Method and apparatus for matching corresponding points based on Gaofen-2 satellite remote sensing imagery

Abstract

This invention discloses a method and apparatus for matching corresponding points based on Gaofen-2 satellite remote sensing imagery. The method includes: performing grayscale processing on the Gaofen-2 satellite remote sensing imagery and its matching reference base map to obtain a processed image; performing convolution operations on the processed imagery using a Gaussian function to obtain the image scale space corresponding to the processed imagery, and performing extremum detection on the image scale space to obtain a feature point set; processing the feature point set according to the distribution characteristics of the gradient directions of the neighboring pixels of the feature points to obtain SIFT feature vectors; obtaining matching feature points based on the calculated ratio of the nearest neighbor distance to the second nearest neighbor distance between the SIFT feature vectors and a set threshold; and determining the corresponding points of the Gaofen-2 satellite remote sensing imagery based on the matching feature points. This invention can improve the reliability and accuracy of image matching and effectively reduce the workload of measuring corresponding points.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and apparatus for matching corresponding points based on Gaofen-2 satellite remote sensing images. Background Technology

[0002] With the continuous development of remote sensing technology and the increasing number of remote sensing information sources, the application fields of remote sensing technology have become more and more extensive, and the relationship between remote sensing technology and the national economy and ecological protection has become closer and closer. Remote sensing has played a significant role in land resource surveys, ecological environment monitoring, agricultural monitoring and crop yield estimation, disaster forecasting and disaster assessment, marine environmental surveys, as well as activities closely related to daily life such as weather forecasting, air quality monitoring, electronic maps and navigation.

[0003] Among remote sensing images of various resolutions, high spatial resolution images (referred to as "high-resolution images") have become an important data foundation for tasks such as land cover classification, target extraction and identification, and change detection because they contain rich information on the texture, shape, structure, and neighborhood relationships of ground features. As a widely used sub-meter-level remote sensing satellite in my country, Gaofen-2 requires accurate identification of corresponding points in its remote sensing images during data orthorectification and registration—a crucial foundational step in subsequent image processing.

[0004] Image matching is an image processing technique based on image matching. Currently, existing image matching methods can be broadly categorized into two types: region-based and feature-based. Region-based image matching can be further subdivided into grayscale-based and transform-domain-based methods. While relatively simple to implement, this method is computationally intensive and time-consuming due to the sensitivity of images to illumination intensity and nonlinear transformations. This drawback is even more pronounced for information-rich sub-meter level remote sensing images like those from the Gaofen-2 satellite. Feature-based image matching, unlike the aforementioned methods, first extracts features from the target region of the image using feature extraction operators, then uses matching algorithms to match the local features. This approach offers advantages such as low computational cost, robustness, and adaptability. Commonly used feature descriptors include Harris corners, SIFT (Scale-invariant feature transformation), and SURF (Speed-Up Robust Features). The SIFT feature matching algorithm is widely used due to its advantages such as high uniqueness, scale invariance, scalability, and extensibility. However, while its 128-dimensional feature description vector brings rich information, it also increases computational complexity. Furthermore, the original matching points still contain many noise and error points, and direct use of them can cause significant distortion in the matching results. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a method and apparatus for matching corresponding points based on Gaofen-2 satellite remote sensing images.

[0006] The technical solution of this invention is:

[0007] In a first aspect, embodiments of the present invention provide a method for matching corresponding points based on Gaofen-2 satellite remote sensing imagery, the method comprising:

[0008] The Gaofen-2 satellite remote sensing image and its matching reference base map are processed into grayscale to obtain the processed image;

[0009] The processed image is convolved using a Gaussian function to obtain the image scale space corresponding to the processed image, and extreme value detection is performed on the image scale space to obtain the feature point set.

[0010] The feature point set is processed based on the distribution characteristics of the gradient direction of the neighboring pixels of the feature point to obtain the SIFT feature vector.

[0011] Based on the calculated ratio of the nearest neighbor distance to the second nearest neighbor distance between the Euclidean distances of the SIFT feature vectors and a set threshold, matching feature points are obtained;

[0012] Based on the matching feature points, the corresponding points of the Gaofen-2 satellite remote sensing image are determined.

[0013] Optionally, the step of performing a convolution operation on the processed image using a Gaussian function to obtain the image scale space corresponding to the processed image, and performing extremum detection on the image scale space to obtain a feature point set, includes:

[0014] The processed image is convolved with Gaussian kernel functions of different scale factors to form a Gaussian pyramid and generate images of different scales.

