PolSAR-GeoSIFT multi-source data matching method based on SAR geometry configuration

Abstract

This invention relates to a PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry, comprising: obtaining feature points to be matched in the image to be matched and initial reference feature points in the reference image according to the PolSAR-Harris function; processing the initial reference feature points with geographic information to obtain target feature points in the image to be matched and reference feature points in the reference image; wherein the target feature points and reference feature points are mutually matched points. Through the above technical solution, feature points are extracted according to the PolSAR-Harris function, and then the feature points are processed according to geographic information to obtain mutually matched target feature points and reference feature points, thereby achieving the purpose of heterogeneous data matching. This method can be applied to feature point matching of heterogeneous SAR images in different bands, different time phases, and different downward viewing angles, and has high accuracy and computational efficiency.

Description

Technical Field

[0001] This invention relates to the field of radar data processing technology, and more specifically, to a PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry. Background Technology

[0002] Synthetic Aperture Radar (SAR) is a microwave imaging radar system. Because it is not limited by weather and lighting conditions, existing ground-based, airborne, and spaceborne systems have widely applied SAR in fields such as environmental monitoring, topographic mapping, and resource exploration.

[0003] Polarization is an important property of electromagnetic waves, which can describe additional information about the characteristics of ground surfaces, such as vegetation structure, soil moisture, and land cover type. Polarimetric Synthetic Aperture Radar (PolSAR) has the capability to measure the full polarization scattering information of targets and is often used in fields such as ground feature decomposition, ground feature classification and interpretation, and scattering mechanism analysis.

[0004] SAR image registration is a prerequisite for the joint analysis and processing of multi-band or multi-source SAR data. In Interferometric Synthetic Aperture Radar (InSAR), SAR image registration is mainly based on image grayscale information or image features. Registration based on image grayscale information generally utilizes a sliding window search to estimate the offset of corresponding points through an optimized measure function. Measure functions mainly include correlation coefficients, spectral functions, and wave functions. Registration methods based on image grayscale information are sensitive to image noise and rotational distortion, and are mainly suitable for SAR data from the same sensor without nonlinear distortion. For heterogeneous data, factors such as different bands or incident angles can lead to inconsistencies in image grayscale features. Different observation geometry and operating modes can also cause spatial variations in the image. In such cases, the corresponding transformation relationship of the image can be established by extracting invariant features from the image. In 1999, scholars first proposed the Scale Invariant Feature Transform (SIFT) algorithm; this algorithm has rotation and scale invariance and has been widely used in computer vision and image processing. In 2014, some scholars proposed a SIFT-like algorithm for SAR data—SAR-SIFT. This method uses an exponentially weighted mean ratio to calculate the image gradient, which can effectively reduce the false alarm rate of SIFT feature selection. In 2016, some scholars introduced an enhanced feature matching method—PSO-SIFT algorithm—based on the SIFT algorithm, which has good adaptability in multispectral and multi-source sensor image matching.

[0005] SIFT features possess scale, rotation, and affine transformation invariance, offering significant advantages for SAR data registration. However, their high algorithm complexity, computational efficiency, and memory consumption limit their application to some extent. Furthermore, for heterogeneous SAR data, the drastic rotation and scale variations between images pose even greater challenges to the accuracy and computational efficiency of local feature point detection. Therefore, achieving accurate and efficient heterogeneous image registration and feature point extraction is an indispensable task for addressing the needs of SAR data cross-processing and joint applications. Summary of the Invention

[0006] To address the aforementioned problems in the prior art, this invention provides a PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry.

[0007] According to a first aspect of the present invention, a PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry is provided, the method comprising:

[0008] The PolSAR-Harris function is used to obtain the feature points to be matched in the image to be matched and the initial reference feature points in the reference image;

[0009] The initial reference feature points are processed using geographic information to obtain the target feature points in the image to be matched and the reference feature points in the reference image; wherein the target feature points and the reference feature points are mutually matched points.

[0010] Optionally, after processing the feature points to be matched and the initial reference feature points using geographic information to obtain mutually matching target feature points and reference feature points, the method further includes:

[0011] The first sub-feature point and the second sub-feature point are obtained from the target feature point and the reference feature point, respectively.

[0012] The transformation matrix is ​​obtained based on the first sub-feature point and the second sub-feature point;

[0013] The transformation matrix is ​​used to verify whether there is a match between the target feature point and other target feature points and other reference feature points.

[0014] Optionally, obtaining the feature points to be matched in the image to be matched and the initial reference feature points in the reference image according to the PolSAR-Harris function includes:

[0015] Obtain the first image value of the first pixel in the image to be matched and the second image value of the second pixel in the reference image;

[0016] The first exponentially weighted average ratio of the rectangular regions on both sides of the first pixel is obtained based on the first image value, and the first exponentially weighted average ratio of the rectangular regions on both sides of the second pixel is obtained based on the second image value.

