CN104392456B - SAR (synthetic aperture radar) image segmentation method based on depth autoencoders and area charts - Google Patents

Abstract

The invention discloses a SAR image segmentation method based on a depth self-encoder and a region map, which mainly solves the problems of inaccurate and detailed segmentation in the prior art. The implementation steps are: 1. Obtain the SAR pixel sketch map according to the initial sketch model, complete the sketch line segments to obtain the region map, and map the region map to the original image to obtain the aggregated, homogeneous and structural regions; 2. Separate the aggregated and homogeneous regions Train with different depth autoencoders to obtain representations corresponding to all points, and the last two layers of the cascaded coding layer are used as the point features; 3. Construct dictionaries for the aggregated and homogeneous regions respectively, and project the features of each point to the corresponding dictionary and aggregate each point. Regional characteristics of sub-regions; 4. Clustering of sub-regional features of two types of regions; 5. Segmentation of structural regions using superpixel merging under the guidance of sketch line segments; 6. Merging the segmentation results of each region to complete SAR image segmentation. The invention has the advantages of accurate and detailed segmentation and can be used for target recognition.

Description

SAR Image Segmentation Method Based on Depth Autoencoder and Region Map

technical field

The invention belongs to the technical field of image processing, and in particular relates to a method for segmenting SAR images by a deep self-encoder in a deep learning method and a region map of a SAR image, which can be used to further identify and classify SAR images.

Background technique

Synthetic Aperture Radar (SAR) is a radar that uses the relative motion of the radar and the target to synthesize a smaller real antenna aperture into a larger equivalent antenna aperture by data processing. It has all-weather and all-time stable high-resolution imaging The SAR image is an image formed by a larger radar antenna through radar aperture synthesis, and its characteristics make it a very valuable image that is widely used in many fields such as military affairs, agriculture, navigation, and geography. Image segmentation is the technology and process of dividing an image into several interconnected and disjoint regions with unique properties according to its grayscale, texture, structure, aggregation and other characteristics. It is a key link in further image recognition and classification. SAR image segmentation is particularly important as an important application field in image segmentation.

In the traditional SAR image segmentation methods, there are clustering algorithms such as Kmeans, FCM, and spectral clustering. Due to the unique imaging mechanism and replicability of the SAR image itself, the resulting image contains a large amount of coherent speckle noise, resulting in its The separability is extremely poor, and the segmentation results of traditional clustering algorithms are not ideal; there are also some segmentation methods after semi-supervised feature extraction, and the segmentation results are better than the aforementioned clustering segmentation algorithms, but it requires human-computer interaction for labeling Data guides the division of data, and SAR map sources are relatively scarce and cover a wide range, so there is very little marked data, and the cost of marking new ones is high. Affected by speckle noise, the existing SAR image segmentation methods are difficult to suppress its influence, and it is difficult to obtain ideal results, which cannot meet the requirements of the general application of SAR images. In order to reduce the degree of manual participation, unsupervised feature learning methods have become the mainstream of feature extraction methods, and the use of unsupervised feature learning has become an inevitable development method for feature extraction in SAR image segmentation. The selection of features in SAR image segmentation is very important, and it is very difficult to obtain better features in an unsupervised extraction method.

As an emerging technology in unsupervised feature learning methods, deep learning has attracted strong attention due to its successful application in various fields. Deep learning can learn the inherent characteristics of the data itself. Not only can feature learning be performed in an unsupervised manner, but the learned features are better than most other features extracted based on the assumption of its distribution. Introduce deep learning method theory in the field of SAR image segmentation can get better SAR image segmentation results. However, in practical applications, first of all, it will be in trouble due to the inability to obtain ideal training data, and secondly, it is difficult to apply the learned features to the region for segmentation and representation. Therefore, how to effectively combine deep learning methods for SAR image segmentation It is a very difficult problem to obtain a very ideal effect, and there are very few studies based on this technology.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the above-mentioned existing technologies and methods, and propose a SAR image segmentation method based on a deep self-encoder and a region map in deep learning to improve the effect of SAR image segmentation.

