High-precision registration method for complex images measured by InSAR interferometry

By improving the nonlinear registration method of feature collaborative transformation and SPS energy evaluation, the low accuracy and high error rate problems in InSAR image registration are solved, and high-precision image registration and high-quality interference map generation are achieved, improving the The application potential of InSAR technology.

CN115115678BActive Publication Date: 2024-11-01SHAANXI KAIYUE SURVEYING & DESIGN CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210297307.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-24
Publication Date
2024-11-01
Estimated Expiration
2042-03-24

AI Technical Summary

Technical Problem

The existing InSAR image registration methods have low accuracy and high error rates, which cannot meet the high accuracy requirements of InSAR data. Especially when dealing with the deformation, rotation and scaling of SAR images, it is difficult to generate high-quality DEM.

Method used

InSAR image registration is carried out based on improved feature collaborative transformation method, combining polynomial fitting and resampling, amplitude information is extracted, and accuracy is improved through blocked feature point matching; on this basis, non-linear matching combined with additional SPS energy evaluation The quasi-method eliminates SPS points in the interference graph and improves local registration accuracy.

Benefits of technology

It significantly improves the accuracy and quality of InSAR complex image registration, and the generated interference map and DEM signal-to-noise ratio are improved, reducing registration time, and enhancing the robustness and application range of InSAR technology.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115115678B_ABST
    Figure CN115115678B_ABST
Patent Text Reader

Abstract

The present application proposes a method for improving the registration accuracy in a step-by-step manner in combination with the characteristics of InSAR images. In the first stage, an improved feature collaborative transformation method is used to perform InSAR image registration. Polynomial fitting is used to obtain the offset polynomials of the master and slave images in the distance and azimuth directions, and then resampling and interference calculation are performed to obtain an interference map. The key points of the master and slave images are checked and registered, and the images are divided into several areas of appropriate sizes for feature point matching. In the second stage, on the basis of the improved feature collaborative transformation method, a nonlinear InSAR registration method based on additional SPS energy evaluation is combined to improve the accuracy of local image registration, eliminate SPS points in the interference map obtained by coherent calculation after registration, and perform nonlinear registration on the images based on a large number of singular rotation points. The method has a step-by-step progressive effect, generates high-quality interference maps and DEMs, has high image registration accuracy, low error rate, and has great practical application value.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to an InSAR complex image registration method, and in particular to an InSAR interferometric complex image high-precision registration method, belonging to the technical field of InSAR image registration. Background Art

[0002] Synthetic Aperture Radar Interferometry (InSAR) is a rapidly developing geodetic technology that can obtain high-precision, continuous coverage of ground elevation and surface deformation information around the clock. In recent years, InSAR has been widely used in related fields such as terrain mapping (ground elevation model construction), environmental monitoring, disaster monitoring and assessment, energy resource exploration (oil and gas field exploitation, mineral resource exploitation, groundwater extraction, etc.), and has achieved a series of important results. The great progress of high-resolution, multi-dimensional, multi-station, and multi-network SAR systems, and the gradual maturity of new airborne SAR and ground-based SAR systems have significantly enhanced the capabilities of terrain mapping and surface deformation monitoring. Among them, multi-station InSAR and multi-baseline InSAR technologies have promoted the development of high-precision terrain mapping, while the development of time-series InSAR technology has significantly improved the accuracy and reliability of surface deformation measurement, and has played an increasingly important role in earth science research and development applications.

[0003] However, due to the limitations of data source quality, quantity and related parameter information, most InSAR imaging interferometric data have not yet been widely used and are in urgent need of greater development. There is a lack of methods to study limited SAR complex image data, implement image registration technology in InSAR data processing and subsequent unwrapping work, and finally roughly restore the DEM.

[0004] Image registration is the core, difficulty and hot spot of digital image processing technology. With the development of computer vision and information processing technology, the complexity of image matching methods has become increasingly higher, from simple template matching at the beginning, to angle-rotated image matching, and then to matching methods based on feature point extraction. This series of developments has greatly improved image matching performance.

[0005] The registration methods can be classified according to different criteria and the complexity of the transformation model, and the implementation steps of the registration technology are summarized: feature three-dimensional, similarity measurement, search three-dimensional and strategy.

[0006] For the registration method of InSAR images, the most commonly used method is to first determine the control points, then fit the polynomials with the control points, and at the same time perform coordinate system conversion and resampling on the data to be registered. The most commonly used registration method is the correlation coefficient method in template or pattern matching. Although the registration accuracy cannot meet the requirements of InSAR for data, its results can provide effective initial values ​​for the subsequent further registration of the image. Image registration is the core of the image processing field. If the registration only uses point features, it is difficult to meet the requirements of various methods at present. The existing technology methods have been widely used in remote sensing image processing, computer vision, pattern recognition, etc. The method based on feature points is a hot and difficult problem, and the main methods of this type include Harris registration method and maximum stable extreme region (MSER) ​​registration method, etc. These methods have been successfully applied to optical image registration. Due to the different mechanisms between SAR images and optical images, if the feature-based registration method is directly used in SAR images, some defects in image registration will not be avoided, such as low precision and high error rate.

[0007] There are large deformations in InSAR image registration, and there are large differences between grayscale images of the same feature. Grayscale-based image registration methods are ineffective in SAR image registration. So far, there are many image registration methods, such as the same-name point matching method based on image edge feature points, but this method is very complicated. Texture-driven SAR image automatic registration methods can be used for SAR image registration with more bright spots or feature lines, but these features cannot be found under normal circumstances. The image automatic registration method based on invariant moments initially needs to extract image features in closed areas and then register the same name points, but due to the low signal-to-noise ratio of SAR images, the feature points of closed areas cannot be obtained from the SAR image processing process. The homologous SAR image registration method combining wavelet iterative refinement and information fusion is effective when the SAR image distortion is small, but it is difficult to register images with different resolutions. The ordinary image registration method of the prior art is no longer suitable for real-time SAR image registration when there are large deformations, rotations and scaling problems between the main image and the secondary image.

[0008] In order to synthesize SAR interferograms, two complex images are obtained by observing two identical locations, namely the "master" and "slave" maps. InSAR can obtain two or more images of the same monitored target location at a slight time difference or even at the same time. In obtaining the interferogram, the interference noise is eliminated, so that the information on the interferogram is only related to the location of the monitored target location, thereby making full use of the data information in the obtained SAR image. Compared with other methods of measuring terrain deformation, its difference is that it can accurately, economically, efficiently and quickly solve the thorny problems of terrain and deformation in a large area. Compared with other optical images, the outstanding advantages of InSAR are: First, InSAR can adapt to various weather climates and working hours, has a wide monitoring range, and can obtain information data quickly; second, InSAR can not only use a variety of methods to obtain terrain information of target locations with high monitoring difficulties, but also can be used for high-quality surface deformation monitoring; third, the SAR image data that InSAR can provide has a large time span, and its accuracy in monitoring ground elevation changes can reach millimeter level or even higher. It can obtain continuous and high-precision observation information at a lower cost than other methods, and can be used for continuous observation of ground subsidence.

[0009] However, there are many key factors that affect the quality of InSAR data. One of the important factors is InSAR image registration, which determines the robustness and accuracy of InSAR technology, and thus determines its application and future development. Improving the accuracy of InSAR image registration will promote the development of InSAR technology and expand its practical application range.

[0010] In summary, the existing InSAR complex image registration technology has obvious shortcomings. Its main defects and design difficulties include:

[0011] First, for the registration method of InSAR images, the existing technology generally determines the control points first, and then fits the polynomial with the control points, and at the same time performs coordinate system conversion and resampling on the data to be registered. The registration adopts the correlation coefficient method, and the accuracy cannot meet the requirements of InSAR for data. Image registration only uses point features, which is difficult to meet the requirements. The existing technology based on feature points can generally only be applied to optical image registration. Due to the different mechanisms between SAR images and optical images, if the feature-based registration method is directly used on SAR images, some defects in image registration will not be avoided, such as low accuracy and high error rate; and due to the limitations of data source quality, quantity and related parameter information, the existing technology lacks the registration of limited SAR complex image data, and cannot realize InSAR image registration technology and subsequent unwrapping work, and cannot generate high-quality DEM;

[0012] Second, there are large deformations in InSAR image registration, and there are large differences between the same feature grayscale images. The existing grayscale-based image registration method is invalid in SAR image registration; the same-name point matching method based on image edge feature points is very complicated; the texture-driven SAR image automatic registration method can be used for SAR image registration with more bright spots or feature lines, but these features cannot be found under normal circumstances; due to the low signal-to-noise ratio of SAR images, the feature points of the closed area of ​​the image automatic registration method based on invariant moments cannot be obtained from the SAR image processing process; the homologous SAR image registration method combining wavelet iterative refinement and information fusion is difficult to register images with different resolutions; the existing image registration method is no longer suitable for real-time SAR image registration when there are large deformations, rotations and scaling problems between the main image and the secondary image; image registration directly determines the robustness and accuracy of InSAR technology, and determines the application and future development of InSAR. The low accuracy of current InSAR image registration restricts the development of InSAR technology and its practical application scope;

[0013] Third, the prior art lacks a method for polynomial fitting to obtain the offset polynomials of the master and slave images in the distance and azimuth directions, and is unable to obtain the interference pattern by resampling and interference calculation, and is unable to extract the amplitude of the InSAR complex image, and is unable to perform key point inspection and registration on the obtained master and slave images in blocks. The prior art lacks reliable methods for scale three-dimensional image extreme value inspection, InSAR image feature point positioning, InSAR direction angle determination, image feature point InSAR descriptor, and SAR image registration, and is unable to fully utilize the phase and amplitude information of the image, and is unable to balance high efficiency with the accuracy of the registration results. Inaccurate registration leads to serious decoherence, and there is a lot of noise at the edges of the stripes in the obtained interference phase map, and the stripes are blurred, requiring additional information such as orbital parameters, and the registration takes a long time and has low accuracy.