[0015] The DOG pyramid function is obtained by subtracting the image functions of adjacent scale spaces in the Gaussian pyramid, and a Gaussian difference pyramid is created based on the DOG pyramid function.

[0016] Based on the difference-of-Gaussian pyramid, determine the extreme feature points corresponding to the processed image;

[0017] The extreme feature points are interpolated and curve-fitted to obtain the feature point set.

[0018] Optionally, the step of processing the feature point set according to the distribution characteristics of the gradient directions of the neighboring pixels of the feature points to obtain the SIFT feature vector includes:

[0019] Based on the distribution characteristics of the gradient direction of the neighboring pixels of the feature point, the direction of the feature point in the feature point set is determined.

[0020] Based on the feature points, determine the direction and vector representation, and determine the target image features in the feature point set;

[0021] Based on the target image features, a SIFT feature vector with rotation invariance and scale invariance is generated.

[0022] Optionally, determining the corresponding points of the Gaofen-2 satellite remote sensing image based on the matching feature points includes:

[0023] Sort the initially matched point set in ascending order of similarity value, so that the initially matched point set becomes an ordered sequence;

[0024] Based on the ordered sequence, determine the solution point set of the homography matrix;

[0025] Based on the homography matrix, the point set is solved to determine the corresponding points of the Gaofen-2 satellite remote sensing image.

[0026] Secondly, embodiments of the present invention provide a matching device for corresponding points based on Gaofen-2 satellite remote sensing imagery, the device comprising:

[0027] The image acquisition module is used to perform grayscale processing on the remote sensing images from the Gaofen-2 satellite and their matching reference base maps to obtain the processed images.

[0028] The feature point set acquisition module is used to perform convolution operation on the processed image using a Gaussian function to obtain the image scale space corresponding to the processed image, and to perform extreme value detection on the image scale space to obtain the feature point set.

[0029] The feature vector acquisition module is used to process the feature point set according to the distribution characteristics of the gradient direction of the neighboring pixels of the feature points to obtain SIFT feature vectors.

[0030] The matching feature point acquisition module is used to acquire matching feature points based on the calculated ratio of the nearest neighbor distance to the second nearest neighbor distance between the Euclidean distances of the SIFT feature vectors and a set threshold.

[0031] The corresponding point determination module is used to determine the corresponding points of the Gaofen-2 satellite remote sensing image based on the matching feature points.

[0032] Optionally, the feature point set acquisition module includes:

[0033] The image generation unit is used to perform convolution operations on the processed image with Gaussian kernel functions under different scale factors to form a Gaussian pyramid and generate images at different scales.

[0034] The Difference-of-Gaussian Pyramid (DOG) creation unit is used to obtain the DOG pyramid function by subtracting the image functions of adjacent scale spaces in the Gaussian pyramid, and to create the Difference-of-Gaussian pyramid based on the DOG pyramid function.

[0035] An extreme feature point determination unit is used to determine the extreme feature points corresponding to the processed image based on the difference-of-Gaussian pyramid.

[0036] The feature point set acquisition unit is used to perform interpolation and curve fitting processing on the extreme feature points to obtain the feature point set.

[0037] Optionally, the feature vector acquisition module includes:

[0038] The feature point determination method acquisition unit is used to acquire the feature point determination direction of the feature point set based on the distribution characteristics of the gradient direction of the neighboring pixels of the feature point;

[0039] The target image feature determination unit is used to determine the target image features in the feature point set based on the direction and vector representation of the feature points.

[0040] The feature vector generation unit is used to generate SIFT feature vectors with rotation invariance and scale invariance based on the target image features.

[0041] Optionally, the corresponding point determination module includes:

[0042] The ordered sequence acquisition unit is used to sort the initially matched point set according to the similarity value from smallest to largest, so that the initially matched point set is formed into an ordered sequence;

[0043] The point set determination unit is used to determine the homography matrix and solve for the point set based on the ordered sequence.

[0044] The corresponding point determination unit is used to determine the corresponding points of the Gaofen-2 satellite remote sensing image by solving the point set based on the homography matrix.

