CN103294792B - Based on the polarization SAR terrain classification method of semantic information and polarization decomposing - Google Patents

Abstract

The invention discloses a polarimetric SAR ground object classification method based on semantic information and polarization decomposition. Its implementation includes: performing mean shift on the span graph, extracting the edge and ridge sketch of the span graph, and using the region extraction technology based on semantic information to extract the line segment aggregation area in the edge and ridge sketch; The strategy and the polarization-based feature merging strategy merge the mean shifted over-segmented regions of the span image to obtain the image segmentation result; fuse the image segmentation result based on semantic information and the H/α-Wishart classification result based on MRF to obtain the final classification result. The invention combines semantic information, image processing technology and polarization scattering characteristics, and mainly solves the problem of regional consistency of the classification results of the existing classification technology based on polarization decomposition for objects with aggregation characteristics (such as forests, building groups, etc.). It improves the regional consistency and boundary preservation of the classification results of the features with aggregation characteristics.

Description

Polarimetric SAR Object Classification Method Based on Semantic Information and Polarization Decomposition

technical field

The invention belongs to the technical field of image processing and remote sensing, and relates to the classification of ground objects in polarimetric SAR images, in particular to a classification method of polarimetric SAR ground objects based on semantic information and polarization decomposition, which can be used for low-level ground objects with aggregation characteristics. Resolving ground object classification in polarimetric SAR images.

Background technique

Polarimetric Synthetic Aperture Radar (POLSAR) image processing is an important subject of national defense construction and economic development, and has attracted more and more attention and research. Compared with ordinary single-polarization synthetic aperture radar (Synthetic Aperture Radar, SAR), polarimetric SAR performs full-polarization measurement, which can obtain more abundant object information of the target, and provides important information for more in-depth research on the scattering characteristics of the target. basis. Polarimetric SAR object classification is one of the important tasks of polarimetric SAR image processing, and it is the premise of polarimetric SAR image interpretation. The key and difficulty of polarimetric SAR segmentation or object classification lies in the regional consistency of the same object and the boundary preservation between different objects.

There are many classification methods for polarimetric SAR ground features, which are mainly divided into three types: 1) classification methods based on statistical models; 2) classification methods based on electromagnetic wave scattering mechanism; 3) classification methods based on image processing technology. The methods based on the statistical model mainly include: Lee et al. According to the polarization covariance matrix satisfying the complex wishart distribution, a supervised polarization SAR classification is proposed. But in practical applications, there is very little prior knowledge about the categories of SAR images. There are many methods based on the scattering mechanism of electromagnetic waves. In 1997, Cloude et al. first proposed the H/α classification method. Through decomposition, the scattering entropy H of the ground object and the angle α representing the scattering mechanism of the ground object were obtained, and the unsupervised polarization was realized. SAR image classification. In 1999, Lee et al. introduced the Wishart classifier based on the H/α classification method combined with the statistical distribution, and improved the classification accuracy by performing Wishart iterations on the results of the H/α classification method. In 2004, Lee et al. proposed a classification method that maintains the characteristics of polarization scattering. This method uses the power of the three polarization scattering mechanism components obtained by Freeman decomposition for initial classification, and merges and classifies through Wishart iteration. A better classification effect is achieved.

The above method makes good use of the scattering characteristics and polarization information of polarimetric SAR data for classification, but this pixel-based classification method does not consider the visual characteristics of polarimetric SAR images, and does not combine computer vision methods and image processing methods. method to classify. Therefore, there are many defects in the traditional polarimetric SAR ground object classification method including the above method: (1) the regional consistency of the same ground object is not good, resulting in a classification result map of salt and pepper noise; (2) based on the traditional image processing method The polarimetric SAR ground object classification method, for the ground objects with light and dark gray scale changes, such as the traditional classification method based on pixel point and super pixel combination, it is difficult to classify such ground objects into one category; (3) for Complex ground features, such as building groups, since the ground features themselves contain houses, roads, etc., the scattering characteristics of the ground features are not single, and have light and dark ground features, so it is difficult to divide them into a complete area. Extracting various underlying features and using various methods of region merging are difficult to group these regions together, but for the classification of low-resolution polarimetric SAR images, it should be divided into one from the perspective of human vision and image understanding. kind. Therefore, the extraction of low-level features is already difficult to classify such ground objects together, and the advanced features based on the characteristics of ground objects need to be further excavated for classification.

To sum up, the pixel classification methods of the above-mentioned polarization SAR ground object classification methods are fine, but there are still some defects, especially for the ground objects with aggregation characteristics (such as buildings, forests, etc.). Single, with light and dark ground object scattering characteristics, the consistency of the classification area is poor, and the boundary is easily affected by noise, and it is easy to produce salt-and-pepper classification results.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the above-mentioned existing methods, and propose a polarization SAR feature classification method based on semantic information and polarization decomposition. This method has good regional consistency and precise boundaries for features with aggregation characteristics. The result of classification improves the effect of polarimetric SAR object classification.