[0014] Fourth, the existing technology is often affected by a large number of singular points SPS on the interference pattern, and these SPS points come from the registration results with low precision. There is currently no particularly effective method to eliminate such singular points. The existing technology's method of processing SPS points in the image registration process is not ideal. The existing technology lacks a nonlinear InSAR registration method with additional SPS quantity evaluation, and lacks SPS quantity as a registration evaluation standard. It is impossible to eliminate the SPS points in the interference pattern obtained by coherent calculation after registration, and it is impossible to perform nonlinear registration of images based on a large number of singular rotation points, resulting in low quality of the generated interference pattern and low signal-to-noise ratio of the obtained DEM. Summary of the invention

[0015] In view of the shortcomings of the prior art, the present application proposes a method for effectively improving the registration accuracy in a step-by-step manner in combination with the characteristics of InSAR images. The first stage is to use an improved feature collaborative transformation method to perform InSAR image registration, use polynomial fitting to find the offset polynomials of the master and slave images in the distance and azimuth directions, then perform resampling and interference calculation to obtain an interference map, extract the amplitude of the InSAR complex image, perform key point inspection and registration on the master and slave images, and divide the image into several appropriately sized areas for feature point matching; the second stage is to improve the accuracy of local image registration by combining a nonlinear InSAR registration method based on additional SPS energy evaluation on the basis of the improved feature collaborative transformation method, and perform nonlinear registration of images based on a large number of singular rotation points by eliminating SPS points in the interference map obtained by coherent calculation after registration; it has the effect of gradually improving the registration accuracy of InSAR complex images, and gradually generating high-quality interference maps and DEMs, with high image registration accuracy and low error rate, and has great practical application value.

[0016] In order to achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0017] InSAR interferometric complex image high-precision registration method, combined with the characteristics of InSAR images themselves, a method to effectively improve the registration accuracy is proposed step by step. The first stage is to use the improved feature collaborative transformation method to perform InSAR image registration, and use polynomial fitting to calculate the offset polynomials of the master and slave images in the distance and azimuth directions, and then resample and interfere to obtain the interference map, extract the amplitude of the InSAR complex image, and check and align the key points of the obtained master and slave images. The block processing method is adopted to divide the image into several areas of appropriate size for feature point matching; the second stage is to On the basis of the improved feature collaborative transformation method, a nonlinear InSAR registration method based on additional SPS energy evaluation is combined to improve the accuracy of local image registration. The SPS energy is used as the registration evaluation standard. By eliminating the SPS points in the interference map obtained by coherent calculation after registration, the image is nonlinearly registered based on a large number of singular rotation points, and the quality of the interference map generated after InSAR image registration is improved by combining with the improved feature collaborative transformation registration method. The method proposed in this application has the effect of gradually improving the accuracy of InSAR complex image registration, and gradually generating high-quality interference maps and DEMs.

[0018] In the first stage, the InSAR image registration method based on improved feature cooperative transformation is as follows: first, the scale three-dimensional image extreme value test, second, the location of InSAR image feature points, third, the determination of InSAR direction angle, fourth, the InSAR descriptor of image feature points, and fifth, the SAR image registration based on improved feature cooperative transformation;

[0019] In the second stage, the nonlinear InSAR registration method based on additional SPS energy evaluation is as follows: first, the setting and solution of SPS points; second, the InSAR registration operation based on SPS energy evaluation; and third, the nonlinear InSAR image precise registration with additional SPS.

[0020] InSAR interferometric complex image high-precision registration method, further, scale three-dimensional image extreme value test: construct a scale three-dimensional Gaussian pyramid model, first perform Gaussian blur and downsampling of InSAR images at different scales, and obtain a group of images of different sizes, from large to small, from bottom to top to form a pyramid-shaped structure, there are a total of O groups, each group is divided into S layers, the initial image is placed at the bottom of the pyramid, that is, the first layer, and the image obtained after each downsampling is used as the next layer of the pyramid. The pyramid has a total of n layers, and the total number of layers is determined by the size of the initial InSAR image and the size of the highest layer image. The calculation formula is:

[0021] n=log 2 {min(M,N)}-t, t∈[0,log 2 {min(M, N)} Formula 1

[0022] m, n are the sizes of the initial image, t is the minimum logarithmic value of the highest layer image, and the three-dimensional scale L(x, y, σ) of the target image is defined as:

[0023] L(x,y,σ)=G(x,y,σ)*I(x,y) Equation 2

[0024] G(x,y,σ) is a scale-variable Gaussian function, and I(x,y) is the initial image:

[0025]

[0026] Where (x, y) is the coordinate position, parameter σ represents the smoothness of the image, different σ values ​​represent different scales of the image, and show different characteristics of the image. The resolution of I(x, y) in formula 2 is infinite, σ = 0, but the image with scale zero does not exist. The initial scale of the image is set to 0.5, and the small scale σ 1 Generate large scale σ 2 The process is:

[0027]

[0028] Gaussian 3D difference is used to obtain image key points at different scales, and images of adjacent Gaussian 3D scales are subtracted to obtain Equation 5:

[0029]

[0030] D(x,y,σ) is the response value image, k is the three-dimensional multiple of the adjacent scale, and the difference of the upper and lower adjacent images in the Gaussian pyramid is obtained to obtain the Gaussian difference map. The key points to be sought are composed of the local extreme points of the three-dimensional scale of the Gaussian difference. The first step in solving the key points is to compare the adjacent images in a single group. In order to find the extreme points of the Gaussian three-dimensional difference function, some points with too small pixel values ​​are removed from the Gaussian three-dimensional difference pyramid. The contrast between these pixels is low and they are unstable points. All pixel points must be compared with the points around them to determine whether their image domain and scale domain are different from those of adjacent points. The test point in the center is compared with a total of 26 pixel points of the same scale and adjacent scales to ensure that the extreme points can be successfully tested in the global three-dimensional space. The extreme points of x scales are extracted from each group. Each group of Gaussian difference pyramids has x+2 layers, and each group of Gaussian pyramids has x+3 layers of images, 3≤x≤5.

[0031] InSAR interferometric complex image high-precision registration method, further, the location of InSAR image feature points: fitting the extreme points extracted by discrete three-dimensional, interpolating the obtained discrete three-dimensional extreme points to obtain the real extreme points, and performing three-dimensional expansion of D(x) to obtain:

[0032]

[0033] Where X=(x,y,σ) T , by taking the derivative of equation 6 and making the equation equal to zero, the offset of X is calculated as:

[0034]

[0035] The value of the equation corresponding to the extreme point is:

[0036]

[0037] It is the offset relative to the interpolation center point. When its offset in any dimension is greater than 0.5, the interpolation center has shifted to its neighboring point, and the position of the current key point must be changed. At the same time, interpolation is repeatedly performed at the new position until convergence; if the number of iterations exceeds the set number or exceeds the image boundary range, such points are deleted. If the extreme value |D(x)| is too small, when |D(x)|<0.3, the extreme point is eliminated to obtain the accurate position of the feature point.

[0038] InSAR interferometric complex image high-precision registration method, further, the determination of InSAR direction angle: the local structure in the Gaussian scale image of the layer where each feature point is located is used to obtain its direction reference. The scale σ of the Gaussian scale image is known, and the scale is relative to the scale of the reference layer of the group where the Gaussian pyramid is located. The local structure is to calculate the angle and amplitude of all pixel gradients in the area with a radius of r and the feature point as the center in the Gaussian scale image. The radius r is:

[0039] r=3×1.5σ Formula 9

[0040] Where σ is the scale of the Gaussian scale image relative to the reference layer of the group, and the calculation formula for the magnitude and angle of the pixel gradient is:

[0041]

[0042]

[0043] The scale of L is the scale of each key point. After completing the key point inspection of the image, all key points contain three types of information: position, scale and direction. These information are used to obtain the corresponding improved features and collaboratively transform the feature area.