[0017] The first matrix and the second matrix are obtained based on the first mean ratio and the second mean ratio;

[0018] The feature points to be matched and the initial reference feature points are obtained based on the PolSAR-Harris function, the first matrix, and the second matrix.

[0019] Optionally, obtaining the first image value of the first pixel in the image to be matched and the second image value of the second pixel in the reference image includes:

[0020] The first surface scattering component corresponding to the first pixel and the second surface scattering component corresponding to the second pixel are obtained according to the surface scattering model.

[0021] The first secondary scattering component corresponding to the first pixel and the second secondary scattering component corresponding to the second pixel are obtained according to the secondary scattering model.

[0022] The first image value is obtained based on the first surface scattering component and the first secondary scattering component;

[0023] The second image value is obtained based on the second surface scattering component and the second secondary scattering component.

[0024] Optionally, the step of processing the initial reference feature points with the feature points to be matched using geographic information to obtain the target feature points of the image to be matched and the reference feature points in the reference image includes:

[0025] Obtain the first geographic information of the feature point to be matched and the first coordinates of the feature point to be matched in the image to be matched;

[0026] The second coordinates of the feature point to be matched in the reference image are obtained based on the first geographic information and the first coordinates.

[0027] Obtain the third coordinate of the initial reference feature point corresponding to the feature point to be matched in the reference image;

[0028] The feature points to be matched and the initial reference feature points whose difference between the second coordinate and the third coordinate is less than a preset threshold are respectively used as target feature points and reference feature points.

[0029] Optionally, the PolSAR-Harris function is represented as follows:

[0030] R SH (x,y,σ)=Det(C PH (x,y,σ))-d·Trace(C PH (x,y,σ)) 2 ;

[0031] Among them, R SH (x,y,σ) represents the coordinates of a pixel (x,y), and Det() calculates the determinant of a matrix. PH (x,y,σ) is a matrix relating to pixel (x,y), σ is a scale factor, d is a preset threshold for detecting extreme points, and Trace() represents finding the trace of the matrix.

[0032] Optionally, the matrix C with respect to pixel (x,y) PH (x,y,σ) can be represented as follows:

[0033]

[0034] Where G is the convolution function, Γ represents the convolution operation. x,σ' Γ represents the horizontal gradient of a pixel (x, y). y,σ' σ' represents the vertical gradient of pixel (x,y), and σ' is the scale factor of the exponential function.

[0035] Optionally, the first index-weighted mean ratio and the second index-weighted mean ratio are represented as follows:

[0036]

[0037] Among them, R f,σ' M represents the exponentially weighted ratio of the rectangular regions on either side of pixel (x, y). 1,σ' M represents the exponentially weighted mean of the first rectangular region of pixel (x, y). 2,σ' represents the exponentially weighted mean of the rectangular region on the second side of pixel (x,y), and f represents the rotation angle for calculating the mean.

[0038] Optionally, the exponentially weighted mean of the rectangular regions on both sides of the pixel (x,y) is represented as follows:

[0039]

[0040] in, For a weighted exponential function, x * Let x be the difference between the integral pixel and the pixel at (x, y).* R is the difference between the integral pixel and the y-value of the pixel at (x, y). y' The integral range is represented by I1, which represents the first rectangular region of pixel (x,y), and I2, which represents the second rectangular region of pixel (x,y).

[0041] Optionally, the second coordinates of the feature point to be matched in the reference image are obtained with reference to the following formula:

[0042]

[0043] Y'=(|p t -p s |-R near ) / R cell ;

[0044] Wherein, the first coordinates of the matching feature point in the image to be matched are (X,Y), and the first geographic information is p. t =(X T ,Y T Z T ), p s The reference image at time n in the azimuth direction a The reference image platform position vector, v is the velocity vector of the reference image platform relative to the matching feature point, and f' is the velocity vector of the reference image platform relative to the matching feature point. dc Let f be the Doppler center frequency at a certain location and time, λ be the wavelength, and f be the frequency. dc R is the Doppler center frequency, X' is the azimuth coordinate in the second coordinate system, and R is the center frequency of the Doppler. near R is the nearest slant distance of the reference image. cell Y' is the range sampling interval of the reference image, and Y' is the range coordinate in the second coordinate system.

[0045] The technical solution provided by this invention may include the following beneficial effects:

[0046] The above technical solution extracts feature points based on the PolSAR-Harris function, and then processes the feature points based on geographic information to obtain mutually matching target feature points and reference feature points, thereby achieving the purpose of heterogeneous data matching. It can be applied to feature point matching of heterogeneous SAR images in different bands, different time phases, and different downward viewing angles, and has high accuracy and computational efficiency.