The present invention is a polarimetric SAR ground feature classification method based on semantic information and polarization decomposition. It performs unsupervised ground feature classification on low-resolution polarimetric SAR images obtained in advance. The classification process includes the following steps:

Step 1. Input the data of the polarimetric SAR image to be classified, process the polarimetric SAR data, obtain the amplitude values of the three channels of the polarimetric SAR data, and fuse the amplitude values of the three channels to obtain the backscatter of the polarimetric SAR image The total power map, that is, the span map, uses the mean shift to obtain the over-segmentation result map of the span map; and extracts the edge and ridge sketch of the span map composed of line segments according to the initial sketch (primesketch) sparse representation model, that is, SketchMap.

Step 2. Analyze the semantic information of the line segments in SketchMap, and assign semantic information to the line segments according to the statistical distribution of the line segment aggregation characteristics, that is, two-sided aggregation, one-sided aggregation and isolated line segments.

Step 3. In SketchMap, according to the semantic information assigned to the line segment, use the line segment set solving algorithm to extract several disjoint aggregated line segment sets, and use the region extraction method for each aggregated line segment set to obtain the line segment aggregation area R.

Step 4. Merge regions for the over-segmentation results: For the over-segmentation result map of the span graph obtained in step 1, use the critical region majority vote merging strategy for the over-segmentation region corresponding to the line segment aggregation region R; extract the over-segmentation region where the isolated line segment is located , using the non-merging strategy; for the other regions, that is, the remaining regions, using the region merging strategy based on polarization features to obtain the polarization SAR image segmentation results based on semantic information.

Step 5. Use polarization decomposition to perform H/α-Wishart classification on polarimetric SAR data, and use Markov Random Field (MRF) to perform neighborhood optimization on H/α-Wishart classification results.

Step 6. Fuse the polarization SAR image segmentation result based on semantic information and the H/α-Wishart classification result based on MRF through the majority vote strategy to obtain the final classification of the polarization SAR image object classification to be classified result.

The key technology for realizing the present invention is: aiming at the problem of poor regional consistency in the classification of clustered features (such as building groups, forests, etc.), the analysis shows that low-resolution polarimetric SAR images generally include farmland, urban areas, forests, etc. , mountains, bridges, etc. According to human prior knowledge, it can be known that the structural line segments of the building group should be clustered and distributed in a spherical shape, and the structural line segments of bridges should be linearly distributed, etc., taking these cognitions as prior knowledge, the semantic information contained in the line segments Analyze and endow the line segment with semantic information. By analyzing the semantic information of the line segment, the line segment aggregation area can be extracted, which corresponds to the buildings, forests and other features in the image. By extracting the line segment aggregation area, the consistent area of these features can be obtained. According to the semantic information analysis of the line segment, we can Divide the over-segmented image into line segment aggregation area, isolated line segment area and wireless segment area. The line segment aggregation area corresponds to buildings and other ground objects, the isolated line segment area corresponds to line targets, etc., and the wireless segment area generally corresponds to oceans, farmland, etc. Ground features, the present invention adopts different merging strategies for different regions, and adopts more targeted merging strategies for different types of ground features, so that various ground features can be better merged, and finally the segmentation and classification results are fused, The semantic information and polarization information are organically combined to obtain classification results with good regional consistency and precise boundaries, which solves the problem of poor consistency of classification regions with clustered features.

Compared with the prior art, the present invention has the following advantages:

1. From the analysis of the semantic information, the present invention utilizes the PrimalSketch sparse representation model to obtain the SketchMap of the span graph, analyzes the semantic information contained in the line segment according to the SketchMap, and proposes a region division technology based on the line segment semantic information analysis, which is effective on the SketchMap In order to extract the clustered area of line segments. These line segment gathering areas correspond to urban areas, forests and other ground objects in polarimetric SAR images. These ground features are often classified into multiple categories due to the gray scale changes between light and dark. The present invention overcomes this shortcoming and effectively improves the regional consistency of the classification of line segments gathering areas.

2. From the perspective of image processing technology, when the mean value drifts over the segmentation area for merging, the present invention adopts different merging strategies for different types of feature areas, and the area merging is more targeted, ensuring that different feature areas can be compared. Well merged, a segmentation result based on semantic information is obtained.

3. From the perspective of polarization decomposition, the present invention uses H/α-Wishart classification, and uses MRF to perform neighborhood optimization to obtain pixel-level classification results, and finally fuses the segmentation and classification results, using the segmentation region to guide the regional consistency of the classification, and at the same time The classification results also help the further merging of the segmented regions, and the interaction of segmentation and classification leads to better classification results. The present invention combines image processing technology and technology based on electromagnetic wave scattering mechanism, integrates semantic information and polarization information, combines semantic information, image processing technology and polarization scattering characteristics, and improves the accuracy of polarization SAR ground object classification results. Regional consistency and boundary preservation.