[0044] After calculating the gradient value of the feature point, the results are statistically analyzed and the amplitude corresponding to each gradient direction of the pixel in the feature point area is expressed in the form of a histogram. The histogram is smoothed with a Gaussian function to eliminate noise. The gradient direction indicated by the maximum value of the histogram is the gradient direction of the key point.

[0045] InSAR interferometric complex image high-precision registration method, further, InSAR descriptor of image feature points: partition the neighborhood of the key point, and obtain a separate vector according to the gradient information in each area. The vector is the only description of the image information in its corresponding field. The specific process of determining the feature point descriptor is as follows: first locate a descriptor region for solution, and describe the corresponding feature point on the scale image where the feature point is located. The area around the feature point is divided into d×d small blocks, and each small block is a square with a side length of 3σ, that is, there are 3σ pixels on each side of each small block, where σ is the scale of the lowest layer of the Gaussian pyramid group corresponding to the feature point, and the side length of the area around the feature point is 3σ(d+1), and there are a total of 3σ(d+1)×3σ(d+1) pixels in the area around the feature point.

[0046] The neighborhood around the obtained feature point is rotated to the main direction of the corresponding feature point with the feature as the midpoint. In order to make the square after rotation fully included in the area, the radius of this area is the distance from the midpoint of the square to the diagonal point, that is, 1 / 2 of the diagonal line, that is, formula 12:

[0047]

[0048] The coordinates of the key points after rotation become:

[0049]

[0050] Where [x',y'] T is the coordinate after the coordinate axis rotation;

[0051] All sampling points in the region are dispersed into their respective sub-neighborhoods, and the gradients in the sub-neighborhoods are dispersed into 8 angles according to their values, and their values ​​are calculated. After the region is rotated, the sampling points are assigned to a d×d sub-block in a circle with a radius of r. The gradients and directions of the sampling points in the sub-block are calculated and divided into 8 different directions. The coordinates of the sampling points in the sub-grid block after the rotation operation are expressed by formula 14:

[0052]

[0053] The sampling points obtained by formula (14) are linearly interpolated to calculate their contribution to each seed point. After obtaining the feature vector, it is normalized to eliminate the influence of illumination change on the result. After the description vector is sorted according to the scale of the feature point, the improved feature collaborative transformation feature description vector is generated.

[0054] InSAR interferometric complex image high-precision registration method, further, improve the SAR image registration of feature cooperative transformation: use polynomial fitting to find the offset polynomials of the master and slave images in range and azimuth, then resample and interfere to get the interference map, extract the amplitude of the SAR complex image, generate the amplitude map of the reference image and the input image, check and align the key points of the obtained master and slave images, the pixel offset between the master and slave images to be registered, there is not much offset and scale change in the range and azimuth, that is, the feature points of the master image are near the corresponding positions in the auxiliary image, divide the image into several areas of appropriate size for feature point matching, get the same name point corresponding to each feature point in the other image, set the appropriate critical value according to formula 15 to remove the error points caused by the error:

[0055] U={Δx, Δy|abs(Δx-median{Δx})<R, abs(Δy-median{Δy})<R} Equation 15

[0056] Where U is the set of points matched by the improved feature collaborative transformation, (△x,△y) is the offset of the same name, and R is the critical value. If the matching point satisfies Formula 15, it is considered to be a correct matching point. After removing the mismatched points, the successful matching point is obtained.

[0057] According to the least squares method, a four-parameter quadratic polynomial is fitted to these detected matching points according to Formula 16 to obtain the offset polynomial parameters of the master and slave images. The generated polynomial is used to resample the slave image to generate its corresponding coherence map. After obtaining a certain number of correct registration point pairs, further registration point pair refinement is performed, all registration point pairs are sorted by root mean square error, and the smallest 16 points are selected as fine registration points.

[0058] InSAR interferometric complex image high-precision registration method, further, the setting and solution of SPS points: by reducing the number of SPS points to eliminate local distortion and finally improve the quality of DEM;

[0059] First, the master and slave images are registered, and then the interference map I generated by the master and slave images is filtered using 16-point mean filtering using equation 17:

[0060]

[0061] Explanation of the parameters in Formula 17: The 8×2 pixels on the master and slave images are regarded as a global block M(m,n) and S(m,n) respectively. Each pair of corresponding blocks generates a pixel point in the 16-point set interference map, and in the figure M(p,q;m,n) represents the pixel in the pth row and qth column from top to bottom in M(m,n). The 16-point mean filter reduces noise by solving the average value of the 16 pixels on the interference map, that is, the generated phase map has eight times compression in the azimuth direction and two times compression in the distance direction. Then its SPS point is calculated, and the master and slave images after processing are processed by the local nonlinear registration interpolation method.

[0062] InSAR interferometric complex image high-precision registration method, further, InSAR registration operation based on SPS energy evaluation: the number of SPS points is obtained to perform local nonlinear registration based on SPS point energy. After eight iterative interpolation calculations, the master and slave images are in 1 / 8 pixel coordinate system, while the initial image is in 1 pixel coordinate system, I / 8 (m,n) represents the pixel of the interference pattern generated by the 1 / 8 pixel coordinate system, I(m,n) is the pixel of the interference pattern generated by the 1 pixel coordinate system, (m,n) defines the location in the global image of the interference pattern, (p,q) defines the local location in the 1 / 8 pixel coordinate system, and divides the master and slave images into blocks with (8×8)×(2×8)=1024 sub-pixels each, and then performs calculations to obtain I / 8 (m,n), these 1024 pixel blocks on the master and slave images are called M / 8 (m,n) and S / 8 (m,n);

[0063] First, make a local interference map of a 1 / 8 pixel coordinate system, with the pixel value I / 8 (m,n) is associated with the pixel value I(m,n) in the interference map after 16-point mean filtering, and the relationship is:

[0064]

[0065] M in formula 18 / 8 (p,q;m,n) represents M / 8 The pixel in the pth row and qth column of the (m,n) block, S / 8 (p, q; m, n) Similarly, if in block I / 8 (m,n),I / 8 (m+1,n),I / 8 (m,n+1),I / 8 There is an SPS point in (m+1,n+1) where the value of the closed-loop rotation interpolation is non-zero, that is:

[0066] |(θ / 8 (m+l,n)-θ / 8 (m,n))+(θ / 8 (m,n)-θ / 8 (m,n+1))+(θ / 8 (m+1,n+1)-θ / 8 (m+1,n))+(θ / 8 (m,n+1)-θ / 8 (m+1,n+1))|≥2π Formula 19

[0067] where θ / 8 (m,n)=arg(I / 8 (m,n)). That is, in I / 8 (m,n)-I / 8 There is local distortion in the master and slave images corresponding to the four regions (m+1, n+1). / 8 (m,n)-S / 8 One or more blocks in (m+1,n+1) are moved by a distance of 1 / 8 pixel length to reduce the distortion in the interference pattern. / 8 Point S in (m,n) / 8 (p,q;m,n) produces S / 8 new (p, q; m, n), and then check whether the number of SPS points in the generated interference graph decreases with this operation. If not, continue to move the remaining three blocks separately.

[0068] InSAR interferometric complex image high-precision registration method, further, after using iterative interferometric operation to remove SPS points, there are still some SPS points that cannot be removed. At the same time, 4|(2×2) blocks are moved and regarded as a large block. / 8 (m,n), S / 8 (m,n+1), S / 8 (m+1,n), S / 8 (m+1, n+1) perform the same moving operation, regard these four blocks as a large block, perform the above related operations to obtain the interference diagram, and further erase the SPS points. If the requirements are still not met, use a larger block of 9|(3×3) to perform the operation and check the final effect;

[0069] If the above movements cannot reduce the number of SPS points, move from the image to the other seven directions, and then for the remaining SPS points, try to move a distance of 2 / 8 pixel length, and operate in the previous way. If the moving distance reaches 8 / 8 = 1 pixel, delete all SPS points and start deleting the SPS points of the next interferogram to improve the InSAR registration accuracy.