[0047] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0049] Figure 1 This is a flowchart illustrating a PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry, according to an exemplary embodiment.

[0050] Figure 2 This is a schematic diagram of an experimental scenario according to an exemplary embodiment.

[0051] Figure 3 This is a polarized pseudo-color composite image of GF-3 in an experimental scenario, as illustrated in an exemplary embodiment.

[0052] Figure 4 This is a polarization pseudo-color composite image of TanDEM-X in an experimental scenario, as illustrated in an exemplary embodiment.

[0053] Figure 5 This is a schematic diagram illustrating the SIFT feature point matching comparison results corresponding to a specific embodiment.

[0054] Figure 6 This is a schematic diagram illustrating the SAR-SIFT feature point matching comparison results corresponding to a specific embodiment.

[0055] Figure 7 This is a schematic diagram illustrating the PSO-SIFT feature point matching comparison results corresponding to a specific embodiment.

[0056] Figure 8 This is a schematic diagram illustrating the SAR-GeoSIFT feature point matching comparison results corresponding to a specific embodiment.

[0057] Figure 9 This is a schematic diagram illustrating the PolSAR-GeoSIFT feature point matching comparison results corresponding to a specific embodiment.

[0058] Figure 10 This is a schematic diagram illustrating the SIFT feature point matching comparison results corresponding to Case 2, according to an exemplary embodiment.

[0059] Figure 11 This is a schematic diagram of the SAR-SIFT feature point matching comparison results corresponding to Case 2, as shown in an exemplary embodiment.

[0060] Figure 12This is a schematic diagram illustrating the PSO-SIFT feature point matching comparison results corresponding to Case 2, according to an exemplary embodiment.

[0061] Figure 13 This is a schematic diagram illustrating the SAR-GeoSIFT feature point matching comparison results corresponding to Case 2, as shown in an exemplary embodiment.

[0062] Figure 14 This is a schematic diagram illustrating the PolSAR-GeoSIFT feature point matching comparison results corresponding to Case 2, as shown in an exemplary embodiment.

[0063] Figure 15 This is a schematic diagram illustrating the SIFT feature point matching comparison results corresponding to Case 3 according to an exemplary embodiment.

[0064] Figure 16 This is a schematic diagram illustrating the SAR-SIFT feature point matching comparison results corresponding to Case 3 according to an exemplary embodiment.

[0065] Figure 17 This is a schematic diagram illustrating the PSO-SIFT feature point matching comparison results corresponding to Case 3 according to an exemplary embodiment.

[0066] Figure 18 This is a schematic diagram illustrating the SAR-GeoSIFT feature point matching comparison results corresponding to Case 3 according to an exemplary embodiment.

[0067] Figure 19 This is a schematic diagram illustrating the PolSAR-GeoSIFT feature point matching comparison results corresponding to Case 3 according to an exemplary embodiment. Detailed Implementation

[0068] Figure 1 This is a flowchart illustrating a PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry, according to an exemplary embodiment. Figure 1 As shown, the method includes the following steps.

[0069] S101. Obtain the feature points to be matched in the image to be matched and the initial reference feature points in the reference image according to the PolSAR-Harris function.

[0070] It is understandable that the PolSAR-Harris function, i.e., the polarimetric synthetic aperture radar Harris function, is derived from the polarimetric synthetic aperture radar Harris function. PolSAR target decomposition is a polarimetric data analysis method based on actual physical constraints to interpret target scattering mechanisms. It identifies and classifies different ground targets by decomposing complex polarimetric scattering matrices, coherence matrices, or covariance matrices into several physically meaningful scattering mechanisms. It is worth mentioning that this invention is used in any two images from heterogeneous sources, with one image serving as the image to be matched and the other as the reference image.

[0071] Optionally, S101 may include:

[0072] Obtain the first image value of the first pixel in the image to be matched and the second image value of the second pixel in the reference image;

[0073] The first exponentially weighted average ratio of the rectangular regions on both sides of the first pixel is obtained based on the first image value, and the first exponentially weighted average ratio of the rectangular regions on both sides of the second pixel is obtained based on the second image value.

[0074] The first matrix and the second matrix are obtained based on the first mean ratio and the second mean ratio;

[0075] The feature points to be matched and the initial reference feature points are obtained based on the PolSAR-Harris function, the first matrix, and the second matrix.

[0076] It is understandable that the first image value of the first pixel in the image to be matched and the second image value of the second pixel in the reference image are used to obtain the polarization index weighted average ratio of the first pixel and the second pixel, respectively, and then the matching feature points and the initial reference feature points are obtained.

