Wishart分类方法同实施例1中的步骤5,作为本发明的对比实验,仿真的数据和结果如下: [0092] Classification Based polarimetric SAR semantic information and polarization Decomposition with Examples 1-2, wherein the MRF H / a -Wishart classification based on the same implementation steps in Example 15, the present invention Comparative experiment, simulation data and results are as follows:

[0093] 1.仿真条件 [0093] 1. The simulation conditions

[0094] (1)选取NASA/JPLAIRSAR L 波段的全极化San Francisco 数据; [0094] (1) Select the NASA / JPLAIRSAR L full band polarization San Francisco transactions;

[0095] (2)仿真实验中,基于MRF的H/ a -Wishart分类中邻域窗选择为3*3。 [0095] (2) in the simulation experiment, MRF of H / a -Wishart classification is based on the window selecting neighborhood 3 * 3.

[0096] 2.仿真内容与结果 [0096] 2. Content and simulation results

[0097] 利用NASA/JPL AIRSAR L波段的全极化San Francisco数据,用基于MRF的H/ a -Wishart分类方法进行分类,该方法是基于像素点的分类方法,图12为span图,图13为基于MRF的H/ a -Wi shart分类方法的结果。 [0097] using the NASA / full polarization San Francisco data JPL AIRSAR L band, with the classification of the MRF H / a -Wishart based classification, the classification is based on pixel, FIG. 12 is a span, and FIG. 13 based on a result of MRF H / a -Wi shart classification method. 从图中可以看出,该方法分类精细,但产生椒盐式的分类结果,尤其是对于建筑群这种具有聚集特性的地物,由于其包含建筑物和道路等, 它们的散射类型不一致,因此产生不一致的分类结果,但对于低分辨极化SAR图像,我们在进行图像理解时,希望能够得到一致的建筑群分类结果,因此,该方法对具有聚集特性的地物分类区域一致性较差,边界也易受噪声影响。 As can be seen from the figure, the fine classification method, but it produces salt and pepper type classification results, especially feature for such buildings have aggregation properties, because it contains buildings and roads, they are inconsistent with the type of scattering, and therefore when inconsistent classification results, but the low-resolution polarimetric SAR images, we have appreciated that when the image, hoping to get the same buildings classification result, therefore, the less consistent method of classification regions having the feature characteristic of the aggregate, border is also susceptible to noise.

[0098] 本发明与基于MRF的H/ a -Wishart分类方法的结果对比: [0098] The present invention is based on a comparison result of the MRF and the H / a -Wishart classification methods:

[0099] 将本发明与基于MRF的H/ a -Wishart分类的地物分类结果进行对比。 [0099] A feature of the present invention is compared with a classification result based on MRF H / a -Wishart classification. 实验结果如下,图12是为span图,图13是基于MRF的H/ a -Wishart分类的结果图,图14为本发明的分类结果图。 Results are as follows, FIG. 12 is a view of the classification result span, and FIG. 13 is a result of the MRF FIG H / a -Wishart classification based on the present invention. FIG. 14. 对比图13和图14可以看出,本发明较基于MRF的H/a-Wishart分类,其建筑群区域采用基于语义信息分析的区域提取方法,提高了这类复杂地物的区域一致性, 基于均值漂移过分割结果合并,边界也更精准。 Compare FIGS. 13 and 14 can be seen, the present invention is based on the MRF than H / a-Wishart classification, which complex region extraction method using the semantic information area based on the analysis, to improve the consistency of such complex feature region based on mean shift segmentation results over the merger, the border is also more accurate. 最后和基于Markov Random Field和极化信息的分类方法的融合提高了分类精度。 Finally, and fusion-based classification method Markov Random Field polarization information and the classification accuracy is improved.