[0070] Compared with the prior art, the innovations and advantages of this application are:

[0071] First, the present application proposes a method for effectively improving the registration accuracy in a step-by-step manner in combination with the characteristics of InSAR images. The first stage is to use an improved feature collaborative transformation method to perform InSAR image registration, use polynomial fitting to find the offset polynomials of the master and slave images in the distance and azimuth directions, then perform resampling and interference calculation to obtain an interference map, extract the amplitude of the InSAR complex image, perform key point inspection and registration on the obtained master and slave images, and divide the image into several appropriately sized areas for feature point matching; the second stage is to improve the accuracy of local image registration by combining a nonlinear InSAR registration method based on additional SPS energy evaluation on the basis of the improved feature collaborative transformation method, by eliminating the SPS points in the interference map obtained by coherent calculation after registration, and performing nonlinear registration on the image based on a large number of singular rotation points, and combining with the improved feature collaborative transformation registration method to improve the quality of the interference map generated after InSAR image registration; it has the effect of gradually improving the registration accuracy of InSAR complex images, and gradually generating high-quality interference maps and DEMs, with high image registration accuracy and low error rate, and has great practical application value;

[0072] Second, the present application proposes an InSAR image registration method based on improved feature collaborative transformation, which uses polynomial fitting to obtain the offset polynomials of the master and slave images in the distance and azimuth directions, and then resamples and interferometrically calculates the interference pattern, extracts the amplitude of the InSAR complex image, performs key point inspection and registration on the master and slave images, divides the image into several appropriately sized areas for feature point matching, and fully utilizes the phase and amplitude information of the image through scale three-dimensional image extreme value inspection, InSAR image feature point positioning, InSAR direction angle determination, image feature point InSAR descriptor, and SAR image registration, and balances high efficiency with the accuracy of the registration result. After registration, the interference pattern fringes are clearer and the speckle noise at the edge of the fringes is relatively less. The improved feature collaborative transformation does not require additional information such as orbital parameters for registration. At the same time, the time required for the improved feature collaborative transformation registration is relatively short and has higher accuracy.

[0073] Third, the present application proposes to use a nonlinear InSAR registration method based on additional SPS energy evaluation, combined with an improved feature collaborative transformation registration method to improve the accuracy of InSAR image registration. Through the setting and solution of SPS points, InSAR registration operations based on SPS energy evaluation, and nonlinear InSAR image precision registration with additional SPS, the energy of SPS points is used as an evaluation criterion in the process of local and nonlinear registration of master and slave images, which can generate higher quality interference maps. This is an additional technique for InSAR image registration, which produces higher quality interference maps and improves the signal-to-noise ratio. This method is effective in improving the accuracy of InSAR complex image registration.

[0074] Fourth, the InSAR of the present application can obtain two or even more images of the same monitored target location with a slight time difference or even at the same time, and the obtained interference pattern eliminates interference noise, so that the information on the interference pattern is only related to the location of the monitored target location, thereby making full use of the data information in the obtained SAR image, and can accurately, economically, efficiently and quickly solve the thorny problems of terrain and deformation in a large area, can adapt to various weather climates and working hours, has a wide monitoring range, and can obtain information data quickly; the SAR image data provided has a large time span, and continuous and high-precision observation information can be obtained with less cost. For image registration, which is a key factor affecting the quality of InSAR data, the present application has carried out a series of optimization and improvements, and the robustness and accuracy of InSAR technology have been greatly improved, which promotes the development of InSAR technology and expands its practical application range. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a schematic diagram of the InSAR image registration process with improved feature cooperative transformation.

[0076] Figure 2It is a schematic diagram of master and slave images for generating reference images and input images.

[0077] Figure 3 This is the effect diagram of InSAR image registration based on improved feature collaborative transformation.

[0078] Figure 4 This is the effect diagram of InSAR image registration using the coherence coefficient method.

[0079] Figure 5 It is a schematic diagram of the result of precise alignment of point pairs by improved feature collaborative transformation.

[0080] Figure 6 It is a schematic diagram of a local enlargement of the interference phase image of the registration method.

[0081] Figure 7 This is the interference result diagram after removing the flat ground phase in this application.

[0082] Figure 8 It is a schematic diagram of the fringes and phase unwrapping results of the intercepted part.

[0083] Fig. 9 The DEM model schematic diagram is restored from the entire image after least squares untangling.

[0084] Fig.10 It is a schematic diagram of the movement transformation of the master and slave images when the interference pattern and SPS point are aligned.

[0085] Fig.11 It is a flow chart of the nonlinear InSAR registration method based on additional SPS energy evaluation.

[0086] Fig.12 It is a relationship diagram between the blocks of the present application and the interference image pixels.

[0087] Fig.13 It is a graphic symbol and state definition diagram based on SPS energy evaluation.

[0088] Fig.14 It is a schematic diagram of the InSAR registration operation moving from the image process.

[0089] Fig.15 It is the InSAR registration interferogram and SPS point distribution map for SPS energy evaluation.

[0090] Fig.16 It is the DEM obtained after nonlinear InSAR registration based on additional SPS energy evaluation.

[0091] Fig.17 This is a comparison chart of the average signal-to-noise ratio and peak signal-to-noise ratio before and after InSAR registration evaluated by SPS. Specific implementation methods

[0092] The following further describes the technical solution of the InSAR interferometry complex image high-precision registration method proposed in the present application in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present application and implement it.

[0093] Synthetic Aperture Radar Interferometry (InSAR) is an important technology for obtaining surface elevation and deformation information in the field of remote sensing applications. It can not only adapt to various weather conditions and working hours, but also produce high-quality digital elevation models (DEM) to meet people's growing needs. InSAR can obtain surface elevation and deformation information over a large area, and has achieved a lot of results in monitoring surface elevation values, ground activity monitoring, and ocean current monitoring, showing great potential.

[0094] The InSAR data processing process includes InSAR complex image data registration, interferogram generation and related processing such as flattening effect, interferometric phase unwrapping, and finally elevation calculation geocoding processing. The image registration process of InSAR is to match the pixels of the common area in the master and slave complex images one by one, and then obtain the interferometric phase of the entire common area. The accuracy of the result of this step directly affects the quality of the generated interferogram and the subsequent data processing of InSAR. Therefore, InSAR complex image registration is the most critical and basic step. Generally, the registration accuracy at the pixel level must be achieved, and the subsequent generated interferogram and DEM will have better quality. However, with the development of computer vision and information processing technology, the image registration method of the existing technology can no longer meet the standard of generating high-quality interferograms and DEMs.

[0095] This application proposes a method to effectively improve the registration accuracy in a step-by-step manner in combination with the characteristics of InSAR images. The first stage is to use the improved feature collaborative transformation method to perform InSAR image registration. The second stage is to use the improved feature collaborative transformation method combined with the nonlinear InSAR registration method based on additional SPS energy evaluation to improve the accuracy of local image registration. The SPS energy is used as the registration evaluation standard. As an additional skill in the image registration process, it can effectively improve the quality of the interference map generated after InSAR image registration in combination with the improved feature collaborative transformation registration method. In order to prove the effectiveness of the two-stage method in improving InSAR complex image registration, experiments were conducted using the traditional coherence coefficient method, the InSAR image registration method based on the improved feature collaborative transformation, and the nonlinear InSAR registration method based on the additional SPS energy evaluation. By comparing the quality and signal-to-noise ratio of the generated interference map, an effective conclusion was drawn. The final experimental results show that the method proposed in this application can improve the accuracy of InSAR complex image registration, and has a progressive enhancement effect, and generates high-quality interference maps and DEMs.

[0096] 1. InSAR image registration method based on improved feature cooperative transformation

[0097] Some image registration methods in the existing InSAR technology cannot fully utilize the phase and amplitude information of the image, and cannot strike a balance between high efficiency and the accuracy of the registration results. This application proposes to use an improved feature collaborative transformation method for InSAR complex image registration, and proves the feasibility and effectiveness of the improved feature collaborative transformation InSAR complex image registration through comparative experiments.

[0098] (I) Scaled 3D Image Extreme Value Test

[0099] To construct a scaled three-dimensional Gaussian pyramid model, firstly, the InSAR image is Gaussian blurred and downsampled at different scales to obtain a group of images of different sizes, which form a pyramid-shaped structure from large to small and from bottom to top. There are a total of O groups, and each group is divided into S layers. The initial image is placed at the bottom of the pyramid, that is, the first layer. The image obtained after each downsampling is used as the next layer of the pyramid. The pyramid has a total of n layers, and the total number of layers is determined by the size of the initial InSAR image and the size of the highest layer image. The calculation formula is:

[0100] n=log 2 {min(M,N)}-t, t∈[0,log 2 {min(M, N)} Formula 1

[0101] m, n are the sizes of the initial image, t is the minimum logarithmic value of the highest layer image, and the three-dimensional scale L(x, y, σ) of the target image is defined as:

[0102] L(x,y,σ)=G(x,y,σ)*I(x,y) Equation 2

[0103] G(x,y,σ) is a scale-variable Gaussian function, and I(x,y) is the initial image:

[0104]

[0105] Where (x, y) is the coordinate position, parameter σ represents the smoothness of the image, different σ values ​​represent different scales of the image, and show different characteristics of the image. The resolution of I(x, y) in formula 2 is infinite, σ = 0, but the image with scale zero does not exist. The initial scale of the image is set to 0.5, and the small scale σ 1 Generate large scale σ 2 The process is:

[0106]

[0107] Gaussian 3D difference is used to obtain image key points at different scales, and images of adjacent Gaussian 3D scales are subtracted to obtain Equation 5:

[0108]

[0109] D(x,y,σ) is the response value image, k is the three-dimensional multiple of the adjacent scale, and the difference of the upper and lower adjacent images in the Gaussian pyramid is obtained to obtain the Gaussian difference map. The key points to be sought are composed of the local extreme points of the three-dimensional scale of the Gaussian difference. The first step in solving the key points is to compare the adjacent images in a single group. In order to find the extreme points of the Gaussian three-dimensional difference function, some points with too small pixel values ​​are removed from the Gaussian three-dimensional difference pyramid. The contrast between these pixels is low and they are unstable points. All pixel points must be compared with the points around them to determine whether their image domain and scale domain are different from those of adjacent points. The test point in the center is compared with a total of 26 pixel points of the same scale and adjacent scales to ensure that the extreme points can be successfully tested in the global three-dimensional space. The extreme points of x scales are extracted from each group. Each group of Gaussian difference pyramids has x+2 layers, and each group of Gaussian pyramids has x+3 layers of images, 3≤x≤5.