[0077] In one implementation, taking any pixel (x,y) in the image to be matched or the reference image as an example, the image value I(x,y) of pixel (x,y) includes the image values ​​of polarization data surface scattering and secondary scattering mechanisms, and the image value I(x,y) can be expressed as I(x,y) = f S (x,y)+f D (x,y), f S (x,y) and f D (x, y) represent the surface scattering component and the secondary scattering component at (x, y), respectively; then the exponentially weighted mean of the rectangular regions on both sides of the pixel (x, y) is represented as follows:

[0078]

[0079] in, For a weighted exponential function, x *Let x be the difference between the integral pixel and the pixel at (x, y). * R is the difference between the integral pixel and the y-value of the pixel at (x, y). y' The integral range is represented by I1, which represents the first rectangular region of pixel (x,y), and I2, which represents the second rectangular region of pixel (x,y).

[0080] Optionally, the first exponentially weighted mean ratio and the second exponentially weighted mean ratio can both be represented by the following mean ratios:

[0081]

[0082] Among them, R f,σ' M represents the exponentially weighted ratio of the rectangular regions on either side of pixel (x, y). 1,σ' M represents the exponentially weighted mean of the first rectangular region of pixel (x, y). 2,σ' This represents the exponentially weighted mean of the rectangular region on the second side of pixel (x,y), where f represents the rotation angle for the mean calculation. When f=1, it represents the weighting ratio in the vertical direction; when f=3, it represents the weighting ratio in the horizontal direction.

[0083] Furthermore, the exponentially weighted mean ratio is normalized to obtain the horizontal gradient Γ of the pixel (x,y). x,σ' and vertical gradient Γ y,σ' Please refer to the following representation:

[0084] Γ x,σ' =log(R) 1,σ' )

[0085] Γ y,σ' =log(R) 3,σ' );

[0086] Based on the horizontal gradient Γ of pixel (x,y) x,σ' and vertical gradient Γ y,σ' We can obtain a matrix C for the pixel (x, y). PH (x,y,σ), the matrix C PH (x,y,σ) can be represented as follows:

[0087]

[0088] Where G is the convolution function, Γ represents the convolution operation. x,σ' Γ represents the horizontal gradient of a pixel (x, y). y,σ' σ' represents the vertical gradient of pixel (x,y), and σ' is the scale factor of the exponential function.

[0089] Based on the PolSAR-Harris function and matrix CPH (x,y,σ) obtains the coordinate information of pixel (x,y), that is, the coordinate information of the feature point to be matched in the image to be matched and the initial reference feature point in the reference image.

[0090] The PolSAR-Harris function reference is as follows:

[0091] R SH (x,y,σ)=Det(C PH (x,y,σ))-d·Trace(C PH (x,y,σ)) 2 ;

[0092] Among them, R SH (x,y,σ) represents the coordinates of a pixel (x,y), and Det() calculates the determinant of a matrix. PH (x,y,σ) is a matrix about the pixel (x,y), σ is a scale factor, d is a preset extreme point detection threshold, which can usually be set to 0.04 to 0.06 based on experience, and Trace() means to find the trace of the matrix.

[0093] Optionally, obtaining the first image value of the first pixel in the image to be matched and the second image value of the second pixel in the reference image includes:

[0094] The first surface scattering component corresponding to the first pixel and the second surface scattering component corresponding to the second pixel are obtained based on the surface scattering model.

[0095] The first secondary scattering component corresponding to the first pixel and the second secondary scattering component corresponding to the second pixel are obtained according to the secondary scattering model.

[0096] The first image value is obtained based on the first surface scattering component and the first secondary scattering component;

[0097] The second image value is obtained based on the second surface scattering component and the second secondary scattering component.

[0098] It is understandable that the scattering matrix S1 of the surface scattering model is:

[0099]

[0100] Among them, R H and R V These are the complex scattering coefficients for horizontal and vertical polarization, respectively.

[0101] Furthermore, the covariance matrix C of the surface scattering model can be obtained from the scattering matrix S1 of the surface scattering model. S :

[0102]

[0103] Among them, f S f represents the contribution of the surface scattering component. S =|R V | 2 ,β * Denotes the conjugate matrix of β.

[0104] The secondary scattering model can simulate the scattering of dihedral reflectors made of different dielectric materials. The scattering matrix S2 of the secondary scattering model can be represented as follows:

[0105]

[0106] Among them, R TH and R TV R represents the scattering coefficients of the vertical tree trunk for horizontally polarized waves and vertically polarized waves, respectively. GH and R GV These are the scattering coefficients of the horizontal Earth surface, and These are the horizontal and vertical polarization propagation factors, respectively.

[0107] Furthermore, based on the scattering matrix S2 of the secondary scattering model, the covariance matrix C corresponding to the secondary scattering can be obtained. D :

[0108]

[0109] Among them, f D f represents the contribution of the secondary scattering component. D =|R TV R GV | 2 ,

[0110] S102. Using geographic information, the initial reference feature points are processed to obtain the target feature points in the image to be matched and the reference feature points in the reference image; wherein, the target feature points and the reference feature points are mutually matched points.

