SAR (Synthetic Aperture Radar) change detection method based on global neighborhood information enhanced difference chart

Abstract

The invention discloses an SAR change detection method based on a global neighborhood information enhanced difference chart, and relates to the technical field of image change detection. The invention aims to solve the problem that the existing SAR change detection method is low in detection efficiency and accuracy. The method comprises the following steps: generating an initial difference image by using a dual-time-phase SAR image pair, and then acquiring a segmentation image containing structure information and difference characteristics by using the initial difference image; obtaining the weight AXi, j of the SAR image X and the weight AYi, j of the SAR image Y by using the segmentation map containing the structure information and the difference characteristics, and updating the SAR image X and the SAR image Y by using the weight AXi, j of the SAR image X and the weight AYi, j of the SAR image Y to obtain the SAR image PGX and the SAR image PGY; acquiring a final difference chart by using the SAR image PGX and the SAR image PGY; and carrying out automatic threshold segmentation on the final difference image to obtain an SAR image change region. The method is used for detecting the SAR image change area.

Description

Technical Field

[0001] The present invention relates to the technical field of image change detection, and in particular to a SAR change detection method based on a global neighborhood information enhanced difference map. Background Art

[0002] With the advancement of remote sensing technology and computer vision technology, remote sensing image change detection has been widely used in environmental monitoring, terrain change assessment, natural disaster monitoring and other fields. The difference map of SAR image is widely used in the field of change detection because of its ability to penetrate clouds and have all-weather imaging capabilities. Therefore, SAR change detection based on difference map has become a research focus in this field.

[0003] The difference map generation in current SAR image change detection methods mostly relies on local neighborhood information, which cannot fully capture large-scale image contextual relationships, resulting in inaccurate recognition of changed areas. Although the image modeling method based on graph theory can capture the global relationship between pixels or superpixels by constructing a graph structure, the existing graph theory-based model has a large amount of computation and is still insufficient in capturing subtle change areas, resulting in low efficiency in difference map generation and difficulty in fully utilizing the contextual information in the image when facing high-resolution images with complex backgrounds or noise interference, which in turn leads to low accuracy and efficiency in SAR image change detection. Summary of the invention

[0004] The purpose of the present invention is to solve the problems of low detection efficiency and accuracy in existing SAR change detection methods, and propose a SAR change detection method based on global neighborhood information enhanced difference map.

[0005] The SAR change detection method based on global neighborhood information enhanced difference map includes the following steps:

[0006] Step 1: Generate an initial difference image using a dual-phase SAR image pair, and then use the initial difference image to obtain a segmentation map containing structural information and difference features;

[0007] The dual-phase SAR image pair includes: SAR image X and SAR image Y;

[0008] Step 2: Obtain the weight AX of the SAR image X using the segmentation map containing structural information and difference features i,j and the weight AY of the SAR image Y i,j , and then use the weight AX of the SAR image X i,j and the weight AY of the SAR image Y i,j Update SAR images X and Y to obtain images PGX and PGY;

[0009] Step 3, using the images PGX and PGY obtained in step 2 to obtain the final difference map;

[0010] Step 4: Automatically perform threshold segmentation on the final difference map to obtain the changed area of ​​the SAR image.

[0011] Furthermore, the step 1 uses the dual-phase SAR image pair to generate an initial difference image, and then uses the initial difference image to obtain a segmentation map containing structural information and difference features, specifically:

[0012] Step 11: Generate an initial difference image DI using a dual-phase SAR image pair;

[0013] The pixel value at (m,n) in the initial difference image is as follows:

[0014]

[0015] Where X and Y are dual-phase SAR image pairs, DI(m,n) is the pixel value at (m,n) in the initial difference image, X(m,n) is the pixel value at (m,n) in SAR image X, and Y(m,n) is the pixel value at (m,n) in SAR image Y;

[0016] Step 1 and 2, using the dual-phase SAR image pair and the initial difference image DI to generate a pseudo-color image;

[0017] Step 13: Perform superpixel segmentation on the pseudo-color image to obtain a segmentation map containing structural information and difference features.