The technical solution for realizing the present invention is: obtain the SAR pixel sketch map according to the initial sketch model, complete the sketch line segment in the sketch map to obtain the area map, map the area map to the original image to obtain the aggregation, homogeneity and structure area; The homogeneous area is trained with different depth autoencoders; the trained depth autoencoder is used to obtain the representations corresponding to all points, and the representations of the last two layers of the cascaded coding layer are used as the features of the points; according to the bag-of-words model, the aggregation Construct a dictionary with the homogeneous area, project the features of each point to the corresponding dictionary to obtain the corresponding sub-area features; cluster the sub-area features of the two types of areas respectively; divide the structural area into superpixels and merge them under the guidance of the sketch line segment, and combine them with the uniform Combining the qualitative regions; combining the segmentation results of each region to complete the SAR image segmentation. The specific steps include the following:

(1) According to the initial sketch model, obtain the SAR pixel sketch map, complete the sketch line segment in the sketch map to obtain the area map, and map the area map to the original image to obtain the aggregation area a, homogeneous area b and structural area c;

(2) According to the respective characteristics of the aggregation area a and the homogeneous area b, construct two different depth autoencoders Sa and Sb for the aggregation area a and the homogeneous area b respectively;

(3) According to the positions of the aggregated area a and the homogeneous area b in the area map, sample and train the corresponding depth autoencoders Sa and Sb on the aggregated area a and the homogeneous area b respectively;

(4) Using the two trained deep self-encoders Sa and Sb, obtain the respective multi-layer coding layer representations of all points in the region in the corresponding region type, and concatenate the last two coding layer representations of each point as the point Characteristics;

(5) According to the bag-of-words model, the dictionaries of aggregated area a and homogeneous area b are constructed from the features of all points in the aggregated area a and homogeneous area b, and the features of each point are projected to the corresponding dictionary and the areas of each sub-area are aggregated feature;

(6) Clustering all sub-area features of the aggregation area a and the homogeneous area b, respectively, to obtain the segmentation results of the aggregation area a and the homogeneous area b;

(7) Use the watershed algorithm to divide the structural region c into many superpixels, and merge the superpixels once under the guidance of the sketch lines in the sketch map to obtain the line target and boundary, and then merge other superpixels twice, and Merge the superpixels after the second merging with the sub-regions of the homogeneous region b three times, and the remaining superpixels after the three times merging are independent targets, and complete the segmentation of the structural region c;

(8) Merge the segmentation results of the aggregation area a, the homogeneous area b and the structural area c to obtain the final SAR image segmentation result.

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

1. The present invention fully combines the area map to effectively extract the aggregated area, and uses the depth autoencoder to learn features in both the aggregated area and the homogeneous area, which can extract more essential features of the data, and then obtain more accurate segmentation results.

2. The present invention uses the features learned by the deep self-encoder combined with the bag-of-words model to perform projection to obtain regional features, and uses the hierarchical clustering method for clustering to obtain better regional consistency.

3. The present invention uses superpixel merging method for structure region combined with sketch line segment information, not only can extract line object region, but also can obtain better edge consistency.

Experiments show that the present invention applies the deep self-encoder to SAR image segmentation, and combines region maps, bag-of-words models and superpixel merging to obtain segmentation results with more accurate segmentation, better regional consistency and edge consistency.

Description of drawings

Fig. 1 is the realization overall flowchart of the present invention;

Fig. 2 is the sub-flow chart of segmenting the SAR image gathering area and homogeneous area in the present invention;

Fig. 3 is the SAR image original picture that is used for simulation experiment in the present invention;

Fig. 4 is the sketch figure extracted based on the sketch model in the present invention;

Fig. 5 is the area figure that obtains by sketch figure among the present invention;

Fig. 6 is a structural diagram of two depth autoencoders used in the present invention;

Fig. 7 is the result figure to SAR image gathering region segmentation in the present invention;

Fig. 8 is the result figure of SAR image homogeneous region segmentation in the present invention;

Fig. 9 is a result diagram of using superpixel merging in the SAR image structure region in the present invention;

Fig. 10 is a diagram of the final result of SAR image segmentation using the present invention.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings of the embodiments.