[0111] Understandably, to achieve rotation invariance in the SIFT algorithm, traditional SIFT methods rely on the image's texture structure, determining the principal and secondary directions of feature point distribution based on the gradient values ​​of pixel neighborhoods. Due to the SAR side-looking imaging mechanism, SAR data not only exhibits rotational scale variations but also suffers from image geometric distortion caused by terrain undulations. Since the actual coordinates of feature points representing the same geographical location in two images should be identical, geographic information can be used to project feature points from both images into the same coordinate space. This eliminates rotation and scale transformations between feature points, resulting in target and reference feature points with rotation and scale stability.

[0112] Optionally, S102 may include:

[0113] Obtain the first geographic information of the feature points to be matched and the first coordinates of the feature points in the image to be matched;

[0114] The second coordinates of the feature point to be matched in the reference image are obtained based on the first geographic information and the first coordinates.

[0115] Obtain the third coordinate of the initial reference feature point in the reference image corresponding to the feature point to be matched;

[0116] The feature points to be matched and the initial reference feature points whose difference between the second and third coordinates is less than a preset threshold are respectively used as target feature points and reference feature points.

[0117] Understandably, the feature points to be matched in the image to be matched and the initial reference feature points in the reference image obtained by the PolSAR-Harris function contain the coordinate information of the feature points to be matched and the coordinate information of the initial reference feature points, respectively. Taking a single feature point to be matched in the first image to be matched as an example, assuming that the coordinates of the feature point to be matched in the image to be matched are (X,Y), the first geographic information p is obtained by locating the feature point based on the platform position vector and load parameters of the image to be matched. t =(X T ,Y T Z T In one implementation, the second coordinates of the feature point to be matched in the reference image are obtained with reference to the following formula:

[0118]

[0119] Y'=(|p t -p s |-R near ) / R cell ;

[0120] Wherein, the first coordinates of the matching feature point in the image to be matched are (X,Y), and the first geographic information is p. t=(X T ,Y T Z T ), p s For the reference image, the time in the azimuth direction is n. a The reference image platform position vector, v is the velocity vector of the reference image platform relative to the matching feature points, and f' is the velocity vector of the reference image platform relative to the matching feature points. dc Let f be the Doppler center frequency at a certain location and time, λ be the wavelength, and f be the frequency. dc R is the Doppler center frequency, X' is the azimuth coordinate in the second coordinate system, and R is the center frequency. near R is the nearest slope distance to the reference image. cell Given the range sampling interval of the reference image and Y' as the range coordinate in the second coordinate system, the feature description vector F of the matching feature points in the reference image can be obtained. R = [X'Y'] and the feature description vector F of the matching feature points in the image to be matched M =[XY], further, the feature points to be matched and the initial reference feature points whose difference between the second and third coordinates is less than a preset threshold are respectively used as target feature points and reference feature points.

[0121] Optionally, after S102, the method may further include:

[0122] Obtain the first and second sub-feature points from the target feature points and the reference feature points, respectively;

[0123] The transformation matrix is ​​obtained based on the first and second sub-feature points;

[0124] Verify whether there is a match between the target feature point and other target feature points and other reference feature points based on the transformation matrix.

[0125] As can be understood, feature point matching involves finding corresponding feature points in two or more images to obtain the transformation relationship between the images and achieve image matching. Image transformations typically include affine transformations such as rotation, translation, scaling, and flipping. For SAR images, side-view imaging makes the transformation relationship between images more complex. Therefore, this invention uses perspective transformation to describe the transformation relationship between the reference image and the image to be matched, which can be expressed as:

[0126]

[0127] Where H is the perspective transformation matrix. Represents a linear transformation of an image, [h 31 h 32 ] represents the perspective transformation of the image, h 13 h 23 1 indicates a translation transformation of the image.

[0128] Expanding the above equation, we get:

[0129]

[0130] The perspective transformation matrix H can be solved by matching the feature points of two images. This allows the transformation relationship between the two images to be described using the perspective transformation matrix H, and further verifies whether the extracted target feature points and reference feature points are reasonable.

[0131] In one embodiment, the present invention will be further described below with reference to the accompanying drawings and examples, but the embodiments of the present invention are not limited thereto.