Description of drawings

Fig. 1 is the flow chart of the present invention to the classification of polarimetric SAR data features;

Fig. 2 is the span diagram of the full polarization SanFrancisco data of the NASA/JPLAIRSARL band that the present invention uses;

Fig. 3 is the over-segmentation result figure that mean drift obtains among the present invention;

Fig. 4 is the side ridge sketch that adopts the present invention to obtain, i.e. SketchMap

Fig. 5 is the semantic information tree structure diagram of line segment in the present invention;

Fig. 6 is a sketch of edge and ridge endowed with semantic information obtained by adopting the present invention;

Fig. 7 is a schematic diagram of the extraction process of the line segment aggregation area in the present invention;

Fig. 8 is a diagram of the extraction result of line segment aggregation area based on semantic information analysis in the present invention;

Fig. 9 is a merging result diagram of an over-segmented area corresponding to a line segment gathering area in the present invention;

Fig. 10 is an image segmentation result diagram based on semantic information in the present invention;

Fig. 11 is a schematic diagram of the fusion process of segmentation and classification results in the present invention;

Fig. 12 is the span diagram of the fully polarized SanFrancisco data of the NASA/JPLAIRSARL band used by the present invention;

Fig. 13 is the H/α-Wishart classification result figure based on MRF in the present invention;

Fig. 14 is a diagram of classification results of the present invention.

detailed description

Example 1

The present invention is a polarimetric SAR object classification method based on semantic information and polarization decomposition, and performs unsupervised object classification for low-resolution polarimetric SAR images obtained in advance. Referring to FIG. 1, the implementation steps of the classification process of the present invention include:

Step 1, input the data of the polarimetric SAR image to be classified, process the polarimetric SAR data, obtain the amplitude values of the three channels of the polarimetric SAR data, and fuse the amplitude values of the three channels to obtain the backscatter of the polarimetric SAR image The total power map, as shown in Figure 2, is the span map of the fully polarized SanFrancisco data in the NASA/JPLAIRSARL band. Use the mean shift on the span graph to obtain the over-segmentation result graph of the span graph; and extract the edge and ridge sketch of the span graph composed of line segments according to the primesketch sparse representation model, that is, SketchMap.

First, the polarimetric SAR data is processed to obtain the covariance matrix, and the amplitude values of the three channels are obtained according to the three values of the diagonal elements of the covariance matrix, and the span image of the polarimetric SAR image is obtained by fusing the amplitude values of the three channels. The first operation on the span graph is to use the mean shift to obtain the over-segmentation result graph of the span graph, as shown in Figure 3.

The second operation is to extract SketchMap using the edge-ridge detection sparse coding method, and the extraction steps include:

Firstly, Gaussian first-order guide filters and Gaussian second-order guide filters with N scales and M directions are constructed to form a filter bank. The value of N is 3, and the value of M is 18. As shown in Figure 2, the span image is convolved with the filter bank to obtain the joint response of each pixel, and the maximum value of the joint response is extracted as the edge/ridge strength of the pixel, and the maximum response filter direction as the local direction for that pixel. Apply non-maximum suppression to the edge/ridge intensity map to get the proposed sketch According to the suggested sketch The location of the maximum joint response in the proposed sketch Connect the points connected with this position into a line segment to generate an edge/ridge original model S _{sk, 0} ;

Secondly, add a new line segment to the edge-ridge model, evaluate the encoding length gain ΔL of the image, if ΔL<ε, ε is the threshold value, and the value is 10, then reject the line segment, otherwise accept it, and search, and the sketch will be suggested The dividing line between the end of the new line segment and the remaining pixels within the average fitting error is used as the next new suggested line segment. If there is a new suggested line segment, recalculate the image encoding length gain ΔL after adding the new suggested line segment. If ΔL<ε Then refuse to accept the new suggested line segment, otherwise accept the new suggested line segment, and iteratively add new line segments until there is no new suggested line segment, then the edge and ridge sketch is obtained, as shown in Figure 4, which is SketchMap.

Step 2, analyze the semantic information of the line segments in SketchMap, and assign semantic information to the line segments according to the statistical distribution of the line segment aggregation characteristics, that is, two-sided aggregation, one-sided aggregation and isolated line segments.

2.1 For low-resolution polarimetric SAR images, for ground objects with aggregation characteristics, taking building groups as an example, the line segments are formed by bright buildings and dark ground. Such structures appear repeatedly, forming building groups , the corresponding sketch line segment is usually characterized by dense distribution, and the direction of the line segment is mostly approximately horizontal and vertical. For forest features, the sketch line segments are also densely distributed, but the direction of the line segments is disorderly. For bridges, the sketch line segments are distributed in a manifold, etc. Therefore, the distribution structure of line segments contains certain semantic information. According to the distribution of sketch line segments corresponding to different types of features, it can be concluded that line segments mainly correspond to three types of feature information: line targets, spherically aggregated features and different ground features. border between things.

2.2 The distance between two line segments is defined as the Euclidean distance of the midpoint of the line segment, and the average distance of the K nearest neighbors of the line segment is used to indicate the degree of aggregation of the line segment; according to the statistical distribution of the aggregation of the line segment, the line segment is given semantic information: aggregated line segment and isolated line segment ;According to the topological structure of the aggregated line segment, it can be divided into two-sided aggregation and one-sided aggregation.