With reference to Fig. 1-2, the detailed implementation steps of the present invention are as follows:

Step 1, obtain the area map of the SAR image.

(1.1) Input a SAR image as shown in Figure 3, and get the sketch of the SAR image according to the initial sketch model, as shown in Figure 4, the initial sketch model is derived from computer vision theory, which is a vision-based A model for abstract representation of images, which can be used to extract sketches of images;

(1.2) Use the ray method to complete the sketch line segments in the sketch map to obtain the area map, as shown in Figure 5. The ray method is a method of emitting rays to all the sketch line segments in the sketch map to find the boundary of the area, which can be used in the area graph extraction;

(1.3) Map the region map to the original map to obtain the aggregation region a, the homogeneous region b and the structural region c.

Step 2, according to the respective characteristics of the aggregation area a and the homogeneous area b, construct two different depth autoencoders Sa and Sb, as shown in Figure 6.

(2.1) Set the input window size of the depth autoencoder Sa of the aggregation area a to be 21*21 during training, and the corresponding number of input layer nodes is 441;

(2.2) Set the input window size of the depth autoencoder Sb training in the homogeneous region b to be 15*15, and the corresponding number of input layer nodes is 225;

(2.3) The number of hidden layers of the encoding layer in the depth autoencoder Sa and the depth autoencoder Sb is set to be 6, and the number of hidden layers of the corresponding decoding layer is also 6;

(2.4) Set the abstract window size of each hidden layer of the coding layer of the deep self-encoder Sa to be 25*25, 19*19, 15*15, 13*13, 9*9, 5*5, respectively corresponding to each layer The number of nodes is: 625, 361, 225, 169, 81, 25;

(2.5) Set the abstract window size of each hidden layer of the coding layer of the deep self-encoder Sb to be 19*19, 13*13, 11*11, 9*9, 7*7, 5*5 respectively, which correspond to each layer The number of nodes is: 361, 169, 121, 81, 49, 25;

(2.6) For the decoding layer part of the depth self-encoder Sa, the last layer of the encoding layer and the first layer of the decoding layer are used as the common layer of the decoding layer and the encoding layer, and the number of nodes in the penultimate layer of the encoding layer is used as the second layer of the decoding layer. For the number of layer nodes, use the number of nodes in the last third layer of the coding layer as the number of nodes in the third layer of the decoding layer, and so on, to obtain the number of nodes in each layer of the decoding layer as 25, 81, 169, 225, 361, 625;

(2.7) For the decoding layer part of the depth self-encoder Sb, the last layer of the encoding layer and the first layer of the decoding layer are used as the common layer of the decoding layer and the encoding layer, and the number of nodes in the penultimate layer of the encoding layer is used as the second layer of the decoding layer. For the number of layer nodes, the number of nodes in the last third layer of the coding layer is used as the number of nodes in the third layer of the decoding layer, and so on, and the number of nodes in each layer of the decoding layer is 25, 49, 81, 121, 169, and 361, respectively.

Step 3: According to the locations of the aggregated area a and the homogeneous area b in the area map, sample and train the corresponding depth autoencoders Sa and Sb on the aggregated area a and the homogeneous area b respectively.

(3.1) Use a 21*21 window in the aggregation area a to sample every 4 dotted windows;

(3.2) Use a 15*15 window in the homogeneous area b to sample every 6 dotted windows;

(3.3) Input the samples in the aggregated area a obtained into the input layer of the depth autoencoder Sa, and input the samples in the homogeneous area b obtained into the input layer of the depth autoencoder Sb, and then two Each deep autoencoder uses the following process to generate initial weights between its layers:

(3.3a) A restricted Boltzmann machine RBM network is formed from the input layer and the first layer of the encoding layer, and the weight of the RBM network is obtained by using the RBM fast learning algorithm based on contrastive divergence, which is used as the weight of the input layer and the encoding layer Initial weights between the first layer;