[0018] Furthermore, the pseudo-color image is generated by using the dual-phase SAR image pair and the initial difference image DI in the steps one and two, specifically:

[0019] The pixels X(m,n), Y(m,n) in the dual-phase image pair and the pixels DI(m,n) in the initial difference image are spliced ​​along the channel dimension, and all pixels are spliced ​​to obtain a pseudo-color image.

[0020] Furthermore, the step 1 to 3 performs superpixel segmentation on the pseudo-color image to obtain a segmentation map containing structural information and difference features, specifically: adopting the SLIC algorithm to perform superpixel segmentation on the pseudo-color image.

[0021] Furthermore, the weight AX of the SAR image X is obtained by using the segmentation map containing structural information and difference features in step 2. i,j and the weight AY of the SAR image Y i,j , and then use the weight AX of the SAR image X i,j and the weight AY of the SAR image Y i,j Update SAR images X and Y to obtain images PGX and PGY, specifically:

[0022] Step 21: Obtain the pixel value corresponding to the superpixel position in the segmentation map containing structural information and difference features in the SAR image X, and then obtain the weight AX of the SAR image X i,j :

[0023]

[0024] Among them, X i ′ is the average pixel value in the i-th superpixel region in the SAR image X, X′ j is the average pixel value in the jth superpixel region in the SAR image X, ∈ is the smoothing parameter;

[0025] Step 22: Obtain the pixel value corresponding to the superpixel position in the segmentation map containing structural information and difference features in the SAR image Y, and then obtain the weight AY of the SAR image Y i,j ;

[0026] Step 2: Take the area corresponding to each superpixel in the SAR images X and Y as a node and obtain the self-center graph selfGX of node i in the SAR images X and Y. i 、selfGY i ;

[0027] Step 24: Using the self-centered graph selfGX i ,selfGY i The SAR image is updated to obtain updated SAR images PGX and PGY.

[0028] Furthermore, in step 22, the pixel value corresponding to the superpixel position in the segmentation map containing the structural information and the difference feature is obtained in the SAR image Y, and then the weight AY of the SAR image Y is obtained. i,j , specifically:

[0029]

[0030] Among them, Y i ′ is the average pixel value in the i-th superpixel region in the SAR image Y, Y j ′ is the average pixel value in the jth superpixel region in the SAR image Y.

[0031] Furthermore, in the steps 2 and 3, the area corresponding to each superpixel in the SAR images X and Y is taken as a node to obtain the self-center graph selfGX of the node i in the SAR images X and Y. i ,selfGY i , specifically:

[0032] selfGX i =(XV i ,XEi )

[0033] selfGY i =(YV i ,YE i )

[0034] Among them, selfGX i is the self-centered graph of the i-th node in the SAR image X, XV i is the set of the ith node and its neighboring nodes in the SAR image X, XE i is the set of connecting edges between the ith node and the neighboring nodes of the ith node in the SAR image X, selfGY i is the self-centered graph of the i-th node in the SAR image Y, YV i is the set of the ith node and its neighboring nodes in the SAR image Y, YE i It is a set of connecting edges between the ith node and the neighboring nodes of the ith node in the SAR image Y.

[0035] Furthermore, the use of the self-center graph selfGX in step 24 i ,selfGY i Update the SAR image and obtain the updated SAR images PGX and PGY, specifically:

[0036] X in SAR image X i After update:

[0037]

[0038] Y in SAR image Y i After update:

[0039]

[0040] Among them, PGX i Is the updated X i PGY i is the updated Y i , X j Yes X i Neighborhood nodes, Y j Yes i Neighborhood node of X i The label of the neighboring node, N i Yes X i A collection of neighborhood nodes.

[0041] Furthermore, in step 3, the images PGX and PGY obtained in step 2 are used to obtain a final difference map, and the i-th pixel value in the final difference map is as follows:

[0042]

[0043] Among them, DIfinal i is the i-th pixel value in the final difference map.