[0045] The advantages of this invention compared to existing technologies are as follows: This invention provides a method for matching corresponding points based on Gaofen-2 satellite remote sensing images. First, image initialization processing is performed, converting the Gaofen-2 satellite remote sensing image to be matched and the reference base map into appropriately sized grayscale images. Then, a Gaussian pyramid and a DoG pyramid are obtained by convolving the image with a Gaussian function. Next, extreme points are initially located within the DoG scale space. Then, noise and redundant points caused by edge effects within the point set are filtered out. After obtaining an accurate feature point set, the orientation of the feature points is determined by the distribution characteristics of the pixel gradient direction in the point's neighborhood, generating an improved stable SIFT feature vector. Initial matching of feature point pairs is performed by calculating the ratio of the nearest neighbor to the Euclidean distance. Finally, the optimal model is obtained by iteratively calculating the feature point pairs using the RANSAC algorithm, completing the purification and error elimination of the matched points. This invention improves upon existing technologies by introducing the concept of a circle in the generation of feature descriptor vectors. Circles possess rotational invariance, a property not found in other shapes. Concentric circles are generated at the center of keypoints, each with a unique gradient direction. This ensures the feature descriptor itself has selection invariance while reducing its dimensionality. Compared to the original 128-dimensional feature descriptor, this method reduces computational complexity and improves matching efficiency. Furthermore, after initial point matching, this invention incorporates the RANSAC algorithm. Through iterative optimization of the optimal model, it effectively removes noise and erroneous points from the matched points, improving the reliability and accuracy of the final image matching and significantly reducing the manual workload of measuring corresponding points. Attached Figure Description

[0046] Figure 1A flowchart illustrating the steps of a method for matching corresponding points based on Gaofen-2 satellite remote sensing images, provided in an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of a Gaussian difference pyramid provided in an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of an improved feature description vector provided in an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram illustrating the use of the least squares method to fit the final optimal model after applying the RANSAC algorithm, as provided in an embodiment of the present invention.

[0050] Figure 5 This is a schematic diagram of a matching device based on remote sensing images from the Gaofen-2 satellite, provided as an embodiment of the present invention. Detailed Implementation

[0051] Example 1

[0052] Reference Figure 1 The diagram illustrates a flowchart of a method for matching corresponding points based on Gaofen-2 satellite remote sensing imagery, as provided in an embodiment of the present invention. Figure 1 As shown, the method may include the following steps:

[0053] Step 101: Perform grayscale processing on the Gaofen-2 satellite remote sensing image and its matching reference base map to obtain the processed image.

[0054] In this embodiment of the invention, image initialization processing can be performed first, that is, the selected Gaofen-2 satellite remote sensing image and the required reference base map are converted into a grayscale image of appropriate size, i.e., the processed image, for subsequent processing work.

[0055] After performing grayscale processing on the Gaofen-2 satellite remote sensing image and its matching reference base map to obtain the processed image, step 102 is executed.

[0056] Step 102: Perform convolution operation on the processed image using a Gaussian function to obtain the image scale space corresponding to the processed image, and perform extreme value detection on the image scale space to obtain the feature point set.

[0057] After performing grayscale processing on the Gaofen-2 satellite remote sensing image and its matching reference base map to obtain the image scale space corresponding to the processed image, and performing extremum detection on the image scale space to obtain the image after feature point set processing, the processed image can be convolved using a Gaussian function to obtain the final image. This implementation process can be described in detail below with reference to the specific implementation method.

[0058] In one specific implementation of this invention, step 102 may include:

[0059] Sub-step A1: Convolve the processed image with Gaussian kernel functions of different scale factors to form a Gaussian pyramid and generate images of different scales.

[0060] In this embodiment of the invention, after obtaining the processed image, the processed image can be convolved with Gaussian kernel functions under different scale factors to form a Gaussian pyramid and generate images of different scales.

[0061] After generating images at different scales, sub-step A2 is performed.

[0062] Sub-step A2: Obtain the DOG pyramid function by subtracting the image functions of adjacent scale spaces in the Gaussian pyramid, and create a Gaussian difference pyramid based on the DOG pyramid function.

[0063] In this example, after constructing the Gaussian pyramid, the DOG pyramid function can be obtained by subtracting the image functions of adjacent scale spaces in the Gaussian pyramid. A Gaussian difference pyramid can then be created based on the DOG pyramid function. The created Gaussian difference pyramid font can be generated as follows: Figure 2 As shown.

[0064] After creating the Gaussian difference pyramid, execute sub-step A3.

[0065] Sub-step A3: Determine the extreme feature points corresponding to the processed image based on the Gaussian difference pyramid.

[0066] After creating the Difference-of-Gaussian pyramid, the extreme feature points corresponding to the processed image can be determined based on the Difference-of-Gaussian pyramid.

[0067] After determining the extreme feature points corresponding to the processed image based on the difference of Gaussian pyramid, sub-step A4 is executed.