2.3 According to the statistical distribution of the aggregation of line segments, the semantic information of line segments is represented in a tree structure, as shown in Figure 5, which is a schematic diagram of the tree structure of semantic information of line segments. Two-sided aggregation corresponds to forests, buildings and other features; one-sided aggregation corresponds to the boundary of one side with forests or buildings and other features; isolated line segment corresponds to line targets, bridges and other manifold features or two different features boundary. Figure 6 shows the SketchMap endowed with line segment semantic information, where the gray line segments are aggregated line segments, and the black line segments are isolated line segments.

According to the statistical distribution of line segment aggregation characteristics, the present invention assigns semantic information to line segments, and the endowed semantic information includes two-side aggregation, one-side aggregation, and isolated line segments. Provides a basis for regional division.

Step 3. In SketchMap, according to the semantic information assigned to the line segment, use the line segment set solving algorithm to extract several disjoint aggregated line segment sets, and use the area extraction method for each aggregated line segment set to obtain the line segment aggregation area R.

3.1 Symbol definition: the set of sketch line segments is S; the space constraint threshold δ ₁ ; the line segment growth threshold δ ₂ ; the set of line segments satisfying the space constraints U; the collection of line segments Line segment gathering area R={r ₁ , r ₂ ,..., r _m };

3.2 Firstly, the line segment set solution algorithm is adopted. This algorithm is similar to the method of region growth, but the present invention uses the line segment as the primitive for growth, and obtains the aggregated line segment set, which is beneficial to the extraction of the line segment aggregation area. The specific steps are as follows:

3.2.1 Firstly, get the sketch set S. According to the aggregation of the line segments in the line segment aggregation areas such as forests and buildings, the k nearest neighbors of each line segment are counted, and the k nearest neighbor average distance of each line segment is calculated. From the k nearest neighbor average distance Histogram statistics show whether the image line segment has aggregation. If there is a certain aggregation, it means that there is such a ground object. According to the histogram statistics, the threshold of space constraint δ ₁ and the threshold of line segment growth δ ₂ are obtained.

3.2.2 Initially set T _i as an empty set; get the initial seed line segment according to the threshold of the seed line segment Randomly select a seed segment to grow, at this time, The growth criterion is that if a neighbor of the line segment Satisfy the line segment growth threshold δ ₂ , then grow into a set of aggregated line segments Traverse its k-nearest neighbors until there is no line segment that can be grown, assuming that at this time For the line segments that have not been traversed in T _i at this time, they are grown sequentially as seed line segments, and iteratively grow until all grown line segments can no longer be grown. At this time, a collection of line segments T _i is obtained.

3.2.3 If there are still line segments in the initial seed line segment set U that have not been grown, select a line segment as the seed line segment to continue growing, and iteratively grow until all initial seed line segments are grown. Finally, several disjoint line segment sets T _k are obtained.

3.3 The area extraction method is adopted for each aggregated line segment set: on the basis of the line segment set, the area is extracted with circular primitives to obtain the area where the aggregated line segment set is located.

3.3.1 Circular primitive construction: take the line segment growth threshold δ ₂ as the radius of the circle to construct a disc. A circle is used to maintain the smoothness of the region boundary, and the radius _δ2 is used to ensure the maximum gap between filled line segments. Because the line segment intervals in the same line segment aggregation area should be similar, and the growth threshold δ ₂ represents the maximum line segment interval of the grown line segment set, so here we take δ ₂ as the disk radius.

3.3.2 Closing operation: using structural element B to close the set A, denoted as A·B, defined as

in, Indicates that B performs an expansion operation on A, Indicates that B performs a corrosion operation on A.

This formula shows that the closed operation of using the structural element B to A is to use B to expand A, and then use B to corrode the result. Fig. 7 is a schematic diagram of the extraction process of the line segment aggregation area in the present invention. In Fig. 7(a), the structural element B is the circular primitive constructed above, and the set A is a set composed of line segments. Dilation of set A refers to using structure B to move every point on the line segment in image A, and the set of all displacements is the result after dilation. The dilation operation is shown in Figure 7(b), and the dilation result is shown in Figure 7(c). Erosion operation is performed after dilation. The corrosion operation is shown in Fig. 7(d), and the final result of closing operation is shown in Fig. 7(e). It can be seen from the figure that the closing operation obtains the area where the line segment set A is located, eliminates the narrow and long slits, and obtains a consistent connected area. Region extraction is performed on each collection of line segments to obtain a line segment collection area R. Figure 8 shows the results of the line segment collection area extraction.

Step 4. Merge the over-segmentation results: the over-segmentation area corresponding to the line segment aggregation area R adopts the critical area majority vote merging strategy; extracts the over-segmentation area where the isolated line segment is located, adopts the non-merging strategy; for other areas, adopts the polarization-based The regional merging strategy of features is used to obtain the segmentation results of polarimetric SAR images.