[0110] (II) Positioning of InSAR image feature points

[0111] In the process of improving feature collaborative transformation, it is necessary to obtain accurate scale and position information of feature points, that is, to achieve sub-pixel accuracy. Therefore, the extreme points extracted by discrete three-dimensional model must be fitted, and the real extreme points are obtained by interpolation. D(x) is scaled three-dimensionally to obtain:

[0112]

[0113] Where X=(x,y,σ) T , by taking the derivative of equation 6 and making the equation equal to zero, the offset of X is calculated as:

[0114]

[0115] The value of the equation corresponding to the extreme point is:

[0116]

[0117] It is the offset relative to the interpolation center point. When its offset in any dimension is greater than 0.5, the interpolation center has shifted to its neighboring point, and the position of the current key point must be changed. At the same time, interpolation is repeatedly performed at the new position until convergence; if the number of iterations exceeds the set number or exceeds the image boundary range, such points are deleted. If the extreme value |D(x)| is too small, when |D(x)|<0.3, the extreme point is eliminated to obtain the accurate position of the feature point.

[0118] 3. Determination of InSAR direction angle

[0119] The local structure in the Gaussian scale image of the layer where each feature point is located is used to obtain its direction reference. The scale σ of the Gaussian scale image is known, and the scale is relative to the scale of the reference layer of the group where the Gaussian pyramid is located. The local structure is to calculate the angle and amplitude of all pixel gradients in the area with a radius of r and the feature point as the center in the Gaussian scale image. The radius r is:

[0120] r=3×1.5σ Formula 9

[0121] Where σ is the scale of the Gaussian scale image relative to the reference layer of the group, and the calculation formula for the magnitude and angle of the pixel gradient is:

[0122]

[0123]

[0124] The scale of L in the formula is the scale of each key point. The key point inspection of the image is completed. All key points contain three types of information: position, scale and direction. The corresponding improved features are used to derive the collaborative transformation feature area.

[0125] After calculating the gradient value of the feature point, the results are statistically analyzed and the amplitude corresponding to each gradient direction of the pixel in the feature point area is expressed in the form of a histogram. In order to eliminate the influence of noise on the gradient direction angle, the histogram is smoothed with a Gaussian function. The gradient direction indicated by the maximum value of the histogram is the gradient direction of the key point.

[0126] (IV) InSAR descriptors of image feature points

[0127] The neighborhood of the key point is partitioned, and a separate vector is obtained according to the gradient information in each area. The vector is the only description of the image information in its corresponding field. The specific process of determining the feature point descriptor is as follows: first locate a descriptor region for solution, and describe the corresponding feature point on the scale image where the feature point is located. The area around the feature point is divided into d×d small blocks, each of which is a square with a side length of 3σ, that is, there are 3σ pixels on each side of each small block, where σ is the scale of the lowest layer of the Gaussian pyramid group corresponding to the feature point, the side length of the area around the feature point is 3σ(d+1), and there are a total of 3σ(d+1)×3σ(d+1) pixels in the area around the feature point.

[0128] In order to keep the feature point rotationally invariant, the neighborhood around the feature point is rotated to the main direction of the corresponding feature point with the feature as the midpoint. In order to make the square after rotation fully included in the area, the radius of this area is the distance from the midpoint of the square to the diagonal point, that is, 1 / 2 of the diagonal, that is, formula 12:

[0129]

[0130] The coordinates of the key points after rotation become:

[0131]

[0132] Where [x',y'] T is the coordinate after the coordinate axis rotation;

[0133] All sampling points in the region are dispersed into their respective sub-neighborhoods, and the gradients in the sub-neighborhoods are dispersed into 8 angles according to their values, and their values ​​are calculated. After the region is rotated, the sampling points are assigned to a d×d sub-block in a circle with a radius of r. The gradients and directions of the sampling points in the sub-block are calculated and divided into 8 different directions. The coordinates of the sampling points in the sub-grid block after the rotation operation are expressed by formula 14:

[0134]

[0135] The sampling points obtained by formula (14) are linearly interpolated to calculate their contribution to each seed point. After obtaining the feature vector, it is normalized to eliminate the influence of illumination change on the result. After the description vector is sorted according to the scale of the feature point, the improved feature collaborative transformation feature description vector is generated.

[0136] 5. Improved Feature Co-Transformation for SAR Image Registration

[0137] The two images registered by InSAR do not have much displacement and scale change, and the position difference of key points on the master and slave images is not too large. The block matching mode is used to improve the operation efficiency and accuracy, and improve the matching process of the feature collaborative transformation method. Figure 1 shown.

[0138] After the registration is completed using the improved feature collaborative transformation, polynomial fitting is used to obtain the offset polynomials of the master and slave images in the distance and azimuth directions, and then resampling and interference calculation are performed to obtain the interference map.

[0139] The amplitude of the SAR complex image is extracted to generate the amplitude map of the reference image and the input image, such as Figure 2 shown.

[0140] The obtained master and slave images are inspected and registered for key points. The pixel offset between the master and slave images to be registered does not have too much offset and scale change in the distance and azimuth directions, that is, the feature points of the master image are near the corresponding positions in the auxiliary image. Therefore, a block processing method is adopted to divide the image into several areas of appropriate size for feature point matching. This can greatly improve the registration efficiency and reduce the mismatching rate. After obtaining the corresponding points of the same name for each feature point in the other image, set a suitable critical value according to formula 15 to remove the erroneous points caused by errors:

[0141] U={Δx, Δy|abs(Δx-median{Δx})<R, abs(Δy-median{Δy})<R} Equation 15

[0142] Where U is the set of points matched by the improved feature collaborative transformation, (△x,△y) is the offset of the same name, and R is the critical value. If the matching point satisfies Formula 15, it is considered to be a correct matching point. After removing the mismatched points, the successful matching point is obtained.

[0143] According to the least squares method, a four-parameter quadratic polynomial is fitted to these detected matching points according to Formula 16 to obtain the offset polynomial parameters of the master and slave images.

[0144] The slave images are resampled using the polynomials generated by the method of the present application and the coherence coefficient method to generate their corresponding coherence maps. The higher the coherence coefficient, the better the quality of the interference data. The registered master and slave images are applied to generate the coherence map.

[0145] like Figure 3 and Figure 4As shown. The brighter the position in the coherence map, the higher the coherence coefficient at that location, that is, the better the quality of the interference phase, while the darker position may be caused by the interference phase discontinuity caused by phase noise or the interference phase discontinuity caused by the drastic terrain undulation. From the two coherence maps, it can be seen that the coherence map obtained by the improved feature collaborative transformation matching has more highlight areas than the coherence coefficient. Similarly, the coherence coefficient values ​​obtained by the statistical comparison method can be made to make a correlation coefficient distribution histogram for comparison to evaluate the accuracy of the matching method.

[0146] From the above two figures, we can see that the coherence coefficient values ​​obtained by the coherence coefficient method are mostly distributed between 0.5 and 0.7, while the coherence coefficient values ​​obtained by the improved feature collaborative transformation are mostly distributed between 0.7 and 0.9. The higher the coherence coefficient, the better the registration effect, and the higher the quality of the subsequent interference pattern. Therefore, after comparison, it can be concluded that the improved feature collaborative transformation registration effect is better than the traditional method. The inaccuracy of the coherence coefficient method in the registration process will lead to serious decoherence.

[0147] After obtaining a certain number of correct registration point pairs, further refinement of the registration point pairs is performed. All the registration point pairs are sorted by root mean square error, and the smallest 16 points are selected as fine registration points. The detected registration points are as follows: Figure 5 .