[0068] Sub-step A4: Perform interpolation and curve fitting on the extreme feature points to obtain the feature point set.

[0069] After determining the extreme feature points corresponding to the processed image based on the difference in Gaussian pyramid, interpolation and curve fitting can be performed on the extreme feature points to obtain the corresponding feature point set.

[0070] The above implementation process can be described in detail in conjunction with the following process.

[0071] In this example, image scale space extremum detection may include the following process:

[0072] I. Construction of the Gauss Pyramid

[0073] The initialized image I(x,y) is convolved with Gaussian kernel functions G(x,y,σ) at different scale factors to form Gaussian pyramids, generating images at different scales, as shown in the following formula:

[0074] L(x,y,σ)=G(x,y,σ)*I(x,y)

[0075] In the formula, σ is the scale factor, G(x,y,σ) represents the Gaussian function, I(x,y) represents a point on the image, and * represents the convolution of the scale-variable Gaussian function with the image.

[0076] The order of the Gaussian pyramid is determined by the size of the top layer image and the original image, as shown in the following formula:

[0077] n=log2{min(M, N)}-t, t∈[0, log2{min(M, N)}]

[0078] In the formula, M and N are the width and height of the image, and t is the logarithm of the minimum dimension of the tower top image.

[0079] In this example, the width and height of the selected Gaofen-2 remote sensing image and the reference base map are (1568, 1284) and (1732, 1421) respectively. The minimum dimension of the top layer of the pyramid is set to 16. Substituting into the above formula, we can get t = 3 and n = 7.

[0080] To ensure that each level of the subsequent difference pyramid has two image layers capable of detecting extreme points, in this example, the number of image layers in each level of the Gaussian pyramid is set to s = 2 + 3 = 5. The scaling factor of adjacent layers within the same level is k, meaning the scaling factor of the second layer of the first-level pyramid is kσ, and so on for other layers. The first layer of the second-level pyramid is the intermediate layer of the first level, and its scaling factor is k. 2 σ, the first layer of the third-order pyramid is the intermediate layer of the second-order pyramid, and its scale factor is k. 4 σ, and so on for other orders, where σ = 1.52. Based on the calculation, in this example...

[0081] II. Construction of the DOG Pyramid

[0082] The DOG pyramid function is obtained by subtracting the image functions of adjacent scale spaces in the Gaussian pyramid. This function is then used to create the Difference of Gaussian pyramid, as shown in the following formula:

[0083] D(x,y,σ)=(G(x,y,kσ)-G(x,y,σ))*I(x,y)=L(x,y,kσ)-L(x,y,σ)

[0084] The difference function approximates a normalized Gaussian Laplacian operator, so the local extrema extracted by the difference function can be used as relatively stable image features, and are more stable than commonly used features such as gradients and corners.

[0085] III. Initial Feature Point Detection in DOG Space

[0086] After constructing the DOG scale space, it is necessary to find the extreme points (maximum / minimum values) within this scale space. Each pixel must be compared with the corresponding points in the surrounding images at the same scale and the images at the adjacent scales above and below it to detect whether the point is a local extreme point. That is, the extreme point must be an extreme point among the 26 pixels that are simultaneously among the 8 adjacent pixels in the same layer and the 9 adjacent pixels in the two adjacent layers above and below. The extreme points that finally meet the requirements are those potential feature points with scale invariance.

[0087] IV. Precise Feature Point Location

[0088] The extreme points detected through the above steps are essentially extreme points in discrete space. Therefore, it is necessary to interpolate these discrete space points to obtain continuous space value points, and then use a ternary quadratic function to perform curve fitting on the scale-space DOG function to improve the stability of key points and achieve sub-pixel accuracy. The Taylor expansion of the scale-space function D(x,y,σ) at the local extreme point (x0,y0,σ), i.e., the fitting function, is as follows:

[0089]

[0090] Taking the derivative of the above equation and setting it equal to 0, we can calculate the corrected results for the image's rows, columns, and scale.

[0091]

[0092] This result represents the offset of the actual extreme point relative to the center of the local extreme point. Substituting this into the formula yields the equation for the extreme point:

[0093]

[0094] In this example, the number of iterations is set to 5, and repeated interpolation calculations are performed to obtain the result. The T threshold is set to 0.03, and the judgment condition is used. Keypoints that meet this criterion will be considered unstable, low-contrast keypoints and removed from the keypoint set.