[0100] 综上所述,本发明的基于语义信息和极化分解的极化SAR地物分类方法。 [0100] In summary, the present invention is the decomposition of semantic information and polarization polarimetric SAR classification based on the feature. 其实现包括:对span图进行均值漂移,提取span图的边脊草图,并在边脊草图中用基于语义信息的区域提取技术提取线段聚集区域;采用临界区域众数投票合并策略和基于极化特征合并策略对span图均值漂移过分割区域进行合并,得到分割结果;融合基于语义信息的图像分割结果和基于MRF的H/ a -Wishart分类结果,得到最终分类结果。 Which implement comprising: a span of mean shift diagrams, sketches span FIG ridge edge extraction, and the extraction area based on region extracting line segment aggregates semantic information in the sketch by side ridge; vote using the public number of the critical region and the polarization combined policies based wherein FIG combined mean shift strategy span over the divided regions were combined to obtain the segmentation result; fusion based on image segmentation based on the result of the semantic information and the MRF H / a -Wishart classification results to obtain a final classification result. 本发明将语义信息、图像处理技术和极化散射特性相结合,主要解决现有基于极化分解的分类技术对具有聚集特性地物的分类结果区域一致性较差的问题,提高了具有聚集特性地物(如森林、建筑群等)分类结果的区域一致性和边界保持性,克服了基于像素级分类的缺点,获得了良好的极化SAR 地物分类效果。 The present invention is semantic information, and image processing technology combined polarization scattering properties, mainly to solve the conventional classification techniques based on the Decomposition poor classification results aggregated region having characteristic feature of consistency, having increased aggregation properties and a boundary region similar feature (e.g., forest, buildings, etc.) retention of the classification result, overcome the shortcomings of classification based on the pixel level, to obtain a good polarization SAR feature classification results.

Claims (6)