[0148] (VI) Experimental comparison of improved feature collaborative transformation registration method and coherence coefficient method

[0149] 1. Generate interference patterns and remove flat ground effects

[0150] By conjugate multiplying the complex data of the master and slave images after the coherence coefficient and the improved feature collaborative transformation, and extracting the interference phase, an interference phase map is obtained. The phase value is displayed by the grayscale value, where a white area represents a phase value of π and a black area represents a phase value of -π.

[0151] The interference phase image of the registration method is locally enlarged, such as Figure 6 By comparing the fringes of the two interference phase images, it is found that in the interference phase image obtained by the coherence coefficient method, there is a lot of noise at the edge of the fringes, making the fringes blurry; while after the improved feature collaborative transformation method is used to obtain the interference image, the interference fringes are clearer, the speckle noise at the edge of the fringes is relatively small, and the registration result is better.

[0152] These interference patterns are generated by superimposing the flat ground and ground elevation information. Therefore, after obtaining the interference phase pattern, the flat ground effect is removed to obtain the real phase. The flat ground phase is estimated by the frequency of the fringe and then removed according to formula 16:

[0153]

[0154] In formula 16, x n is the image intensity, is the estimated value of fringe frequency, the column direction is the distance direction; m, n represent the position of the image in the row and column directions respectively, m=0,1,…,M-1, n=0,1,…,N-1.

[0155] The interference pattern after removing the flat ground phase is as follows Figure 7 As shown, compared Figure 6 It can be found that the fringes of the initial interference phase diagram are almost all produced by the flat ground phase, which is parallel to the azimuth direction and extends in the distance direction in a level-by-level distribution, which is consistent with the principle of the longevity of the flat ground phase. Once the flat ground phase is removed, the trend of terrain and elevation changes can be clearly seen from the interference diagram.

[0156] During the registration process, misaligned points will form phase noise on the interference phase image, which will cause errors in the subsequent phase unwrapping and propagate to the entire phase image, seriously affecting the reliability of the phase unwrapping results. Therefore, it is necessary to filter out phase noise points as much as possible to improve the signal-to-noise ratio of the image.

[0157] 2. Phase unwrapping comparison

[0158] The phase value of the interference pattern is obtained by calculating the inverse tangent function of the phase of the complex image. It is only the main phase value of the interference. The phase recovery is required to obtain the real phase value. The resolution of the initial interference pattern is high. It is very time-consuming to unwrap the entire interference pattern, and it is not convenient to compare and analyze the unwrapping results. Therefore, part of the interference fringe pattern is cut from the initial interference pattern and the least squares method is used for phase unwrapping. Since InSAR has very high requirements for image registration accuracy, the accuracy of registration directly affects the quality of the interference pattern, which has a great impact on the unwrapping results. The branch cutting method is used for unwrapping and the final results are compared. The results of the intercepted fringe and phase unwrapping are shown in the figure. Figure 8 As shown:

[0159] 3. Repair the DEM model

[0160] After obtaining the true interference phase value, the interference phase value after removing the flat ground phase is proportional to the surface elevation. The unwrapped phase map reflects the relative height relationship between the points on the surface, that is, the terrain change. The DEM model is restored based on the entire map after unwrapping by the least squares method, such as Fig. 9 .

[0161] Compared with the traditional coherence coefficient method, the InSAR image registered by the improved feature collaborative transformation method has higher quality interference fringes. Moreover, the improved feature collaborative transformation does not require additional information such as orbital parameters. At the same time, the time required for improved feature collaborative transformation registration is relatively short and has higher accuracy.

[0162] 2. Nonlinear InSAR registration method based on additional SPS energy evaluation

[0163] To obtain an accurate digital elevation model (DEM), a high-quality interferogram must be obtained. However, the result is often affected by a large number of singular points (SPS) on the interferogram. These SPS points come from the registration results with low precision. However, there is currently no particularly effective method to eliminate such singular points. Although the prior art has methods for dealing with SPS points in the image registration process, the effect is not ideal. This application proposes to use a nonlinear InSAR registration method based on additional SPS energy evaluation, combined with an improved feature collaborative transformation registration method to improve the InSAR image registration accuracy. This method can generate a higher quality interferogram to produce a higher precision DEM, and experiments show that this method can obtain a DEM with higher signal-to-noise.

[0164] The actual interferogram contains a large number of SPS non-zero rotation points. Most of the SPS will not be able to recover a high-precision DEM. There are two origins for the generation of SPS points: one is the violation of the sampling theorem, which is caused by the phase aliasing phenomenon due to the rough terrain information; the other is due to the difference between the master and slave observation orbits. When observing the same target location from two slightly different perspectives, the reflection, scattering, and perspective values ​​are different, resulting in local phase distortion in the interferogram. Usually, the registration is achieved by maximizing the large-scale image amplitude cross-correlation and maximizing the complex amplitude correlation at a small scale. When solving the correlation, the average is calculated within a certain range to reduce the noise contained. However, the averaging operation in a large area will lose some points used to eliminate distortion in image registration. This cost will limit the accuracy of image registration. The local deformation caused by reflection, scattering and perspective differences will generate SPS points in the cross-correlation process. This application uses the quantity of SPS points as an evaluation criterion in the process of local and nonlinear registration of master and slave images to produce higher quality interferograms. This is an additional technique for InSAR image registration to improve the signal-to-noise ratio.

[0165] (I) Setting and solving SPS points

[0166] By cutting out a small piece of the phase interference pattern for experimental analysis, it is found that processing the master and slave images corresponding to the small piece of the interference pattern can effectively reduce the number of SPS points; Fig.10 It shows how the interference pattern and SPS points change as the image is moved from the best registration position with the main image to the left or right by an integer multiple of 1 / 8 pixel during image registration. The experiment shows that many SPSs move, appear or disappear, but the fringes of the interference pattern only change slightly. In the interference pattern, white SPS points represent clockwise non-zero rotation points, while black ones represent counter-clockwise non-zero rotation points.

[0167] The appearance and disappearance of SPS points on the interferogram using the above method is more local than the result produced by global correlation calculation. The global figure of the interferogram obtained by adding a local nonlinear registration process to the traditional registration method does not change much, but this local precision registration can improve the quality of the local interferogram. In an idealized interferogram without distortion, there are no SPS points. This application eliminates local distortion and finally improves the quality of DEM by reducing the number of SPS points.

[0168] Fig.11 This is a flow chart of the nonlinear InSAR registration method based on additional SPS energy evaluation. First, the master and slave images are registered, and then the interference map I generated by the master and slave images is filtered by 16-point mean using Equation 17:

[0169]

[0170] The parameters in Equation 17 are Fig.12 Further explanation: The 8×2 pixels on the master and slave images are regarded as a global block M(m,n) and S(m,n), and each pair of corresponding blocks is generated in Fig.12 (a), and M(p,q;m,n) represents the pixel in the pth row and qth column from top to bottom in M(m,n). The 16-point mean filter reduces the noise by solving the average value of the 16 pixels on the interference pattern, that is, the generated phase map has eight times compression in the azimuth direction and two times compression in the distance direction. Then its SPS point is calculated, and the master and slave images are processed by the local nonlinear registration interpolation method.

[0171] (II) InSAR registration operation based on SPS energy evaluation

[0172] The number of SPS points is obtained to perform local nonlinear registration based on the SPS point quantity. Fig.12 For illustration, the master and slave images after eight iterations of interpolation calculation are in a 1 / 8 pixel coordinate system, while the initial image is in a 1 pixel coordinate system. Fig.13 List the definitions of each noun, / 8 (m,n) represents the pixel of the interference pattern generated by the 1 / 8 pixel coordinate system, I(m,n) is the pixel of the interference pattern generated by the 1 pixel coordinate system, (m,n) defines the location in the global image of the interference pattern, (p,q) defines the local location in the 1 / 8 pixel coordinate system, and divides the master and slave images into blocks with (8×8)×(2×8)=1024 sub-pixels each, and then performs calculations to obtain I / 8 (m,n), these 1024 pixel blocks on the master and slave images are called M / 8 (m,n) and S / 8 (m,n).

[0173] First, make a local interference map of a 1 / 8 pixel coordinate system, with the pixel value I / 8 (m,n) is associated with the pixel value I(m,n) in the interference map after 16-point mean filtering, and the relationship is:

[0174]

[0175] M in formula 18 / 8 (p,q;m,n) represents M / 8 The pixel in the pth row and qth column of the (m,n) block, S / 8 (p, q; m, n) Similarly, if in block I / 8 (m,n),I / 8 (m+1,n),I / 8 (m,n+1),I / 8 There is an SPS point in (m+1,n+1), such as Fig.12 As shown in (c), the value of the closed-loop rotation interpolation is non-zero, that is:

[0176] |(θ / 8 (m+l,n)-θ / 8 (m,n))+(θ / 8 (m,n)-θ / 8 (m,n+1))+(θ / 8 (m+1,n+1)-θ / 8 (m+1,n))+(θ / 8 (m,n+1)-θ / 8 (m+1,n+1))|≥2π Formula 19

[0177] where θ / 8 (m,n)=arg(I / 8 (m,n)). That is, in I / 8 (m,n)-I / 8 There is local distortion in the master and slave images corresponding to the four regions (m+1, n+1). / 8 (m,n)-S / 8 One or more blocks in (m+1,n+1) are used to reduce the distortion in the interference pattern. Fig.12 As shown in the figure, the distance of 1 / 8 pixel length moves S / 8 Point S in (m,n) / 8 (p,q;m,n) produces S / 8 new (p, q; m, n), and then check whether the number of SPS points in the generated interference pattern has decreased with this operation. If not, continue as Fig.14As shown, the remaining three blocks are moved separately.