4.1 The over-segmentation area corresponding to the line segment aggregation area adopts the critical area majority voting merge strategy: because the area consistency of the line segment aggregation area is good, but the boundary is not accurate, and the boundary of the over-segmentation is accurate, therefore, for the boundary and over-segmentation of the line segment aggregation area If the boundary of the segmentation area does not match, the critical area majority voting merge strategy is adopted; there are two overlapping situations between the line segment aggregation area and the over-segmentation area: one is that some over-segmentation areas are completely covered by the line segment aggregation area; the other is the line segment aggregation area The edge area and the over-segmentation area partially overlap, and the edge overlapping area is called the critical area. For the first case, the mean shift over-segmentation area is directly merged. For the second case, according to the majority voting strategy, if the line segment aggregation area accounts for more than 50% of the over-segmentation area, all the over-segmentation areas are merged into line segment aggregation area, otherwise, it is divided into wireless segment areas; finally, the merged line segment aggregation area is obtained in the over-segmentation graph This ensures that these hard-to-merge over-segmented regions are well merged. Figure 9 shows the results of the merged line segment aggregation area, and it can be seen that the line segment aggregation area of the building group has been well merged.

4.2 For the isolated line segment, extract the over-segmented area where it is located. These regions are not merged. According to the semantic information analysis of the line segment, for the isolated line segment corresponding to the line target in the image or the boundary of two ground objects, when merging the area, if the area where the isolated line segment is located will be merged, the line target will disappear, or the two Merge of different regions. Therefore, the present invention does not perform region merging on the region where the isolated line segment is located.

4.3 For other areas, it is defined as a wireless segment area, and a merging strategy based on polarization characteristics is adopted. First, each over-segmented area obtained by mean shift is regarded as a superpixel, and the polarization characteristics of the superpixel are counted, and the three-channel gray histogram statistics are used as features. For each channel, the gray value is quantized into 16 parts, and then Computes the region histogram in this feature space. There are 16*3=48 copies in total for the three channels. Each region can be represented by a 48-dimensional vector, such as using Hist _p to represent the normalized histogram feature of region P.

According to the Bhattacharyya coefficient calculation formula, calculate the similarity ρ(P, Q) of two regions P and Q, and ρ(P, Q) is defined as follows:

$\mathrm{<span\; class="notranslate">\; \&rho;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 48</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\sqrt{{\mathrm{<span\; class="notranslate">\; Hist</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}^{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; Hist</span>}}_{\mathrm{<span\; class="notranslate">\; Q</span>}}^{\mathrm{<span\; class="notranslate">\; u</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, Hist _P and Hist _Q are the normalized histograms of R and Q, respectively. The superscript u denotes the uth component of the histogram.

Set the merging threshold U, and merge the adjacent regions whose similarity is greater than the threshold. The histogram features of the merged regions are calculated again, and iteratively merged until there is no region that can be merged, and the segmentation result based on semantic information is obtained. Figure 10 shows the segmentation results based on semantic information.

The present invention not only proposes a method for extracting line segment aggregation areas based on semantic information to extract line segment aggregation areas on edge and ridge sketches, but also uses different strategies to merge over-segmented areas: for line segment aggregation areas, the critical area majority voting strategy is used to guide over-segmentation The area of the block is merged; for the over-segmented area where the isolated line segment is located, the non-merging strategy is adopted; the remaining area is the wireless segment area, and the area merging strategy based on polarization information is adopted. The invention combines the semantic information to extract the line-segment aggregation area, adopts different merging strategies for different types of areas, and well solves the problem of difficult classification of the line-segment aggregation area.

Step 5: Use polarization decomposition to perform H/α-Wishart classification on polarimetric SAR data, and use MarkovRandomField to perform neighborhood optimization on H/α-Wishart classification results.

5.1 Use the H/α-Wishart classification method to obtain the initial classification results where S is a set of pixels. The Wishart distance uses the distance measure based on the wishart distribution modified by Kersten et al. Each pixel label in l ^[0] L is the total number of categories. Here L=8.

5.2 Given a set of observations O={T _s |s∈S}, where T _s is the polarization coherence matrix of pixel s. It is known that the covariance matrix obeys the complex wishart distribution. According to the initial classification results, use the observation samples of the i-th class To estimate the distribution parameter σ of the class, and calculate the L×L inter-class distance matrix D:

${\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ]</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ]</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

where D _ij represents the distance between class i and class j, and d represents the Euclidean distance of the average coherence matrix.

5.3 In the MRF-based framework, the data item is the similarity value of each pixel, and the smoothing item is the distance between classes. The minimized energy function is as follows:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; ln</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\underset{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; D.</span>}}_{{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

in, is the class conditional probability of the observed data at pixel s, and N _s is the set of neighboring pixels of pixel s. λ ₁ is a regularization parameter. The total energy in (3) is minimized by the α-expansion algorithm.

Step 6, fusing the segmentation results based on semantic information and the H/α-Wishart classification results based on MRF.