[0132] In this embodiment, Figure 2 This is a schematic diagram of an experimental scenario according to an exemplary embodiment. Figure 3 This is a polarized pseudo-color composite image of GF-3 in an experimental scenario, according to an exemplary embodiment. Figure 4 This invention presents a polarimetric pseudo-color composite image of TanDEM-X in an experimental scenario, according to an exemplary embodiment. The performance of the heterogeneous SAR data matching algorithm is verified using C-band fully polarimetric data from Gaofen-3 in Xi'an and X-band fully polarimetric data from TanDEM-X. The traditional SIFT method is employed, and the enhanced feature matching method PSO-SIFT, commonly used for SAR data, is compared with the method proposed in this invention. To further verify the necessity of using surface scattering and secondary scattering components to improve the accuracy and efficiency of the algorithm, this invention is divided into SAR-GeoSIFT and PolSAR-GeoSIFT methods. The SAR-GeoSIFT method uses SAR-Harris spatial extraction, consistent with SAR-SIFT, to extract feature points, while the PolSAR-GeoSIFT method uses the PolSAR-Harris function to extract feature points related to the polarimetric scattering mechanism. In this embodiment, the algorithm matching performance and computational performance are compared under three different conditions. No manual initialization is performed during feature point matching in any of the methods. The reference images are all from Gaofen-3 data, and the data to be matched are TanDEM-X data at different translation and scaling scales, which have different degrees of geometric deformation compared with the reference images. The specific parameters are shown in Table 1.

[0133] Table 1

[0134]

[0135] To verify the results, to ensure accurate matching of PolSAR-GeoSIFT feature points, in each case, for the initially extracted feature points, the feature points are first screened based on the Nearest Neighbor Distance Ratio (NNDR), and then purified using the Random Sample Consensus (RANSAC) purification algorithm to refine the correctly matched feature points after the initial screening. Finally, the projection transformation matrix is ​​fitted based on the correct feature points.

[0136] Scenario 1:

[0137] Figure 5 This is a schematic diagram illustrating the SIFT feature point matching comparison results corresponding to a specific embodiment. Figure 6 This is a schematic diagram illustrating the SAR-SIFT feature point matching comparison results corresponding to a specific exemplary embodiment. Figure 7 This is a schematic diagram illustrating the PSO-SIFT feature point matching comparison results corresponding to a specific embodiment. Figure 8 This is a schematic diagram illustrating the SAR-GeoSIFT feature point matching comparison results corresponding to a specific exemplary embodiment. Figure 9 This is a schematic diagram of the PolSAR-GeoSIFT feature point matching comparison results corresponding to a scenario shown in an exemplary embodiment. The initial TanDEM-X data is processed by multi-view processing and common area truncation, so that the two sets of data have the same scale but a certain translation. The data sizes are 1143×2172 and 1100×2000, respectively.

[0138] The comparison parameters are shown in Table 2. The results show that, in the absence of significant scale changes, all five SIFT methods can achieve accurate feature point matching. The SIFT method has the fewest matching feature points, which are concentrated in the urban buildings in the center of the scene. The other methods have a similar number of feature points. Among them, the SAR-SIFT, PSO-SIFT, and SAR-GeoSIFT algorithms have similar matching feature point distributions. Compared with the SIFT algorithm, their feature point distribution is more uniform, but they are still concentrated in the buildings in the urban and suburban areas. The PolSAR-GeoSIFT method covers the entire scene with matching feature points, distributed along the riverbanks on the left and upper right sides of the scene, except for the building area. This proves that the perspective transformation matrix based on the matching feature point estimation can be effective for the entire scene.

[0139] The comparison results were further quantitatively analyzed, and the number of initially extracted feature points, the number of matching feature point pairs screened by NNDR (Nearest Neighbor Distance Ratio), the number of matching feature point pairs purified by RANSAC (Random Sample Consensus), and the matching runtime were statistically analyzed, as shown in the table below. In terms of the number of image feature points, the SIFT and PSO-SIFT methods extracted approximately the same number of image feature points. PSO-SIFT significantly improved the final number of matching feature points by sacrificing computational efficiency. The SAR-SIFT, SAR-GeoSIFT, and PolSAR-GeoSIFT methods outperform the other two methods in terms of the number of initial feature point detections and feature point matches. However, the two GeoSIFT methods significantly improve the number of feature points matched by NNDR by nearly two times compared to the SAR-SIFT method. The number of matching points after RANSAC is about 20.43% higher than that of the SAR-SIFT method, and the algorithm's running time is improved by an order of magnitude. This is because the GeoSIFT description vector is a 2D vector that describes the rotation, scale, and displacement invariant features of the reference image and the image to be matched in 3D geographic space, while the SAR-SIFT description vector is 128-dimensional. Furthermore, since the description vector corresponding to each feature point is not unique, confusion can easily occur when performing similarity judgment.