[0178] After using the iterative interference operation described above to remove SPS points, it is found that some SPS points cannot be removed. At the same time, the 4|(2×2) blocks are moved and regarded as a large block. / 8 (m,n), S / 8 (m,n+1), S / 8 (m+1,n), S / 8 (m+1,n+1) perform the same moving operation, regard these four blocks as a large block, perform the above-mentioned related operations to obtain the interference diagram, and further erase the SPS points. If the requirements are still not met, use a larger block of 9|(3×3) to perform the operation and check the final effect.

[0179] If the above moves cannot reduce the SPS points, the image will be moved according to Fig.14 As shown, move to the other seven directions, and then for the remaining SPS points, try to move a distance of 2 / 8 pixel length, and operate in the previous way. If the moving distance reaches 8 / 8 = 1 pixel, delete all SPS points and start deleting the SPS points of the next interferogram to improve the InSAR registration accuracy.

[0180] (III) Nonlinear InSAR image registration with SPS

[0181] right Figure 7 The screenshots are taken in the experiment according to the above steps. The local interference map and the extracted SPS points before the nonlinear InSAR registration method based on additional SPS energy evaluation are shown in Figure 2. Fig.15 As shown in (a), the figure on the right is the SPS points, the total number is 1016. After the first mobile calculation in the above experimental steps, 59% of the SPS points are erased, leaving 401 SPS points, as shown in Fig.15 (b) After the second round of calculation, there are 325 SPS points left. In the subsequent single block movement calculation, no SPS points are eliminated. Therefore, the next three blocks are moved together, that is, the four blocks mentioned above are moved together for subsequent calculations. The interference diagram obtained has only 185 SPS points left, as shown in Fig.15 (d) Even 9 blocks can be moved together, and the resulting interference pattern has even fewer SPS points.

[0182] In summary, the experimental results show that the nonlinear InSAR registration method based on additional SPS energy evaluation can effectively reduce the SPS points on the interferogram. In the example, the erasure rate reached 81.91%, which greatly improved the quality of the interferogram. Figure 7 By visually comparing the initial interferogram of the detector with the original one, it is found that the interferogram fringes are significantly clearer after the nonlinear InSAR registration method based on additional SPS energy evaluation is adopted.

[0183] Select and Figure 7 Phase unwrapping is performed on the same part, and finally the elevation is restored after phase unwrapping with the least squares method. Fig.16 .

[0184] After obtaining the DEM, the average signal-to-noise ratio and peak signal-to-noise ratio before and after the nonlinear InSAR registration method based on the additional SPS energy evaluation are compared to intuitively reflect its role, such as Fig.17 . Fig.17 It can be intuitively shown that the nonlinear InSAR registration method based on additional SPS energy evaluation can simultaneously improve the average signal-to-noise ratio and peak signal-to-noise ratio of DEM, which illustrates the effectiveness of the nonlinear InSAR registration method based on additional SPS energy evaluation in improving the accuracy of InSAR complex image registration.

Claims

1. High-precision registration method for complex images of InSAR interferometry, It is characterized in that In combination with the characteristics of InSAR images, a method for effectively improving the registration accuracy is proposed step by step. The first stage is to use the improved feature collaborative transformation method to perform InSAR image registration, use polynomial fitting to find the offset polynomials of the master and slave images in the distance and azimuth directions, then perform resampling and interference calculation to obtain the interference map, extract the amplitude of the InSAR complex image, check and align the key points of the obtained master and slave images, and adopt a block processing method to divide the image into several areas of appropriate size for feature point matching; the second stage is to improve the accuracy of local image registration by combining the nonlinear InSAR registration method based on the additional SPS energy evaluation on the basis of the improved feature collaborative transformation method, use the SPS energy as the registration evaluation standard, eliminate the SPS points in the interference map obtained by coherent calculation after registration, perform nonlinear registration on the image based on a large number of singular rotation points, and combine with the improved feature collaborative transformation registration method to improve the quality of the interference map generated after InSAR image registration; the method proposed in this application has the effect of gradually improving the registration accuracy of InSAR complex images, and gradually generates high-quality interference maps and DEMs; In the first stage, the InSAR image registration method based on improved feature cooperative transformation is as follows: first, the scale three-dimensional image extreme value test, second, the location of InSAR image feature points, third, the determination of InSAR direction angle, fourth, the InSAR descriptor of image feature points, and fifth, the SAR image registration based on improved feature cooperative transformation; In the second stage, the nonlinear InSAR registration method based on additional SPS energy evaluation is as follows: first, the setting and solution of SPS points; second, the InSAR registration operation based on SPS energy evaluation; and third, the nonlinear InSAR image precise registration with additional SPS.

2. According to claim 1, the InSAR interferometry complex image high-precision registration method, It is characterized in that Scaled 3D image extreme value test: Construct a scaled 3D Gaussian pyramid model. First, perform Gaussian blurring and downsampling on the InSAR image at different scales to obtain a group of images of different sizes, which form a pyramid-shaped structure from large to small and from bottom to top. There are 0 groups in total, and each group is divided into S layers. The initial image is placed at the bottom of the pyramid, that is, the first layer. The image obtained after each downsampling is used as the next layer of the pyramid. The pyramid has a total of n layers, and the total number of layers is determined by the size of the initial InSAR image and the size of the highest layer image. The calculation formula is: n = log 2 {min(M, N)} - t, t ∈ [0, log 2 {min(M, N)} Equation 1 m, n are the sizes of the initial image, t is the minimum logarithmic value of the highest layer image, and the three-dimensional scale L(x, y, σ) of the target image is defined as: L(x,y,σ)=G(x,y,σ)*I(x,y) Equation 2 G(x, y, σ) is a scale-variable Gaussian function, and I(x, y) is the initial image: Where (x, y) is the coordinate position, parameter σ represents the smoothness of the image, different σ values ​​represent different scales of the image, and show different characteristics of the image. The resolution of I(x, y) in formula 2 is infinite, σ = 0, but the image with scale zero does not exist. The initial scale of the image is set to 0.5, and the small scale σ 1 Generate large scale σ 2 The process is: Gaussian 3D difference is used to obtain image key points at different scales, and images of adjacent Gaussian 3D scales are subtracted to obtain Equation 5: D(x,y,σ) is the response value image, k is the three-dimensional multiple of the adjacent scale, and the difference of the upper and lower adjacent images in the Gaussian pyramid is obtained to obtain the Gaussian difference map. The key points to be sought are composed of the local extreme points of the three-dimensional scale of the Gaussian difference. The first step in solving the key points is to compare the adjacent images in a single group. In order to find the extreme points of the Gaussian three-dimensional difference function, some points with too small pixel values ​​are removed from the Gaussian three-dimensional difference pyramid. The contrast between these pixels is low and they are unstable points. All pixel points must be compared with the points around them to determine whether their image domain and scale domain are different from those of adjacent points. The test point in the center is compared with a total of 26 pixel points of the same scale and adjacent scales to ensure that the extreme points can be successfully tested in the global three-dimensional space. The extreme points of x scales are extracted from each group. Each group of Gaussian difference pyramids has x+2 layers, and each group of Gaussian pyramids has x+3 layers of images, 3≤x≤5.

3. According to claim 1, the InSAR interferometry complex image high-precision registration method, It is characterized in that Positioning of InSAR image feature points: Fit the extreme points extracted by discrete three-dimensional, interpolate the obtained discrete three-dimensional extreme points to obtain the real extreme points, and expand D(x) in three dimensions to obtain: Where X=(x,y,σ) T , by taking the derivative of equation 6 and making the equation equal to zero, the offset of X is calculated as: The value of the equation corresponding to the extreme point is: It is the offset relative to the interpolation center point. When its offset in any dimension is greater than 0.5, the interpolation center has shifted to its neighboring point, and the position of the current key point must be changed. At the same time, interpolation is repeatedly performed at the new position until convergence; if the number of iterations exceeds the set number or exceeds the image boundary range, such points are deleted. If the extreme value |D(x)| is too small, when |D(x)|<0.3, the extreme point is eliminated to obtain the accurate position of the feature point.