[0044] Furthermore, the automatic threshold segmentation of the final difference map in step 4 is implemented by using K-means clustering method, maximum entropy method or iterative threshold method.

[0045] The beneficial effects of the present invention are:

[0046] The method of the present invention designs a graph structure construction and feature aggregation and propagation architecture from the perspective of global information propagation of images, makes full use of the spatial context information of SAR images and the difference feature information between phases, realizes the context information enhancement of SAR images in complex noise environments, improves the accuracy and robustness of difference map generation, and thus improves the accuracy of SAR change detection. At the same time, the present invention reduces the amount of calculation and improves the detection efficiency of SAR change areas. The present invention as a whole includes three parts, initial difference image generation and pseudo-color image construction, global structure map construction, feature information propagation and difference information optimization; the initial difference image generation and pseudo-color image construction part is responsible for performing a logarithmic ratio operator on the dual-phase SAR image to generate an initial difference map, and fuses the dual-phase image and the difference image to generate a pseudo-color map for subsequent super-pixel segmentation; the global structure map construction part gives a super-pixel map containing structural information generated by super-pixel segmentation of the pseudo-color image, and constructs a global robust structure map. The feature information propagation and difference information optimization part aggregates and propagates super-pixel features based on the global neighborhood information propagation mechanism to generate a reconstructed SAR image, and generates a final difference map based on the difference information optimized by the reconstructed image, and finally obtains the SAR change area. The present invention obtains the global contextual relationship in the image through the global information propagation mechanism, and effectively suppresses high noise interference while enhancing image features. Therefore, the present invention can achieve effective context enhancement for dual-phase SAR images in application scenarios such as high noise and complex terrain, improves the accuracy and stability of change area detection, meets actual needs, and is easy to implement. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the present invention;

[0048] Figure 2 Construct flowcharts for difference image generation and pseudo-color images;

[0049] Figure 3 Schematic diagram for superpixel partitioning and global robust structure graph construction;

[0050] Figure 4Schematic diagram of graph structure feature aggregation for the global neighborhood information propagation mechanism. DETAILED DESCRIPTION

[0051] Specific implementation method 1: Figure 1 As shown, the specific process of the SAR change detection method based on the global neighborhood information enhanced difference map in this embodiment is as follows:

[0052] Step 1: Generate an initial difference image using the dual-phase SAR image pair, and then use the initial difference image to obtain a segmentation map containing structural information and difference features, specifically:

[0053] Step 1: Figure 2 As shown, the logarithmic ratio operator is used to generate the initial difference image DI using the dual-phase SAR image pair:

[0054] The pixel value at (m,n) in the initial difference image is as follows:

[0055]

[0056] Where X and Y are dual-phase SAR image pairs, DI(m,n) is the pixel value at (m,n) in the initial difference image, X(m,n) is the pixel value at (m,n) in SAR image X, Y(m,n) is the pixel value at (m,n) in SAR image Y, and || is the absolute value sign;

[0057] Step 1 and 2: Generate a pseudo-color image using the dual-phase SAR image pair and the initial difference image DI:

[0058] Merge the pixels X(m,n), Y(m,n) in the dual-phase image pair and the pixels DI(m,n) in the initial difference image by multi-channel cascade integration (stitching along the channel dimension) to obtain a pseudo-color image;

[0059] Step 13: Perform superpixel segmentation on the pseudo-color image to obtain a segmentation map containing structural information and difference features:

[0060] The superpixel segmentation is implemented by a simple linear iterative clustering (SLIC) algorithm to reduce the number of nodes and computational complexity;

[0061] Among them, the number of superpixels is 5000, the superpixel compactness factor is 1, and the superpixel color space is Lab color space.