1. 一种基于语义信息和极化分解的极化SAR地物分类方法,其特征在于:包括如下步骤: 步骤1.输入待分类的极化SAR图像的数据,对该极化SAR数据进行处理,得到极化SAR数据Ξ个通道的幅度值,融合Ξ个通道幅度值得到极化SAR图像的功率图,即span图,使用均值漂移得到span图的过分割结果图;并根据primesketch稀疏表示模型提取span图由线段组成的边脊草图,即SketchMap; 步骤2.对SketchMap中的线段进行语义信息分析,根据线段聚集特性的统计分布,对线段赋予语义信息即两侧聚集、单侧聚集和孤立线段; 步骤3.在SketchMap中,根据对线段赋予的语义信息,采用线段集合求解算法提取若干个不相交的聚集线段集合,并对每个聚集线段集合采用区域提取方法得到线段聚集区域R; 步骤4.对过分割结果进行区域合并:将线段聚集区域R对应的过分割区域采用临界区域众数投 A polarimetric SAR Classification Based on polarization and semantic information based on decomposition, characterized by: comprising the following steps: Step 1. polarimetric SAR image input data to be classified, the data is processed polarimetric SAR , to obtain an amplitude value polarimetric SAR data channels Ξ, Ξ channel amplitude fusion power is worth to FIG polarimetric SAR images, i.e. span FIG using mean-shift segmentation result obtained through the graph of FIG span; primesketch sparse representation of the model, and FIG extraction side ridges span sketch of line segments, i.e. SketchMap; semantic information analyzing step 2. SketchMap line segment in accordance with the statistical distribution of the aggregate characteristic line, i.e. on both sides of the line segment to impart semantic information aggregated, aggregation and unilateral isolation line; step 3. SketchMap, the semantic information given segment, using segment collection algorithm for extracting a plurality of sets of disjoint segments aggregates, and each aggregate segment region extraction method using a set of line segments obtained aggregate area R & lt; step 4. the results of over-segmentation region merging: the aggregate segment region R corresponding to the divided region through the critical region using a mode of administration 票合并策略;孤立线段所在过分割区域不合并;对于其他区域采用基于极化特征的区域合并策略,得到基于语义信息的极化SAR图像分割结果; 步骤5.利用极化分解对极化SAR数据进行H/α-Wishart分类,并用MarkovRandom Field对Η/α-Wishart分类结果进行邻域优化; 步骤6.融合基于语义信息的分割结果和基于MRF的H/α-Wishart分类结果,采用众数投票,对于分割图中的每个区域,选择对应的分类结果中像素个数最多的类别作为该区域的类别,并将该类别赋予最终分类结果图中对应的区域,得到待分类的极化SAR图像地物分类的最终分类结果图。 The combined policies votes; isolated segment located over the divided regions are not merged; Step 5 using the Decomposition polarimetric SAR data; polarimetric SAR image segmentation based on the results of the semantic information for other regions using region merging strategy based on the polarization characteristics, to give for H / α-Wishart classification, and for Η / α-Wishart classification results neighborhood optimization MarkovRandom field; step 6. fusion based segmentation result semantic information based on the MRF H / α-Wishart classification result using the public votes , for each divided region in the figure, corresponding to the selected classification result in the largest number of pixels as class category of the region, and the region to impart a final classification result of the category corresponding to FIG obtain polarimetric SAR images to be classified the final classification results map feature classification.
2. 根据权利要求1所述的基于语义信息和极化分解的极化SAR地物分类方法,其特征在于:其中步骤2中对线段进行语义信息分析,包括如下步骤: 2. 1根据不同地物类型对应的sketch线段的分布不同,得出线段对应于Ξ种地物信息:线目标,球形聚集分布的地物和地物之间的边界; 2. 2两线段之间的距离定义为线段中点的欧式距离,用线段K近邻的平均距离表示线段的聚集性;根据线段的聚集性的统计分布,将线段赋予语义信息:聚集线段和孤立线段; 根据聚集线段的拓扑结构分为两侧聚集和单侧聚集; 2. 3根据线段聚集性的统计分布,将线段的语义信息W树型结构表示,两侧聚集对应于森林、建筑群地物;单侧聚集对应于一边有森林、建筑群地物的边界;孤立线段对应于线目标,桥梁流形地物或两种地物的边界。 According to claim polarimetric SAR Classification Based on polarization and semantic information based on decomposition, characterized in that said 1: wherein the two pairs of line segments in the step of analyzing semantic information, comprising the steps of: 2.1 depending on the different sketch segment corresponding to the type of composition distribution, derived segments corresponding to object information Ξ farming: a boundary line between the targets and the feature distribution of spherical agglomerates of feature; defined distance between two segments as segments 2.2 Euclidean distance midpoint of a line segment represented by aggregation line average distance K neighbors; according to the statistical distribution of the aggregated segment, the semantic information given segment: segments and isolated segments aggregate; topology aggregation according to both sides of divided segment unilateral aggregation and aggregation; 2.3 the statistical distribution of the aggregated segment, the semantic information W represents a segment tree structure, corresponding to both sides of aggregation forest, buildings feature; unilateral corresponding side gathering forests, construction boundary of the feature group; isolated segments corresponding to the target line, the bridge manifold or both feature a feature boundary.