[0140] Table 2

[0141]

[0142] Scenario 2:

[0143] Figure 10 This is a schematic diagram illustrating the SIFT feature point matching comparison results corresponding to Case 2, according to an exemplary embodiment. Figure 11 This is a schematic diagram illustrating the SAR-SIFT feature point matching comparison results corresponding to Case 2, according to an exemplary embodiment. Figure 12 This is a schematic diagram illustrating the PSO-SIFT feature point matching comparison results corresponding to Case 2, according to an exemplary embodiment. Figure 13 This is a schematic diagram illustrating the SAR-GeoSIFT feature point matching comparison results corresponding to Case 2, according to an exemplary embodiment. Figure 14 This is a schematic diagram of the PolSAR-GeoSIFT feature point matching comparison results corresponding to Case 2 according to an exemplary embodiment. Only the common area of ​​the initial TanDEM-X data is truncated, resulting in obvious scale transformation and translation phenomena between the two sets of data. The data sizes are 1143×2172 and 2200×2000, respectively.

[0144] The comparison parameters are shown in Table 3. The results show that the introduction of scale variation leads to a sharp deterioration in the performance of SIFT, PSO-SIFT, and SAR-SIFT methods. SIFT and PSO-SIFT methods each obtain only one correctly matched feature point, resulting in the failure of perspective matrix estimation and the inability to achieve normal data matching. SAR-SIFT method obtains only 7 matching points, all of which are concentrated near urban buildings in the lower part of the scene. Conversely, scale transformation increases the number of feature points and further improves the number of matching feature point pairs extracted by GeoSIFT method. Compared with Case 1, the distribution range of the feature points matched by SAR-GeoSIFT method expands to the riverbanks on the left and upper right sides of the scene. PolSAR-SIFT method also covers the entire scene compared with Case 1, and the distribution density of the matched feature points is significantly improved.

[0145] A quantitative analysis of the comparison results is shown in the table below. It can be seen that the number of feature points detected in the images to be matched increases proportionally for each method. However, the initial feature point matching performance of SIFT, SAR-SIFT, and PSO-SIFT methods deteriorates to varying degrees compared to Case 1. SIFT and PSO-SIFT even fail to refine to the minimum number of correctly matched feature points required for perspective matrix estimation. In contrast, the GeoSIFT method can still achieve accurate feature point matching, and the PolSAR-GeoSIFT method exhibits better scale robustness than the SAR-GeoSIFT method in this case.

[0146] Table 3

[0147]

[0148] Scenario 3:

[0149] Figure 15 This is a schematic diagram illustrating the SIFT feature point matching comparison results corresponding to case three, based on an exemplary embodiment. Figure 16 This is a schematic diagram illustrating the SAR-SIFT feature point matching comparison results corresponding to Case 3 according to an exemplary embodiment. Figure 17 This is a schematic diagram illustrating the PSO-SIFT feature point matching comparison results corresponding to case three according to an exemplary embodiment. Figure 18 This is a schematic diagram illustrating the SAR-GeoSIFT feature point matching comparison results corresponding to Case 3 according to an exemplary embodiment. Figure 19This is a schematic diagram of the PolSAR-GeoSIFT feature point matching comparison results corresponding to Case 3 according to an exemplary embodiment. No operation is performed on the initial TanDEM-X data. At this time, the two sets of data not only have obvious scale transformation, but also have a large translation phenomenon. The data sizes are 1143×2172 and 2550×2766, respectively.

[0150] The comparison parameters are shown in Table 4. The results show that the SIFT and SAR-SIFT methods obtained 1 and 2 pairs of correctly matched feature points, respectively, while the PSO-SIFT method could no longer obtain correctly matched feature points. The GeoSIFT method can still robustly obtain correctly matched feature points. The matched feature points obtained by the PolSAR-GeoSIFT method are more evenly distributed and have a higher distribution density around the scene compared to the SAR-GeoSIFT method.

[0151] The quantitative analysis of Case 3 is shown in the table below. The results show that the number of feature points detected in the images to be matched by each method still increases proportionally. However, except for SIFT and PSO-SIFT methods, the SAR-SIFT method also cannot extract the minimum number of correctly matched feature points that satisfy the perspective matrix estimation. The GeoSIFT method can still achieve accurate feature point matching, and the number of matched points is approximately the same as in Case 2. This is because the size of the reference image has not changed, and the GeoSIFT method has completely extracted accurate feature points from the reference image in both Case 2 and Case 3.

[0152] Table 4

[0153]

[0154] Comparative analysis of various SIFT methods under three conditions reveals the following: Under consistent scale conditions, SIFT, SAR-SIFT, and PSO-SIFT methods can all achieve correct feature point matching. SIFT has the highest computational efficiency but matches the fewest feature points. SAR-SIFT is more suitable for SAR data, matching the most feature points, but its matching time is the longest. PSO-SIFT achieves a balance between efficiency and performance, matching more feature points in a shorter time. Under inconsistent scale conditions, SIFT, SAR-SIFT, and PSO-SIFT methods deteriorate significantly, almost failing to achieve accurate image matching. The PolSAR-Harris feature point detection proposed in this invention effectively extracts stable feature points from heterogeneous SAR data, while the Geo-SIFT description vector achieves true scale rotation invariance and reduces the vector dimension from 128 to 2, overcoming the two core problems of accuracy and efficiency in heterogeneous SAR data matching.