[0062] Step 2: Obtain the weight AX of the SAR image X using the segmentation map containing structural information and difference features i,j and the weight AY of the SAR image Y i,j , and then use the weight AX of the SAR image X i,j and the weight AY of the SAR image Yi,j Update SAR images X and Y to obtain SAR images PGX and PGY:

[0063] Step 21: Get the pixel value corresponding to the superpixel position in the segmentation map in the SAR image X, and then get the weight AX of the SAR image X i,j , specifically:

[0064]

[0065] Among them, X i ′ is the average pixel value in the i-th superpixel region in the SAR image X, X′ j is the average pixel value in the jth superpixel region in the SAR image X, ∈ is a smoothing parameter, which is a constant not equal to 0; the superpixel X in the SAR image X i and weight AX i,j The global robust structure diagram that constitutes the SAR image X;

[0066] Step 2: Figure 3 As shown in the figure, the pixel value corresponding to the superpixel position in the segmentation map is obtained in the SAR image Y, and then the weight AY of the SAR image Y is obtained. i,j , specifically:

[0067]

[0068] Among them, Y i ′ is the average pixel value in the i-th superpixel region in the SAR image Y, Y j ′ is the average pixel value in the jth superpixel region in the SAR image Y, ∈ is the smoothing parameter, which is a constant not equal to 0;

[0069] Superpixel Y in SAR image Y i and weight AY i,j The global robust structure graph of the SAR image Y is constructed. The global robust structure graph takes each superpixel region in the segmentation graph as a node and constructs the neighborhood connection relationship by using the feature similarity between nodes.

[0070] Step 2 and 3: Figure 4 As shown in the figure, each superpixel corresponding area in the SAR images X and Y is taken as a node, and the self-center graph selfGX of the node i in the SAR images X and Y is obtained. i ,selfGY i :

[0071] selfGX i =(XV i ,XE i )

[0072] selfGYi =(YV i ,YE i )

[0073] Among them, selfGX i is the self-centered graph of the i-th node in the SAR image X, XV i is the set of the ith node and its neighboring nodes in the SAR image X, XE i is the set of connecting edges between the i-th node and its neighboring nodes in the SAR image X, selfGY i is the self-centered graph of the i-th node in the SAR image Y, YV i is the set of the ith node and its neighboring nodes in the SAR image Y, YE i It is the set of connecting edges between the i-th node and its neighboring nodes in the SAR image Y;

[0074] Step 24: Global neighborhood information propagation mechanism using self-centered graph selfGX i ,selfGY i The superpixels are updated through information propagation, thereby updating the SAR image and obtaining the updated SAR images PGX and PGY:

[0075] SAR image X in X i After the update:

[0076]

[0077] SAR image Y i After the update:

[0078]

[0079] Among them, PGX i Is the updated X i PGY i is the updated Y i , X j Yes X i Neighborhood nodes, Y j Yes i Neighborhood node of X i The label of the neighboring node, N i Yes X i A collection of neighborhood nodes.

[0080] Step 3: Based on the global neighborhood information propagation mechanism, feature aggregation and propagation are performed in the graph structure to reconstruct the SAR image with enhanced context information, and the final difference map DIfinal is obtained using the SAR images PGX and PGY obtained in step 2;

[0081] The i-th pixel value in the final difference map is as follows:

[0082]

[0083] Among them, DIfinal i is the i-th pixel value in the final difference map.

[0084] The images PGX and PGY are constructed by homogeneously assigning the updated representation to each superpixel PGX i and PGY i The pixel set covered maintains the consistency and expansibility of the features in the region, thereby achieving stable transmission of global neighborhood features in the local space. This processing mechanism not only ensures the robustness of the representation of each region, but also improves the fidelity of the structural details of the overall image under the framework of global information propagation.

[0085] The construction process of image PGX is essentially to update the representation PGX of each superpixel area through assignment operation i Uniformly propagates to all pixels in the corresponding area to achieve the internal consistency of the features in the area. This process implicitly assumes that the local features within each superpixel maintain spatial homogeneity, thereby filling the corresponding update value PGX within the range of each superpixel i , gradually completing the process of constructing the overall image PGX. Correspondingly, the generation mechanism of PGY is similar to PGX, that is, the updated feature PGY of each superpixel i Expand to the pixel group under its jurisdiction to achieve a balanced distribution of the features in the region after the update. This operation not only consolidates the feature stability of each region, but also retains the structural details. Under the framework of global neighborhood information propagation, it achieves a feature representation that is both spatially consistent and robust for image features.