[0095] After removing low-contrast keypoints, since the DOG operator has a strong edge response, further edge response points need to be removed to improve the stability of feature points. The DOG extrema of non-feature points have larger principal curvatures at horizontal edges and smaller principal curvatures at vertical edges. The principal curvature of D is proportional to the eigenvalues ​​of the Hessian matrix; therefore, a 2×2 Hessian matrix H is used to calculate the principal curvature of D.

[0096]

[0097] In the formula, D xx D yy To perform two derivatives in the x and y directions of an image at a certain scale in the DOG pyramid, D xy To find the partial derivatives with respect to x and then with respect to y.

[0098] Let α and β be two eigenvalues ​​of the Hessian matrix H, representing the gradients in the x and y directions respectively, and α > β, then we have:

[0099] Tr(H)=D xy +D yy =α+β

[0100] Det(H) = D xx D yy -(D xy ) 2 =αβ

[0101] In the formula, Tr(H) is the trace of the Hessian matrix H, and Der(H) is the determinant of the Hessian matrix H. Let α = rβ, then we have:

[0102]

[0103] Because of (r+1) 2 The ratio / r increases as r increases, and reaches its minimum when α = β. Therefore, setting an appropriate threshold r is sufficient to detect whether the DOG extreme value is an unstable edge response point. The judgment formula is:

[0104]

[0105] In this example, the threshold r is set to 10. Feature points that satisfy the above condition are retained, while feature points that satisfy the above condition are removed, in order to eliminate edge response points and retain stable feature points.

[0106] After obtaining the feature point set, proceed to step 103.

[0107] Step 103: Process the feature point set according to the distribution characteristics of the gradient direction of the neighboring pixels of the feature point to obtain the SIFT feature vector.

[0108] After obtaining the feature point set, the feature point set can be processed according to the distribution characteristics of the gradient direction of the neighboring pixels of the feature points to obtain the SIFT feature vector. The implementation process can be described in detail in conjunction with the specific implementation method below.

[0109] In one specific implementation of the present invention, step 103 may include:

[0110] Sub-step B1: Based on the distribution characteristics of the gradient directions of the neighboring pixels of the feature points, determine the direction of the feature points in the feature point set.

[0111] In this embodiment, after obtaining the feature point set, the feature point determination direction can be obtained based on the distribution characteristics of the gradient direction in the neighborhood of the feature points.

[0112] Sub-step B2: Determine the direction and vector representation based on the feature points, and determine the target image features in the feature point set.

[0113] After determining the orientation of the feature points in the feature point set, the target image features in the feature point set can be determined based on the orientation and vector representation of the feature points.

[0114] Sub-step B3: Generate SIFT feature vectors with rotation invariance and scale invariance based on the target image features.

[0115] After determining the target image features in the feature point set, SIFT feature vectors with rotation invariance and scale invariance can be generated based on the target image features.

[0116] The above implementation process can be described in detail in conjunction with the following flow.

[0117] This example creates a 16×16 rectangular region centered on a keypoint, and divides a circular region with a radius of 8 into four concentric circles with radii differing by 2. The gradient direction and gradient value for each concentric circle are calculated using the following formula:

[0118]

[0119] θ(x, y) = tan -1 ((L(x+1,y)-L(x-1,y)) / (L(x,y+1)-L(x,y-1)))

[0120] In the formula, m(x, y) is the gradient value, θ(x, y) is the gradient direction, L is the scale of each feature point, and (x, y) is the pixel position corresponding to the specified layer.

[0121] In four concentric circles, eight cumulative gradient directions are evenly distributed from 0° to 360° using histograms. Within each concentric circle, the eight gradient directions are transformed into directions ranging from 0° to 360°, accumulated at 45-degree intervals. The gradient values ​​for each direction are then processed using a Gaussian weighting function. This generates eight directional feature vectors. The gradient values ​​in the concentric circle feature vectors are arranged from largest to smallest. Then, starting from the keypoint, the feature vectors of each concentric circle are sequentially connected and numbered from the innermost circle to the outermost circle. Finally, standard normalization is used to reduce noise interference. The formula is as follows:

[0122]

[0123] In the formula, Let j be the vector of the i-th concentric circle. This is a normalized feature description vector.