(3.3b) Use the first layer of the encoding layer and the second layer of the encoding layer to form a new RBM network, and use the RBM fast learning algorithm based on contrastive divergence to obtain the weight of the RBM network as the weight between the two layers Initial weights, generate a new RBM network in turn, and finally get the initial weights between every two layers of all coding layers;

(3.3c) Use the transposition of the initial weight value between the last layer of the coding layer and the penultimate layer as the initial weight value between the first layer and the second layer of the decoding layer, and use the penultimate layer and the penultimate layer of the coding layer The transposition of the initial weights between the third layers is used as the initial weights between the second layer and the third layer of the decoding layer, and so on, to obtain the initial weights between every two layers of all decoding layers;

(3.4) After the initial weights are generated, the two deep autoencoders use the BP algorithm to iteratively adjust the weights of their respective networks.

Step 4: Use the two trained deep self-encoders Sa and Sb to obtain the respective multi-layer coding layer representations of all points in the region in the corresponding region type, and concatenate the last two coding layer representations of each point as the point Characteristics.

(4.1) For all points in the aggregation area a, combine the 10 points around each point to form a sample of 21*21 window size corresponding to the point;

(4.2) For all points in the homogeneous region b, combine the 7 points around each point to form a sample of 15*15 window size corresponding to the point;

(4.3) Input the samples corresponding to each point in the obtained aggregation area a to the input layer of the depth autoencoder Sa, and obtain the corresponding samples of the point according to the weights between the layers of the trained depth autoencoder Sa The multi-layer coding layer representation of , the last two coding layer representations in these representations are concatenated to obtain a 106-dimensional vector as the point feature;

(4.4) Input the samples corresponding to each point in the obtained homogeneous region b to the input layer of the depth autoencoder Sb, and obtain the corresponding The multi-layer encoding layer representation of the sample, the last two encoding layer representations in these representations are concatenated, and a 74-dimensional vector is obtained as the feature of the point.

Step 5, according to the bag-of-words model, construct the dictionaries of aggregation area a and homogeneity area b from the characteristics of all points in aggregation area a and homogeneity area b respectively, and obtain the respective values of aggregation area a and homogeneity area b from the corresponding dictionary. Subregion characteristics.

The bag-of-words model includes processes such as extracting features, constructing a visual dictionary, projecting quantified image features to the visual dictionary, and using word frequency to represent images.

This step is based on the bag-of-words model, and the specific process is as follows:

(5.1) For each sub-area in the aggregation area a, select the 20 features closest to the feature mean as dictionary atoms, and the dictionary atoms selected in all sub-areas together form the dictionary of the aggregation area a;

(5.2) For each sub-region in the homogeneous region b, select the 10 features closest to the feature mean value as dictionary atoms, and the dictionary atoms selected in all sub-regions together form the dictionary of the homogeneous region b;

(5.3) Project the features of all points in the aggregation area a to the dictionary of the aggregation area a using locally constrained linear coding, and obtain the sparse coding of all points in the aggregation area a;

(5.4) Project the features of all points in the homogeneous region b to the dictionary of the aggregation region a using locally constrained linear coding, and obtain the sparse coding of all points in the homogeneous region b;

(5.5) For each sub-area in the aggregation area a and the homogeneous area b, the sparse coding of all points in the sub-area is aggregated into an encoding matrix, and the vector obtained by taking the highest component of each dimension in the encoding matrix is used as the sub-area regional characteristics.

Step 6, clustering all the sub-area features of the aggregation area a and the homogeneous area b respectively.

(6.1) Use the hierarchical clustering method to cluster all the sub-area features of the aggregation area a, and obtain the segmentation result of the aggregation area a, as shown in Figure 7;

(6.2) Use the hierarchical clustering method to cluster all the sub-region features of the homogeneous region b, and obtain the segmentation results of the homogeneous region b, as shown in Figure 8.

Step 7, segment the structural region c.