3. 根据权利要求2所述基于语义信息和极化分解的极化SAR地物分类方法,其特征在于:其中步骤3中在SketchMap中,根据对线段赋予的语义信息,采用线段集合求解算法提取聚集线段集合,并对聚集线段集合采用区域提取方法得到线段聚集区域R,包括: 3. 1符号定义定义:sketch线段集合为S;空间约束阔值δ1;线段生长阔值δ2;满足空间约束线段集合U;聚集线段集合 3. The SAR 2 Classification Based on polarization and polarization semantic information based on decomposition, according to claim wherein: in step 3 SketchMap wherein, the semantic information given segment, using segment collection algorithm extract segment collection aggregation, and aggregated segment collection region extraction method using a line segment obtained aggregation region R, comprising: symbol Definition 3.1: Sketch segment collection is S; wide space constraint value Delta] 1; line width growth values ​​of Delta] 2; segment meet space constraints collection U; gather segment collection
Figure CN103294792BC00021
;线段聚集区域R=h,。 ; Segment aggregate area R = h ,. ,…, rj ; 3. 2采用线段集合求解算法,得到若干个不相交的聚集线段集合Tk; 3. 3对每个聚集线段集合采用区域提取方法,得到线段聚集区域R。 , ..., rj; 3. 2 algorithm uses a set of line segments, to obtain a plurality of disjoint aggregated segment collection Tk; 3. 3 region extraction method using a set of aggregation for each segment, to obtain aggregated segment region R.
4. 根据权利要求3所述的基于语义信息和极化分解的极化SAR地物分类方法,其特征在于:其中步骤4中对过分割结果进行区域合并,得到基于语义信息的极化SAR图像分割结果;包括如下步骤: 4. 1对线段聚集区域对应的过分割区域采用临界区域众数投票合并策略:对于线段聚集区域和过分割区域的重叠情况有两种:对于过分割区域被线段聚集区域全部覆盖区域直接合并,对于过分割区域和线段聚集区域边界区域部分重叠情况采用临界区域众数投票合并策略,其中,将边缘的部分重叠区域叫做临界区域,对临界区域如果线段聚集区域占过分割区域的50%W上,则将运个过分割区域全部合并为线段聚集区域,否则,将其划分为无线段区域;最后在过分割图中得到合并的线段聚集区域灰; 4. 2对于孤立线段,提取其所在的过分割区域;对运些区域不进行合并,保 4. The polarimetric SAR Classification Based on polarization and semantic information based on decomposition, characterized in that said according to claim 3: wherein the step 4 of the segmentation results through region merging, to obtain semantic information polarimetric SAR images based on segmentation result; comprising the steps of: 4.1 pairs of line segments through the aggregation area corresponding to the number of divided regions critical region using all votes consolidation strategy: for line and the collecting area through the overlap region is divided in two ways: through the divided region to be aggregated segment areas cover all direct consolidation area, for over the divided region and segment aggregation area boundary region overlaps case of several critical areas congregation voted to merge strategy, which will be partially overlapping edge of the region called the critical region, the critical area if the line aggregation region accounts for over 50% W on the divided regions, the divided regions will be transported through a merging line segments all the collecting area, otherwise, the radio segment will be divided into areas; finally resulting merged segment ash collecting area through the segmentation map; 4.2 for isolated line segments extracted in the region over which it is divided; op these regions are not merged, Paul 留孤立线段所在区域; 4. 3对于其他区域,定义为无线段区域,采用基于极化特征的合并策略;首先将均值漂移的每个过分割块看作超像素,统计超像素的极化特性,采用Ξ通道灰度直方图统计作为特征,对于每个通道,将灰度值量化为16份,然后计算在运个特征空间的区域直方图;Ξ个通道共有16X3 = 48份;每个区域用一个48维的向量表示,对直方图特征进行归一化,根据化attacharyya系数计算公式,计算两个区域的相似性;设定合并阔值U,相似性大于阔值的相邻区域进行合并,合并后的区域再次计算直方图特征,迭代合并直到没有可合并的区域为止; 经过上述Ξ种合并策略,得到最终区域合并结果,即基于语义信息的极化SAR图像分害结果。 