4. According to claim 1, the InSAR interferometry complex image high-precision registration method, It is characterized in that Determination of InSAR direction angle: The local structure in the Gaussian scale image of the layer where each feature point is located is used to obtain its direction reference. The scale σ of the Gaussian scale image is known, and the scale is relative to the scale of the reference layer of the group where the Gaussian pyramid is located. The local structure is to calculate the angle and amplitude of all pixel gradients in the area with a radius of r and the feature point as the center in the Gaussian scale image. The radius r is: r=3×1.5σ Formula 9 Where σ is the scale of the Gaussian scale image relative to the reference layer of the group, and the calculation formula for the magnitude and angle of the pixel gradient is: The scale of L is the scale of each key point. After completing the key point inspection of the image, all key points contain three types of information: position, scale and direction. These information are used to obtain the corresponding improved features and collaboratively transform the feature area. After calculating the gradient value of the feature point, the results are statistically analyzed and the amplitude corresponding to each gradient direction of the pixel in the feature point area is expressed in the form of a histogram. The histogram is smoothed with a Gaussian function to eliminate noise. The gradient direction indicated by the maximum value of the histogram is the gradient direction of the key point.

5. According to claim 1, the InSAR interferometry complex image high-precision registration method, It is characterized in that InSAR Descriptor of Image Feature Points: Partition the neighborhood of the key point, and obtain a separate vector according to the gradient information in each region. This vector is the unique description of the image information in its corresponding neighborhood. The specific process of determining the feature point descriptor is as follows: First, locate a region for solving the descriptor, and perform the corresponding feature point description on the image at the scale where the feature point is located. Divide the area around the feature point into d×d small blocks, and each small block is a square with a side length of 3σ, that is, there are 3σ pixel points on each side of each small block, where σ is the scale size of the lowest layer of the Gaussian pyramid group corresponding to the feature point. The side length of the area around the feature point is 3σ(d + 1), and there are a total of 3σ(d + 1)×3σ(d + 1) pixel points in the area around the feature point; Rotate the obtained neighborhood around the feature point to the main direction of the corresponding feature point. To ensure that the rotated square is completely included in the area, make the radius of this area the distance from the midpoint of the square to the diagonal point, that is, 1 / 2 of the diagonal, namely Equation 12: The coordinates of the rotated key point become: Where [x',y'] T is the coordinate after the coordinate axis rotation; Disperse all the sampling points in the area to their respective sub-neighborhoods, and disperse the gradients in the sub-neighborhoods to 8 angles according to their values to obtain their values. The sampling points are assigned to a d×d sub-block within a circle with a radius of r after the area is rotated. Calculate the gradients and directions of the sampling points that will change within the sub-block, and divide them into 8 different directions. The coordinates of the sampling points within the sub-square block after the rotation operation are expressed by Equation 14: Linearly interpolate the sampling points obtained from Equation 14, calculate their contribution values to each seed point, and after obtaining the feature vector, perform a normalization operation to eliminate the influence of illumination changes on the result. After sorting the description vectors according to the feature point scale, generate an improved feature co-transformation feature description vector.

6. The high-precision registration method for InSAR interferometric complex images according to claim 1, characterized in that SAR image registration with improved feature co-transformation: Use polynomial fitting to find the offset polynomials of the master and slave images in the range and azimuth directions respectively, and then perform resampling and interference calculation to obtain the interferogram. Extract the amplitude of the SAR complex image to generate the amplitude maps of the reference image and the input image. Check and register the key points of the obtained master and slave images. The pixel offset between the master and slave images to be registered has little offset and scale change in the range and azimuth directions, that is, the feature points of the master image are near the corresponding positions in the slave image. Divide the image into several appropriately sized regions for feature point matching. After obtaining the corresponding homologous points of each feature point in the other image, set an appropriate critical value according to Equation 15 to remove the error-caused wrong points: U={Δr,Δy|abs(Δx-median{Δx})<R,abs(Δy-median{Δy})<R} Equation 15 In the formula, U is the set of points matched by the improved feature co-transformation, (△x,△y) is the offset of the homologous points, and R is the critical value. If the matching points satisfy Equation 15, they are considered correct matching points. After removing the mis-matched points, successful matching points are obtained; According to the least squares method, a four-parameter quadratic polynomial is fitted to these detected matching points according to Formula 16 to obtain the offset polynomial parameters of the master and slave images. The generated polynomial is used to resample the slave image to generate its corresponding coherence map. After obtaining a certain number of correct registration point pairs, further registration point pair refinement is performed, all registration point pairs are sorted by root mean square error, and the smallest 16 points are selected as fine registration points.

7. The high-precision registration method for complex images measured by InSAR interferometry according to claim 1, It is characterized in that Setting and solving SPS points: Eliminate local distortion and improve the quality of DEM by reducing the number of SPS points; First, the master and slave images are registered, and then the interference map I generated by the master and slave images is filtered using 16-point mean filtering using equation 17: Explanation of the parameters in Formula 17: The 8×2 pixels on the master and slave images are regarded as a global block M(m,n) and S(m,n) respectively. Each pair of corresponding blocks generates a pixel point in the 16-point set interference map, and in the figure M(p,q;m,n) represents the pixel in the pth row and qth column from top to bottom in M(m,n). The 16-point mean filter reduces noise by solving the average value of the 16 pixels on the interference map, that is, the generated phase map has eight times compression in the azimuth direction and two times compression in the distance direction. Then its SPS point is calculated, and the master and slave images after processing are processed by the local nonlinear registration interpolation method.

8. According to claim 1, the InSAR interferometry complex image high-precision registration method, It is characterized in that InSAR registration operation based on SPS energy evaluation: The number of SPS points is obtained for local nonlinear registration based on SPS point energy. After eight iterative interpolation calculations, the master and slave images are in a 1 / 8 pixel coordinate system, while the initial image is in a 1 pixel coordinate system. / 8 (m, n) represents the pixel of the interference pattern generated by the 1 / 8 pixel coordinate system, I(m, n) is the pixel of the interference pattern generated by the 1 pixel coordinate system, (m, n) defines the location in the global image of the interference pattern, and (p, q) defines the local location in the 1 / 8 pixel coordinate system. The master and slave images are divided into blocks with (8×8)×(2×8)=1024 sub-pixels each, and then the calculation is performed to obtain I / 8 (m, n), these 1024 pixel blocks on the master and slave images are called M / 8 (m, n) and S / 8 (m, n); First, make a local interference map of a 1 / 8 pixel coordinate system, with the pixel value I / 8 (m, n) is associated with the pixel value I(m, n) in the interference map after 16-point mean filtering, and the relationship is: M in formula 18 / 8 (p, q; m, n) represents M / 8 The pixel in the pth row and qth column of the (m, n) block, S / 8 (p, q; m, n) Similarly, if in block I / 8 (m, n), I / 8 (m+1, n), I / 8 (m, n+1), I / 8 There is an SPS point in (m+1, n+1) where the value of the closed-loop rotation interpolation is non-zero, that is: |(θ / 8 (m + 1, n) - θ / 8 (m, n)) + (θ / 8 (m, n) - θ / 8 (m, n + 1)) + (θ / 8 (m + 1, n + l) - θ / 8 (m + 1, n)) + (θ / 8 (m, n + 1) - θ / 8 (m + 1, n + 1))| ≥ 2π Equation 19 where θ / 8 (m,n)=arg(I / 8 (m, n)). That is, in I / 8 (m, n)-I / 8 There is local distortion in the master and slave images corresponding to the four regions (m+1, n+1). / 8 (m, n)-S / 8 One or more blocks in (m+1, n+1) are moved by a distance of 1 / 8 pixel length to reduce the distortion in the interference pattern. / 8 Point S in (m, n) / 8 (p, q; m, n) produces S / 8 new (p, q; m, n), and then check whether the number of SPS points in the generated interference graph decreases with this operation. If not, continue to move the remaining three blocks separately.

9. The high-precision registration method for complex images measured by InSAR interferometry according to claim 8, It is characterized in that After using iterative interference operation to remove SPS points, some SPS points still cannot be removed. At the same time, 4|(2×2) blocks are moved and regarded as a large block. / 8 (m, n), S / 8 (m, n+1), S / 8 (m+1, n), S / 8 (m+1, n+1) perform the same moving operation, regard these four blocks as a large block, perform the above related operations to obtain the interference diagram, and further erase the SPS points. If the requirements are still not met, use a larger block of 9|(3×3) to perform the operation and check the final effect; If the above movements cannot reduce the number of SPS points, move from the image to the other seven directions, and then for the remaining SPS points, try to move a distance of 2 / 8 pixel length, and operate in the previous way. If the moving distance reaches 8 / 8 = 1 pixel, delete all SPS points and start deleting the SPS points of the next interferogram to improve the InSAR registration accuracy.