(7.1) Use the watershed algorithm to over-segment the SAR image to obtain many superpixels of the image:

(7.1a) Calculate the gradient of the image to obtain the gradient map of the image;

(7.1b) Use watershed transformation on the gradient map of the image to obtain the annotation of the image;

(7.1c) Segment the structural region c of the image into many superpixels according to the annotation of the image;

(7.2) Merge the superpixels once under the guidance of the sketch lines in the sketch:

(7.2a) Among the sketch lines in the sketch image, two sketch lines that are parallel and less than 7 pixels apart are determined as a class of line target sketch lines, and the superpixels between the class of line target sketch lines are merged as one line-like target;

(7.2b) Among the general sketch lines, the sketch lines belonging to the same sub-area on both sides are determined as the second-class line target sketch lines, and each side of the second-class line target sketch line is expanded by one pixel as the second-class line target, and other sketch lines as a sketch line to describe the boundary;

(7.3) The secondary merging of other superpixels is to merge the superpixels adjacent to each superpixel except for the line target and the boundary, and the difference between the average gray value is less than 25, and merge iteratively until no Two superpixels that are adjacent and whose gray mean value difference is less than 25, the result is shown in Figure 9;

(7.4) Merge the superpixels after the second merging three times, and merge each superpixel into the sub-region of the homogeneous region b whose difference with the average gray value of the superpixel is the smallest and less than 25, and the remaining superpixels for independent goals.

Step 8: Merge the segmentation results of the aggregation area a, homogeneous area b, and structural area c to obtain the final SAR image segmentation result, as shown in Figure 10.

It can be seen from Fig. 10 that the original SAR image is segmented into 10 categories using the method of the present invention, and the resulting segmentation results are more accurate, detailed, and have better regional consistency and edge consistency.

The above description is only a specific example of the present invention, and does not constitute any limitation to the present invention. Obviously, for those skilled in the art, after understanding the content and principles of the present invention, it is possible to make various modifications and changes in form and details without departing from the principles and structures of the present invention, but these are based on the present invention. The modification and change of the inventive concept are still within the protection scope of the claims of the present invention.

Claims (9)