Remain isolated region segment is located; 4.3 for the other regions, defined as regions of a wireless section, combined using the strategy based on polarization characteristics; each division block through the first superpixel considered as the mean shift, the polarization characteristics of the statistical superpixel using a Cascade channel grayscale histogram as a feature, for each channel, the gradation value is quantized to 16 parts, and then calculates the operation region histogram feature space; a Cascade channels total parts 16X3 = 48; each region with a 48-dimensional vector representation of the normalized histogram feature, according to the calculation formula of attacharyya coefficient, calculating the similarity of the two regions; combined width setting value U, is greater than the width of the adjacent region of similarity values ​​are merged , the combined region feature histogram calculation again, until the combined area of ​​iterations until there is no combinable; Ξ species through the above strategies were combined, the results were combined to give the final area, i.e. sub-polarization SAR harm results based on semantic information.
5. 根据权利要求3所述基于语义信息和极化分解的极化SAR地物分类方法,其特征在于:线段集合求解过程包括: 3. 2. 1首先得到sketch线段集合S,依据森林、建筑群区域的线段具有聚集性,对每条线段的k近邻进行统计,计算每条线段的k近邻平均距离,从k近邻平均距离的直方图统计看出图像线段是否具有聚集性,如果具有某种聚集性,说明存在运样的地物,根据直方图统计,得到空间约束阔值δ1和线段生长的阔值δ2; 3. 2. 2初始设Ti为空集;根据空间约束阔值得到满足空间约束线段集合ί/=如典,…,随机选取种子线段A进行生长,此时,7;=如};生长的准则为,如果线段的某个近邻φ/满足线段生长阔值δ2,则生长为聚集线段集合7;=如,,遍历其k近邻直到没有可生长的线段,假设此时7:=如,,A,脚对此时Ti中没有遍历过的线段,依次作为种子线段进行生长, 5. The polarimetric SAR 3 Classification Based on polarization and semantic information based on decomposition, according to claim wherein: a set of line segments solving process comprising: 3 to give 2.1 sketch first segment set S, according to the forest, buildings line segment group having the aggregation region, the k-nearest neighbor each segment statistics to calculate the average nearest neighbor distance of each segment k, k-nearest neighbor seen from the histogram of the average distance image segment whether aggregation, if having a certain aggregation, indicating the presence of a sample transport feature, according to the histogram, a value obtained wide space constraints and line width growth values ​​δ1 δ2; 3. 2. 2 Ti initial set is the empty set; is met according to space constraints wide space segment collection constraints ί / = such as Code, ..., randomly selected seeds grown line a, this time, 7; as =}; guidelines for growth, if a neighbor segment φ / width value Delta] 2 satisfying the growth segment, the growth 7 is a set of aggregated segment; ,, = traverse segment as its k nearest neighbors until no growth, assuming 7: ,, a = as, Ti foot at this time is not a line segment traversed successively grown as a seed segment , 迭代生长直到所有生长进来的线段不能再生长为止,得到一个聚集线段集合Ti; 3. 2. 3若满足空间约束线段集合U中还有线段未进行生长,则选一条线段为种子线段继续迭代生长,直到所有的初始种子线段都得到生长;最后得到若干个不相交的聚集线段集合Tk。 Growth Growth iteration until all segments come up regrowth, to give a set of aggregated segment Ti; 3. 2. 3 meet space constraints if there are line segments set U is not grown, the seed is selected from a line segment growth continue iterating until all of the initial seed segments have been growing; and finally get some disjoint gathered segment collection Tk.
6. 根据权利要求3所述基于语义信息和极化分解的极化SAR地物分类方法,其特征在于:对聚集线段集合进行区域提取过程包括: 3. 3. 1圆形基元构造:取线段生长阔值δ2为圆的半径构造圆盘,即圆形基元; 3. 3. 2闭操作:使用结构元素Β对集合A进行闭操作,结构元素Β为构造的圆形基元, 集合A是得到的聚集线段集合;经过闭操作得到线段聚集区域R。 6. The 3 polarimetric SAR Classification Based on polarization and semantic information based on decomposition, according to claim wherein: region extraction process to aggregate set of line segments comprises: 3.3.1 circular configuration primitives: Take growth line width value δ2 radius of the circle disc configuration, i.e. circular primitives; 3.3.2 closing operations: using the structural elements of the set a Β for closing operations, structural elements Β element is constructed of a circular base, a collection of a segment is a collection of aggregates obtained; closing operation obtained through line aggregation region R.
CN201310192057.1A 2013-05-13 2013-05-13 Based on the semantic information and the Decomposition Classification Based polarization sar CN103294792B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201310192057.1A CN103294792B (en) 2013-05-13 2013-05-13 Based on the semantic information and the Decomposition Classification Based polarization sar