[0155] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

[0156] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0157] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. A PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry, characterized in that, The method includes: The PolSAR-Harris function is used to obtain the feature points to be matched in the image to be matched and the initial reference feature points in the reference image; The initial reference feature points are processed using geographic information to obtain the target feature points in the image to be matched and the reference feature points in the reference image; wherein the target feature points and the reference feature points are mutually matched points; The step of obtaining the feature points to be matched in the image to be matched and the initial reference feature points in the reference image according to the PolSAR-Harris function includes: Obtain the first image value of the first pixel in the image to be matched and the second image value of the second pixel in the reference image; The first exponentially weighted mean ratio of the rectangular regions on both sides of the first pixel is obtained based on the first image value, and the second exponentially weighted mean ratio of the rectangular regions on both sides of the second pixel is obtained based on the second image value. The first matrix and the second matrix are obtained based on the first index-weighted mean ratio and the second index-weighted mean ratio; The feature points to be matched and the initial reference feature points are obtained based on the PolSAR-Harris function, the first matrix, and the second matrix. The step of obtaining the first image value of the first pixel in the image to be matched and the second image value of the second pixel in the reference image includes: The first surface scattering component corresponding to the first pixel and the second surface scattering component corresponding to the second pixel are obtained according to the surface scattering model. The first secondary scattering component corresponding to the first pixel and the second secondary scattering component corresponding to the second pixel are obtained according to the secondary scattering model. The first image value is obtained based on the first surface scattering component and the first secondary scattering component; The second image value is obtained based on the second surface scattering component and the second secondary scattering component; The step of processing the initial reference feature points with the feature points to be matched using geographic information to obtain the target feature points in the image to be matched and the reference feature points in the reference image includes: Obtain the first geographic information of the feature point to be matched and the first coordinates of the feature point to be matched in the image to be matched; The second coordinates of the feature point to be matched in the reference image are obtained based on the first geographic information and the first coordinates. Obtain the third coordinate of the initial reference feature point corresponding to the feature point to be matched in the reference image; The feature points to be matched and the initial reference feature points whose difference between the second coordinate and the third coordinate is less than a preset threshold are respectively used as target feature points and reference feature points. The second coordinates of the feature point to be matched in the reference image are obtained with reference to the following formula: ； ； Wherein, the first coordinate of the matching feature point in the image to be matched is The first geographical information is , The reference image at time in the azimuth direction is The reference image platform position vector at that time, The velocity vector of the reference image platform relative to the matching feature points. The Doppler center frequency at a certain location and time. For wavelength, The center frequency of the Doppler wave. This refers to the azimuth coordinate in the second coordinate system. The nearest slant distance of the reference image. The distance sampling interval of the reference image. The distance coordinate in the second coordinate system; The PolSAR-Harris function is represented as follows: ； in, Represents pixels Coordinate information, This indicates finding the determinant of a matrix. For about pixels The matrix, As a scale factor, The preset extreme point detection threshold, This indicates finding the trace of a matrix.

2. The PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry according to claim 1, characterized in that, After processing the initial reference feature points with the feature points to be matched using geographic information to obtain mutually matching target feature points and reference feature points, the method further includes: The first sub-feature point and the second sub-feature point are obtained from the target feature point and the reference feature point, respectively. The transformation matrix is ​​obtained based on the first sub-feature point and the second sub-feature point; The transformation matrix is ​​used to verify whether there is a match between the target feature point and other target feature points and other reference feature points.

3. The PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry according to claim 1, characterized in that, Regarding pixels matrix Please refer to the following representation: ； in, It is a convolution function. This represents the convolution operation. Represents pixels The horizontal gradient, Represents pixels The vertical gradient, It is the scaling factor for the exponential function.

4. The PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry according to claim 3, characterized in that, The first index-weighted mean ratio and the second index-weighted mean ratio are represented as follows: ； in, Represents pixels The ratio of the exponentially weighted means of the two rectangular regions. Represents pixels The exponentially weighted mean of the first rectangular region Represents pixels The exponentially weighted mean of the second rectangular region This indicates the rotation angle used for calculating the mean.

5. The PolSAR-GeoSIFT heterogeneous data matching method based on SAR geometry according to claim 4, characterized in that, pixel The exponentially weighted average of the two rectangular regions is represented as follows: ； in, For a weighted exponential function, For the integral of pixels and pixels middle The difference, For the integral of pixels and pixels middle The difference, Indicates the range of integration. Represents pixels The first rectangular area, Represents pixels The second rectangular area.

Images