The present invention combines the advantages of the regional consistency of the segmentation results and the pixel-level accuracy of the classification results to obtain better classification results. This fusion strategy combines unsupervised segmentation and pixel-based classification results, and classifies based on the majorityvote strategy to obtain the final classification result of the polarization SAR image object classification to be classified. Figure 11 is a schematic diagram of the fusion process of segmentation and classification results, the main steps of which include:

6.1 Segmentation: To obtain consistent regions, the number of regions should be greater than the number of final categories, and slightly higher than the number of categories; Figure 11(a) is a schematic diagram of the segmentation of 4 segmentation regions, where 1 to 4 are used to represent the 4 segmentation regions;

6.2 Pixel-based classification: Pixel-level classification is performed based on the scattering characteristics of the image. Figure 11(b) is a schematic diagram of pixel-based classification, in which white, black and gray represent three categories.

6.3 Fusion segmentation and classification: Using the majorityvote strategy, for each region in the segmentation map, select the category with the largest number of pixels in the corresponding classification result as the category of this region, and mark the corresponding region of the final classification result map as this category. This greatly improves the regional consistency of the classification results. It should be noted that in majorityvote, the neighborhood of pixels is not a fixed neighborhood window, but pixels belonging to the same area are segmented. Figure 11(c) is a schematic diagram of the fusion of the segmentation map and the classification result based on the pixel points, and the majority voting strategy is adopted for each region in the picture, and the classification result shown in Figure 11(d) is obtained. After fusing the segmentation and classification results, the final classification result map of the polarimetric SAR image object classification to be classified is obtained, as shown in Figure 14.

The present invention uses the PrimalSketch sparse representation model to obtain the SketchMap of the span graph, analyzes the semantic information contained in the line segment according to the SketchMap, proposes a line segment aggregation area extraction technology based on the line segment semantic information analysis, and effectively extracts the line segment aggregation area on the SketchMap. These line segment gathering areas correspond to urban areas, forests and other ground objects in polarimetric SAR images. These ground features are often classified into multiple categories due to the gray scale changes between light and dark. The present invention overcomes this shortcoming and effectively improves the regional consistency of the classification of line segments gathering areas. At the same time, in order to maintain the polarization scattering characteristics, the H/α-Wishart classification is performed on the polarization SAR data, and the neighborhood optimization is carried out with MRF. The classification result based on polarization decomposition is fine, but there are many noise points. Therefore, the present invention fuses the segmentation result and the H/α-Wishart classification result based on MRF to obtain the ground object classification result of the polarization SAR image to be classified. The final classification result is obtained by effectively fusing semantic information and polarization decomposition.

Example 2

The classification method of polarimetric SAR ground features based on semantic information and polarization decomposition is the same as that in embodiment 1, and the simulation data and images are described as follows:

1. Simulation conditions

(1) Select the full polarization SanFrancisco data of NASA/JPLAIRSARL band;

(2) In the simulation experiment, the value of the parameter N in the PrimalSketch sparse representation model is 3, the value of M is 18, and the value of the threshold ε is 20;

(3) In the simulation experiment, the number of neighbors k is 9;

(4) In the simulation experiment, the seed line segment threshold δ ₁ is taken as 20; the line segment growth threshold δ ₂ is taken as 12;

(5) In the simulation experiment, the region merging threshold U is set to 0.7;

(6) In the simulation experiment, the neighborhood window is selected as 3*3 in the MRF-based H/α-Wishart classification.

2. Simulation content and results

Utilize the full-polarization SanFrancisco data of NASA/JPLAIRSARL band, use the present invention to classify it. Fig. 12 is a span diagram, which is the same picture as Fig. 2. For the convenience of evaluating the classification results, Fig. 12, Fig. 13, and Fig. 14 are displayed together, and Fig. 14 is a classification result diagram of the present invention. It can be seen from the figure that the regional consistency of the classification results of the present invention is good and the boundary part is also relatively accurate, especially for the building group area, a large consistent area can be obtained, which is more in line with the understanding of human vision for images, and the bridge is more accurate. Line target, the strategy of the present invention can obtain better classification results, and bridges can be well separated. To sum up, due to the addition of semantic information, the present invention can obtain classification results that are more suitable for human beings to understand images, and the regional consistency and edge accuracy of ground objects have been improved.

Example 3

The polarimetric SAR ground object classification method based on semantic information and polarization decomposition is the same as embodiment 1-2, wherein the H/α-Wishart classification method based on MRF is the same as step 5 in embodiment 1, as a comparative experiment of the present invention, simulation The data and results are as follows:

1. Simulation conditions

(1) Select the full polarization SanFrancisco data of NASA/JPLAIRSARL band;

(2) In the simulation experiment, the neighborhood window selection in the MRF-based H/α-Wishart classification is 3*3.