[0086] Step 4: Automatically perform threshold segmentation on the final difference map to obtain the changed area of ​​the SAR image;

[0087] The automatic threshold segmentation method includes but is not limited to the Otsu method, the K-means clustering method, the maximum entropy method, the iterative threshold method and the threshold method based on machine learning.

[0088] In this embodiment, the dual-phase SAR image can be a SAR image that has undergone preprocessing operations such as registration and radiation correction; the dual-phase SAR image is two SAR images of the same area, the same size, and different phases, which are input into the dual-phase SAR difference map generation architecture based on global neighborhood information propagation, and the final difference map is obtained at the output layer; the dual-phase SAR difference map generation architecture based on global neighborhood information propagation includes two logarithmic ratio operator operations, one superpixel segmentation operation, and one feature aggregation and propagation operation. The first logarithmic ratio operator operation is to generate an initial difference map for the dual-phase SAR image, and to construct a pseudo-color image in combination with the initial difference map. The superpixel segmentation operation processes the pseudo-color image and constructs a global robust structure map based on feature similarity. The feature aggregation and propagation operation processes the global robust structure map to obtain a reconstructed dual-phase SAR image. The second logarithmic ratio operator operation processes the reconstructed dual-phase SAR image to generate a final difference map, and finally obtains the SAR change area.

[0089] The present invention innovatively proposes a global neighborhood information propagation model to generate dual-phase SAR difference maps. The input of the dual-phase SAR difference map generation model of global neighborhood information propagation is a dual-phase SAR image pair, and the output is a difference map. It includes a global neighborhood information propagation module and a difference feature extraction module, which provides a guarantee for high-precision image analysis such as change detection, anomaly detection, and target recognition, and has certain method advantages. The present invention uses deep learning technology and image processing technology to better meet the change detection and monitoring in different scenarios. The dual-phase SAR difference map generation model of global neighborhood information propagation maintains a certain degree of stability in feature extraction, edge preservation, and noise suppression of the input image.

Claims

1. A SAR change detection method based on global neighborhood information enhanced difference map, characterized by: The specific process of the method is: Step 1: Generate an initial difference image using a dual-phase SAR image pair, and then use the initial difference image to obtain a segmentation map containing structural information and difference features; The dual-phase SAR image pair includes: SAR image X and SAR image Y; Step 2: Obtain the weight AX of the SAR image X using the segmentation map containing structural information and difference features i,j and the weight AY of the SAR image Y i,j , and then use the weight AX of the SAR image X i,j and the weight AY of the SAR image Y i,j Update SAR images X and Y to obtain images PGX and PGY; Step 3: Use the images PGX and PGY obtained in step 2 to obtain the final difference map; Step 4: Automatically perform threshold segmentation on the final difference map to obtain the changed area of ​​the SAR image.

2. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 1, characterized in that: In the step 1, the dual-phase SAR image pair is used to generate an initial difference image, and then the initial difference image is used to obtain a segmentation map containing structural information and difference features, specifically: Step 11: Generate an initial difference image DI using a dual-phase SAR image pair; The pixel value at (m,n) in the initial difference image is as follows: Where X and Y are dual-phase SAR image pairs, DI(m,n) is the pixel value at (m,n) in the initial difference image, X(m,n) is the pixel value at (m,n) in SAR image X, and Y(m,n) is the pixel value at (m,n) in SAR image Y; Step 1 and 2, using the dual-phase SAR image pair and the initial difference image DI to generate a pseudo-color image; Step 13: Perform superpixel segmentation on the pseudo-color image to obtain a segmentation map containing structural information and difference features.

3. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 2, characterized in that: The step 1 or 2 of generating a pseudo color image using the dual-phase SAR image pair and the initial difference image DI is specifically as follows: The pixels X(m,n), Y(m,n) in the dual-phase image pair and the pixels DI(m,n) in the initial difference image are spliced ​​along the channel dimension, and all pixels are spliced ​​to obtain a pseudo-color image.

4. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 3 is characterized in that: The step 1-3 of performing superpixel segmentation on the pseudo-color image to obtain a segmentation map containing structural information and difference features specifically comprises: performing superpixel segmentation on the pseudo-color image using the SLIC algorithm.

5. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 4, characterized in that: The weight AX of the SAR image X is obtained by using the segmentation map containing structural information and difference features in step 2. i,j and the weight AY of the SAR image Y i,j , and then use the weight AX of the SAR image X i,j and the weight AY of the SAR image Y i,j Update SAR images X and Y to obtain images PGX and PGY, specifically: Step 21: Obtain the pixel value corresponding to the superpixel position in the segmentation map containing structural information and difference features in the SAR image X, and then obtain the weight AX of the SAR image X i,j : Among them, X′ i is the average pixel value in the i-th superpixel region in the SAR image X, X′ j is the average pixel value in the jth superpixel region in the SAR image X, ∈ is the smoothing parameter; Step 22: Obtain the pixel value corresponding to the superpixel position in the segmentation map containing structural information and difference features in the SAR image Y, and then obtain the weight AY of the SAR image Y i,j ; Step 2: Take the area corresponding to each superpixel in the SAR images X and Y as a node and obtain the self-center graph selfGX of node i in the SAR images X and Y. i ,selfGY i ; Step 24: Using the self-centered graph selfGX i 、selfGY i The SAR image is updated to obtain updated SAR images PGX and PGY.

6. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 5, characterized in that: In the step 22, the pixel value corresponding to the superpixel position in the segmentation map containing the structural information and the difference feature is obtained in the SAR image Y, and then the weight AY of the SAR image Y is obtained. i,j , specifically: Among them, Y′ i is the average pixel value in the i-th superpixel region in the SAR image Y, Y′ j is the average pixel value in the jth superpixel area in the SAR image Y.

7. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 6, characterized in that: In the steps 2 and 3, the area corresponding to each superpixel in the SAR images X and Y is taken as a node, and the self-center graph selfGX of the node i in the SAR images X and Y is obtained. i 、selfGY i , specifically: selfGX i =(XV i ,XE i ) selfGY i =(YV i ,YE i ) Among them, selfGX i is the self-centered graph of the i-th node in the SAR image X, XV i is the set of the ith node and its neighboring nodes in the SAR image X, XE i is the set of connecting edges between the ith node and the neighboring nodes of the ith node in the SAR image X, selfGY i is the self-centered graph of the i-th node in the SAR image Y, YV i is the set of the ith node and its neighboring nodes in the SAR image Y, YE i It is a set of connecting edges between the ith node and the neighboring nodes of the ith node in the SAR image Y.

8. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 7, characterized in that: The use of the self-centered graph selfGX in step 24 i 、selfGY i Update the SAR image and obtain the updated SAR images PGX and PGY, specifically: X in SAR image X i After update: Y in SAR image Y i After update: Among them, PGX i Is the updated X i PGY i is the updated Y i , X j Yes X i Neighborhood nodes, Y j Yes i Neighborhood node of X i The label of the neighboring node, N i Yes X i A collection of neighborhood nodes.

9. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 8, characterized in that: In step 3, the images PGX and PGY obtained in step 2 are used to obtain the final difference map. The i-th pixel value in the final difference map is as follows: Among them, DIfinal i is the i-th pixel value in the final difference map.

10. The SAR change detection method based on global neighborhood information enhanced difference map according to claim 9, characterized in that: The automatic threshold segmentation of the final difference map in step 4 is implemented by using K-means clustering method, maximum entropy method or iterative threshold method.

Images