1. A SAR image segmentation method based on depth self-encoder and region map, comprising the steps: (1) According to the initial sketch model, obtain the SAR pixel sketch map, complete the sketch line segment in the sketch map to obtain the area map, and map the area map to the original image to obtain the aggregation area a, homogeneous area b and structural area c; (2) According to the respective characteristics of the aggregation area a and the homogeneous area b, construct two different depth autoencoders Sa and Sb for the aggregation area a and the homogeneous area b respectively; (3) According to the positions of the aggregated area a and the homogeneous area b in the area map, sample and train the corresponding depth autoencoders Sa and Sb on the aggregated area a and the homogeneous area b respectively; (4) Using the two trained deep self-encoders Sa and Sb, obtain the respective multi-layer coding layer representations of all points in the region in the corresponding region type, and concatenate the last two coding layer representations of each point as the point Characteristics; (5) According to the bag-of-words model, the dictionaries of aggregated area a and homogeneous area b are constructed from the features of all points in the aggregated area a and homogeneous area b, and the features of each point are projected to the corresponding dictionary and the areas of each sub-area are aggregated feature; (6) Clustering all sub-area features of the aggregation area a and the homogeneous area b, respectively, to obtain the segmentation results of the aggregation area a and the homogeneous area b; (7) Use the watershed algorithm to divide the structural region c into several superpixels, and merge the divided superpixels for the first time under the guidance of the sketch lines in the sketch map to obtain the line target and boundary, and then perform other superpixels The pixels are merged for the second time, and the superpixels after the second merger are combined with the sub-regions of the homogeneous region b for the third time. After the third merger, the remaining superpixels are independent targets, and the structure region c is completed. segmentation; (8) Merge the segmentation results of the aggregation area a, the homogeneous area b and the structural area c to obtain the final SAR image segmentation result. 2. The SAR image segmentation method according to claim 1, wherein according to the respective characteristics of the aggregation area a and the homogeneous area b described in step (2), two different depths are respectively constructed for the aggregation area a and the homogeneous area b Autoencoders Sa and Sb, proceed as follows: (2.1) Set the input window size of the depth autoencoder Sa of the aggregation area a to be 21*21 during training, and the corresponding number of input layer nodes is 441; (2.2) Set the input window size of the depth autoencoder Sb training in the homogeneous region b to be 15*15, and the corresponding number of input layer nodes is 225; (2.3) The number of hidden layers of the encoding layer in the depth autoencoder Sa and the depth autoencoder Sb is set to be 6, and the number of hidden layers of the corresponding decoding layer is also 6; (2.4) Set the abstract window size of each hidden layer of the coding layer of the deep self-encoder Sa to be 25*25, 19*19, 15*15, 13*13, 9*9, 5*5, respectively corresponding to each layer The number of nodes is: 625, 361, 225, 169, 81, 25; (2.5) Set the abstract window size of each hidden layer of the coding layer of the deep self-encoder Sb to be 19*19, 13*13, 11*11, 9*9, 7*7, 5*5 respectively, which correspond to each layer The number of nodes is: 361, 169, 121, 81, 49, 25; (2.6) For the decoding layer part of the depth self-encoder Sa, the last layer of the encoding layer and the first layer of the decoding layer are used as the common layer of the decoding layer and the encoding layer, and the number of nodes in the penultimate layer of the encoding layer is used as the second layer of the decoding layer. For the number of layer nodes, use the number of nodes in the last third layer of the coding layer as the number of nodes in the third layer of the decoding layer, and so on, to obtain the number of nodes in each layer of the decoding layer as 25, 81, 169, 225, 361, 625; (2.7) For the decoding layer part of the depth self-encoder Sb, the last layer of the encoding layer and the first layer of the decoding layer are used as the common layer of the decoding layer and the encoding layer, and the number of nodes in the penultimate layer of the encoding layer is used as the second layer of the decoding layer. For the number of layer nodes, the number of nodes in the last third layer of the coding layer is used as the number of nodes in the third layer of the decoding layer, and so on, and the number of nodes in each layer of the decoding layer is 25, 49, 81, 121, 169, and 361, respectively. 3. The SAR image segmentation method according to claim 1, wherein according to the positions of the aggregation area a and the homogeneous area b in the area map described in step (3), sampling is performed on the aggregation area a and the homogeneous area b respectively To train the corresponding deep autoencoders Sa and Sb, proceed as follows: (3.1) Use a 21*21 window in the aggregation area a to sample every 4 dotted windows; (3.2) Use a 15*15 window in the homogeneous area b to sample every 6 dotted windows; (3.3) Input the samples in the acquired aggregation area a to the input layer of the depth autoencoder Sa, input the samples in the acquired aggregation area b to the input layer of the depth autoencoder Sb, and then use two depth Each autoencoder uses the following process to generate initial weights between its layers: (3.3a) A restricted Boltzmann machine RBM network is formed from the input layer and the first layer of the encoding layer, and the weight of the RBM network is obtained by using the RBM fast learning algorithm based on contrastive divergence, which is used as the weight of the input layer and the encoding layer Initial weights between the first layer; (3.3b) Use the first layer of the encoding layer and the second layer of the encoding layer to form a new RBM network, and use the RBM fast learning algorithm based on contrastive divergence to obtain the weight of the RBM network as the weight between the two layers Initial weights, generate a new RBM network in turn, and finally get the initial weights between every two layers of all coding layers; (3.3c) Use the transposition of the initial weight value between the last layer of the coding layer and the penultimate layer as the initial weight value between the first layer and the second layer of the decoding layer, and use the penultimate layer and the penultimate layer of the coding layer The transposition of the initial weights between the third layers is used as the initial weights between the second layer and the third layer of the decoding layer, and so on, to obtain the initial weights between every two layers of all decoding layers; (3.4) After the initial weights are generated, the two deep autoencoders use the BP algorithm to iteratively adjust the weights of their respective networks. 