2. Simulation content and results

Using the full-polarization SanFrancisco data of NASA/JPLAIRSARL band, the H/α-Wishart classification method based on MRF is used for classification. This method is a classification method based on pixels. Results of the α-Wishart classification method. It can be seen from the figure that this method is fine in classification, but it produces salt-and-pepper classification results, especially for building groups, which have aggregation characteristics. Since they include buildings and roads, their scattering types are inconsistent, so Inconsistent classification results are produced, but for low-resolution polarimetric SAR images, we hope to obtain consistent building group classification results when performing image understanding. Therefore, this method has poor consistency in the classification of ground features with aggregation characteristics. Boundaries are also susceptible to noise.

The present invention compares with the result of the H/α-Wishart classification method based on MRF:

The present invention is compared with the ground object classification result of the MRF-based H/α-Wishart classification. The experimental results are as follows, Fig. 12 is a span diagram, Fig. 13 is a result diagram of H/α-Wishart classification based on MRF, and Fig. 14 is a classification result diagram of the present invention. Comparing Figure 13 and Figure 14, it can be seen that the present invention is based on the H/α-Wishart classification based on MRF, and its building group area adopts the area extraction method based on semantic information analysis, which improves the regional consistency of such complex features. The mean shift merges the segmentation results, and the boundaries are more precise. Finally, the fusion with the classification method based on MarkovRandomField and polarization information improves the classification accuracy.

To sum up, the method for classifying polarimetric SAR ground features based on semantic information and polarization decomposition of the present invention. Its implementation includes: performing mean shift on the span graph, extracting the edge and ridge sketch of the span graph, and using the region extraction technology based on semantic information to extract the line segment aggregation area in the edge and ridge sketch; The feature merging strategy merges the mean shifted over-segmented regions of the span image to obtain the segmentation result; fuses the image segmentation result based on semantic information and the H/α-Wishart classification result based on MRF to obtain the final classification result. The invention combines semantic information, image processing technology and polarization scattering characteristics, mainly solves the problem of poor regional consistency of classification results of ground objects with aggregation characteristics by existing classification technology based on polarization decomposition, and improves The regional consistency and boundary preservation of the classification results of ground objects (such as forests, building groups, etc.) overcome the shortcomings of pixel-level classification and obtain good polarization SAR ground object classification results.

Claims (6)

1. A polarized SAR terrain classification method based on semantic information and polarization decomposition is characterized in that: the method comprises the following steps:

step 1, inputting data of a polarized SAR image to be classified, processing the polarized SAR data to obtain amplitude values of three channels of the polarized SAR data, fusing the amplitude values of the three channels to obtain a power map, namely a span map, of the polarized SAR image, and obtaining an over-segmentation result map of the span map by using mean shift; extracting a side ridge sketch consisting of line segments of the span graph, namely SketchMap, according to the primeskatch sparse representation model;

step 2, semantic information analysis is carried out on the line segments in the SketchMap, and semantic information, namely two-side aggregation, one-side aggregation and isolated line segments, is given to the line segments according to the statistical distribution of the line segment aggregation characteristics;

step 3, in the SketchMap, extracting a plurality of disjoint aggregation line segment sets by adopting a line segment set solving algorithm according to semantic information given to the line segments, and obtaining a line segment aggregation region R by adopting a region extraction method for each aggregation line segment set;

and 4, carrying out region merging on the over-segmentation result: adopting a critical area mode voting merging strategy for the over-segmentation area corresponding to the line segment gathering area R; the over-segmentation areas where the isolated line segments are located are not merged; adopting a region merging strategy based on polarization characteristics for other regions to obtain a polarized SAR image segmentation result based on semantic information;

step 5, carrying out H/alpha-Wishart classification on the polarized SAR data by utilizing polarization decomposition, and carrying out neighborhood optimization on the H/alpha-Wishart classification result by using MarkovRandomField;

and 6, fusing the segmentation result based on the semantic information and the H/alpha-Wishart classification result based on the MRF, adopting mode voting, selecting the class with the largest number of pixels in the corresponding classification result as the class of the region for each region in the segmentation graph, and endowing the class to the corresponding region in the final classification result graph to obtain the final classification result graph of the ground feature classification of the polarized SAR image to be classified.

2. The polarized SAR terrain classification method based on semantic information and polarization decomposition according to claim 1, characterized in that: the semantic information analysis of the line segment in the step 2 comprises the following steps:

2.1 obtaining three types of feature information corresponding to the line segment according to different distribution of sketch line segments corresponding to different feature types: a line object, a spherical aggregate distribution feature and a boundary between the features;

2.2 the distance between two line segments is defined as the Euclidean distance of the middle points of the line segments, and the average distance of the adjacent line segments K represents the aggregative property of the line segments; and according to the statistical distribution of the aggregation of the line segments, the line segments are endowed with semantic information: aggregating line segments and isolated line segments; dividing the aggregation into two-side aggregation and one-side aggregation according to the topological structure of the aggregation line segment;

2.3 according to the statistical distribution of the line segment aggregations, expressing the semantic information of the line segments in a tree structure, and aggregating the semantic information at two sides to correspond to forest and building group ground objects; gathering boundaries corresponding to forest and building group ground objects on one side; the isolated line segment corresponds to a line target, a bridge manifold feature, or a boundary of two features.