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201310192057.1A CN103294792B (en) 2013-05-13 2013-05-13 Based on the semantic information and the Decomposition Classification Based polarization sar

Publications (2)

Publication Number Publication Date
CN103294792A CN103294792A (en) 2013-09-11
CN103294792B true CN103294792B (en) 2016-02-10

Family

ID=49095654

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201310192057.1A CN103294792B (en) 2013-05-13 2013-05-13 Based on the semantic information and the Decomposition Classification Based polarization sar

Country Status (1)

Country Link
CN (1) CN103294792B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103675818B (en) * 2013-12-03 2017-01-18 中国科学院深圳先进技术研究院 Method and system for measuring polarimetric synthetic aperture radar data ground object three-dimensional distance
CN104123555B (en) * 2014-02-24 2017-05-24 西安电子科技大学 Super-pixel polarimetric SAR land feature classification method based on sparse representation
CN104463882B (en) * 2014-12-15 2017-07-28 西安电子科技大学 SAR image segmentation method based on shape completion administrative division map and feature coding
CN104809690B (en) * 2015-04-10 2017-10-27 西北工业大学 A kind of intelligent method for fusing of many sketches
CN107944370B (en) * 2017-11-17 2019-07-02 西安电子科技大学 Classification of Polarimetric SAR Image method based on DCCGAN model
CN109740475A (en) * 2018-12-25 2019-05-10 杭州世平信息科技有限公司 A kind of remote sensing images ground scene classification method

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102622598A (en) * 2012-01-13 2012-08-01 西安电子科技大学 SAR (Synthesized Aperture Radar) image target detection method based on zone markers and grey statistics
CN102708167A (en) * 2012-04-27 2012-10-03 上海交通大学 Web-based semantic annotation system and Web-based semantic annotation method for high resolution SAR (synthetic aperture radar) image interpretation

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7006656B2 (en) * 2001-10-15 2006-02-28 The Research Foundation Of Suny Lossless embedding of data in digital objects

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102622598A (en) * 2012-01-13 2012-08-01 西安电子科技大学 SAR (Synthesized Aperture Radar) image target detection method based on zone markers and grey statistics
CN102708167A (en) * 2012-04-27 2012-10-03 上海交通大学 Web-based semantic annotation system and Web-based semantic annotation method for high resolution SAR (synthetic aperture radar) image interpretation

Also Published As

Publication number Publication date
CN103294792A (en) 2013-09-11

Similar Documents

Publication Publication Date Title
Yu et al. Object-based detailed vegetation classification with airborne high spatial resolution remote sensing imagery
Fauvel et al. A spatial–spectral kernel-based approach for the classification of remote-sensing images
Huang et al. An adaptive mean-shift analysis approach for object extraction and classification from urban hyperspectral imagery
Cao et al. A SVM-based method to extract urban areas from DMSP-OLS and SPOT VGT data
Ghamisi et al. Automatic spectral–spatial classification framework based on attribute profiles and supervised feature extraction
CN102903110B (en) A method of dividing an image having a depth image information
Li et al. FoSA: F* seed-growing approach for crack-line detection from pavement images
Tso et al. A contextual classification scheme based on MRF model with improved parameter estimation and multiscale fuzzy line process
Mirzapour et al. Improving hyperspectral image classification by combining spectral, texture, and shape features
CN102096825B (en) Graph-based semi-supervised high-spectral remote sensing image classification method
Chen et al. Spectral–spatial preprocessing using multihypothesis prediction for noise-robust hyperspectral image classification
Celik A Bayesian approach to unsupervised multiscale change detection in synthetic aperture radar images
Li et al. A shape-based approach to change detection of lakes using time series remote sensing images
Wang et al. A review of road extraction from remote sensing images
Ban et al. Object-based fusion of multitemporal multiangle ENVISAT ASAR and HJ-1B multispectral data for urban land-cover mapping
US9117105B2 (en) 3D face recognition method based on intermediate frequency information in geometric image
Li et al. Object-oriented classification of high-resolution remote sensing imagery based on an improved colour structure code and a support vector machine
Zhong et al. A support vector conditional random fields classifier with a Mahalanobis distance boundary constraint for high spatial resolution remote sensing imagery
Chen et al. A new process for the segmentation of high resolution remote sensing imagery
Song et al. Shadow detection and reconstruction in high-resolution satellite images via morphological filtering and example-based learning
Gaetano et al. Marker-controlled watershed-based segmentation of multiresolution remote sensing images
Malinverni et al. Hybrid object-based approach for land use/land cover mapping using high spatial resolution imagery
CN101770584B (en) Extraction method for identification characteristic of high spectrum remote sensing data
Chen et al. Single image shadow detection and removal based on feature fusion and multiple dictionary learning
CN101359365B (en) Iris positioning method based on Maximum between-Cluster Variance and gray scale information

Legal Events

Date Code Title Description
C06 Publication
C10 Entry into substantive examination
C14 Grant of patent or utility model