4. SAR image segmentation method according to claim 1, wherein the two depth self-encoders Sa and Sb that use training described in step (4) obtain the respective multi-layer codes of all points in the region in the corresponding region type layer representation, and concatenate the last two layers of encoding layer representations of each point as the feature of the point, proceed as follows: (4.1) Use 21*21 window sampling for all points in the aggregation area a, as the corresponding sample of the point; (4.2) Use a 15*15 window to sample all points in the homogeneous area b as the corresponding sample of the point; (4.3) Input the samples corresponding to each point in the obtained aggregation area a to the input layer of the depth autoencoder Sa, and obtain the corresponding samples of the point according to the weights between the layers of the trained depth autoencoder Sa The multi-layer coding layer representation of , the last two coding layer representations in these representations are concatenated to obtain a 106-dimensional vector as the point feature; (4.4) Input the samples corresponding to each point in the obtained homogeneous region b to the input layer of the depth autoencoder Sb, and obtain the corresponding The multi-layer encoding layer representation of the sample, the last two encoding layer representations in these representations are concatenated, and a 74-dimensional vector is obtained as the feature of the point. 5. SAR image segmentation method according to claim 1, wherein according to the bag-of-words model described in step (5), the feature construction aggregation area a and the homogeneous area are respectively constructed by the features of all points of the aggregation area a and the homogeneous area b In the dictionary of b, the features of each point are projected to the corresponding dictionary and the regional features of each sub-region are gathered, and the steps are as follows: (5.1) For each sub-area in the aggregation area a, select the 20 features closest to the feature mean as dictionary atoms, and the dictionary atoms selected in all sub-areas together form the dictionary of the aggregation area a; (5.2) For each sub-region in the homogeneous region b, select the 10 features closest to the feature mean value as dictionary atoms, and the dictionary atoms selected in all sub-regions together form the dictionary of the homogeneous region b; (5.3) Project the features of all points in the aggregation area a to the dictionary of the aggregation area a using locally constrained linear coding, and obtain the sparse coding of all points in the aggregation area a; (5.4) Project the features of all points in the homogeneous region b to the dictionary of the aggregation region a using locally constrained linear coding, and obtain the sparse coding of all points in the homogeneous region b; (5.5) For each sub-area in the aggregation area a and the homogeneous area b, the sparse coding of all points in the sub-area is aggregated into an encoding matrix, and the vector obtained by taking the highest component of each dimension in the encoding matrix is used as the sub-area regional characteristics. 6. SAR image segmentation method according to claim 1, wherein the described structure region c of step (7) is divided into several superpixels using the watershed algorithm, and is carried out as follows: (6.1) Calculate the gradient of the image to obtain the gradient map of the image; (6.2) Use watershed transformation to the gradient map of the image to obtain the annotation of the image; (6.3) Segment the structural region c of the image into several superpixels according to the annotation of the image. 7. SAR image segmentation method according to claim 1, wherein described in step (7) under the guidance of the sketch line in the sketch map, these superpixels after the segmentation are merged for the first time to obtain line targets and boundaries, Proceed as follows: (7.1) In the sketch lines of the sketch map, two sketch lines parallel and less than 7 pixels apart are determined as the first type of line target sketch lines, and the superpixels between the first type of line target sketch lines are merged as first line target; (7.2) Among the general sketch lines, the sketch lines belonging to the same type of sub-region on both sides are determined as the second-type line target sketch lines, and each side of the second-type line target sketch line is expanded by one pixel as the second-type line target, and the other The sketch line is used as a sketch line to describe the boundary. 8. SAR image segmentation method according to claim 1, wherein said in step (7) carries out the second merging to other superpixels, is to in each superpixel except line object and boundary, its adjacent and The superpixels whose gray mean difference is less than 25 are merged, and merged iteratively until there are no two adjacent superpixels whose gray mean difference is less than 25. 9. The SAR image segmentation method according to claim 1, wherein the superpixels after the second merge and the sub-region of the homogeneous region b are merged for the third time described in step (7), which is to merge the superpixels for the second time Each merged superpixel is respectively merged into a subregion of the homogeneous region b whose difference with the average gray value of the superpixel is the smallest and less than 25.