3. The polarized SAR terrain classification method based on semantic information and polarization decomposition as claimed in claim 2, characterized in that: in the step 3, in the SketchMap, according to semantic information given to the line segment, a line segment aggregation solving algorithm is adopted to extract an aggregated line segment set, and an area extraction method is adopted to obtain a line segment aggregation area R for the aggregated line segment set, including:

3.1 symbol definition

Defining: set of sketch line segments is S; spatial constraint threshold₁(ii) a Line segment growth threshold₂(ii) a Satisfying the space constraint line segment set U; aggregating collections of line segmentsSegment aggregation region R ═ { R ═ R₁，r₂，…，r_m}；

3.2 obtaining a plurality of disjoint aggregation line segment sets T by adopting a line segment set solving algorithm_k；

And 3.3, obtaining a line segment aggregation region R by adopting a region extraction method for each aggregation line segment set.

4. The polarized SAR terrain classification method based on semantic information and polarization decomposition according to claim 3, characterized in that: carrying out region merging on the over-segmentation result in the step 4 to obtain a polarized SAR image segmentation result based on semantic information; the method comprises the following steps:

4.1 adopting a critical area mode voting merging strategy for the over-segmentation area corresponding to the line segment aggregation area: for theThere are two overlapping cases of the line segment aggregation area and the over-segmentation area: directly merging the over-segmentation areas by the whole coverage areas of the line segment aggregation areas, and adopting a critical area mode voting merging strategy for the condition that the over-segmentation areas and the boundary areas of the line segment aggregation areas are partially overlapped, wherein the partially overlapped areas at the edges are called critical areas, if the line segment aggregation areas account for more than 50% of the over-segmentation areas, the over-segmentation areas are completely merged into the line segment aggregation areas, and if not, the line segment aggregation areas are divided into wireless segment areas; finally, obtaining a merged line segment gathering area in the over-segmentation graph

4.2, extracting the over-segmentation area of the isolated line segment; the areas are not merged, and the area where the isolated line segment is located is reserved;

4.3 for other areas, defined as wireless section areas, adopting a combination strategy based on polarization characteristics; firstly, regarding each over-segmentation block with mean shift as a super-pixel, counting the polarization characteristic of the super-pixel, adopting three-channel gray level histogram statistics as a feature, quantizing the gray level into 16 parts for each channel, and then calculating a region histogram in the feature space; the total of the three channels is 16 multiplied by 3 which is 48 parts; each region is represented by a 48-dimensional vector, the characteristics of the histogram are normalized, and the similarity of the two regions is calculated according to a Bhattacharyya coefficient calculation formula; setting a merging threshold U, merging adjacent regions with similarity larger than the threshold, calculating histogram features of the merged regions again, and iteratively merging until no regions which can be merged exist;

and obtaining a final region merging result, namely a polarized SAR image segmentation result based on semantic information, by the three merging strategies.

5. The polarized SAR terrain classification method based on semantic information and polarization decomposition as claimed in claim 3, characterized in that: the line segment set solving process comprises the following steps:

3.2.1 first get the sketch line segment setS, counting k neighbors of each line segment according to the fact that the line segments of forest and building group areas have aggregation, calculating the average distance of the k neighbors of each line segment, finding out whether the image line segments have aggregation or not through histogram statistics of the average distance of the k neighbors, if the image line segments have certain aggregation, indicating that the ground features exist, and obtaining a space constraint threshold value according to the histogram statistics₁And threshold for line segment growth₂；

3.2.2 initial setting of T_iIs an empty set; obtaining a line segment set meeting the space constraint according to the space constraint thresholdRandomly selecting seed line segmentsThe growth is carried out, at which time,the criterion for growth is if a certain neighbor of a line segment is presentSatisfying a line segment growth threshold₂Grow into a collection of aggregated line segmentsTraverse its k neighbors until there are no line segments that can grow, assuming this timeFor T at this time_iThe segments which are not traversed in the process are sequentially used as seed segments to grow, iterative growth is carried out until all the grown segments can not be regrown, and an aggregation segment set T is obtained_i；

3.2.3 if the space constraint line segment set U is satisfied and line segments do not grow, selecting one line segment as the seed line segment to continue iterative growth until all the initial seed line segments are grown; finally obtaining a plurality ofSet of intersecting aggregated line segments T_k。

6. The polarized SAR terrain classification method based on semantic information and polarization decomposition as claimed in claim 3, characterized in that: the process of extracting the region of the aggregated line segment set comprises the following steps:

3.3.1 circular cell construction: taking a line segment growing threshold₂Constructing a circular disc for the radius of the circle, i.e. a circular element;

3.3.2 closing operation: performing a closing operation on a set A by using a structural element B, wherein the structural element B is a constructed circular primitive, and the set A is an obtained aggregation line segment set; and obtaining the line segment aggregation region R through the closing operation.