[0124] By subtracting the absolute values ​​of gradient directions that differ by 180°, the dimensionality of each concentric circle feature vector is reduced from 8 to 4, as shown in the following formula:

[0125]

[0126] The improved feature descriptor in this example consists of four concentric circles, each containing a 4-dimensional feature vector, for a total of 16-dimensional vector representation. This feature descriptor vector improves rotation invariance while reducing dimensionality, thus lowering computational complexity while maintaining the accuracy of feature points. Finally, normalization is used to generate the feature descriptor vector:

[0127]

[0128] The improved feature descriptor vector can be as follows: Figure 3 As shown.

[0129] After processing the feature point set to obtain the SIFT feature vector based on the distribution characteristics of the gradient direction of the neighboring pixels of the feature point, step 104 is executed.

[0130] Step 104: Based on the calculated ratio of the nearest neighbor distance to the second nearest neighbor distance between the Euclidean distances of the SIFT feature vectors and a set threshold, obtain the matching feature points.

[0131] After processing the feature point set to obtain SIFT feature vectors based on the distribution characteristics of the gradient directions of the neighboring pixels of the feature points, matching feature points can be obtained by using the ratio of the Euclidean distance between the calculated SIFT feature vectors to the nearest neighbor distance and the second nearest neighbor distance, along with a set threshold. This implementation process can be described in detail below.

[0132] SIFT feature vector matching mainly includes:

[0133] Let M represent the standard image and N represent the image to be matched. In this example, the feature vector sets of the two images are extracted as A. m and B n :

[0134]

[0135]

[0136] Using Euclidean distance, we search for the minimum distance d(1) and the second minimum distance d(2) between elements of two feature vector sets. The Euclidean distance formula is as follows:

[0137]

[0138] In the formula, m and n represent the dimensions of each set of feature vectors.

[0139] The ratio d(1) / d(2) of the nearest Euclidean distance to the second nearest Euclidean distance is compared with a set threshold a. In this example, the threshold a is 0.6. Matching point pairs with a ratio less than the threshold a are judged as successful matching points, while matching point pairs with a ratio less than the threshold a are judged as unsuccessful matching points.

[0140] After obtaining the matching feature points based on the ratio of the nearest neighbor distance to the second nearest neighbor distance of the Euclidean distance between the calculated SIFT feature vectors and the set threshold, step 105 is executed.

[0141] Step 105: Based on the matching feature points, determine the corresponding points of the Gaofen-2 satellite remote sensing image.

[0142] After obtaining matching feature points based on the ratio of the nearest neighbor distance to the second nearest neighbor distance in the Euclidean distance between SIFT feature vectors and a set threshold, corresponding points in the Gaofen-2 satellite remote sensing image can be determined based on the matching feature points. This implementation process can be described in detail below with reference to the specific implementation method.

[0143] In another specific implementation of the present invention, step 105 may include:

[0144] Sub-step C1: Sort the initially matched point set in ascending order of similarity value, so that the initially matched point set becomes an ordered sequence.

[0145] In this embodiment of the invention, after obtaining the matching feature points, the initial matching point set can be sorted in ascending order of similarity value, so that the initial matching point set becomes an ordered sequence.

[0146] After obtaining the ordered sequence, execute sub-step C2.

[0147] Sub-step C2: Determine the homography matrix solution point set based on the ordered sequence.

[0148] After obtaining the ordered sequence, the solution point set of the homography matrix can be determined based on the ordered sequence.

[0149] After determining the homography matrix solution point set based on the ordered sequence, sub-step C3 is executed.

[0150] Sub-step C3: Based on the homography matrix, solve for the point set and determine the corresponding points of the Gaofen-2 satellite remote sensing image.

[0151] After determining the homography matrix solution point set based on the ordered sequence, the corresponding points of the Gaofen-2 satellite remote sensing image can be determined based on the homography matrix solution point set.

[0152] The implementation process can be described in detail in conjunction with the following flow.

[0153] In this example, the RANSAC algorithm can be used for purification and error removal:

[0154] The RANSAC algorithm is used for purification and error removal, mainly including:

[0155] First, the points set after initial matching are sorted in ascending order of similarity value (ratio of nearest neighbor to second nearest neighbor), making the initial matching points set an ordered sequence. Then, the top 25% of this ordered matching point set is taken as the solution set for the homography matrix. In this top 25% matching point set, four pairs of feature points are selected sequentially to calculate the current homography matrix. The transformation formula is as follows:

[0156]

[0157] In the formula, x,y and x',y' are the coordinates of the matching points on the reference base map and the image to be matched, respectively, and h 11 ~h 33 H represents the corresponding homography matrix, and s represents the scale parameter.

[0158] In this example, h is used. 33 =1 to normalize the homography matrix H. The remaining 8 parameters in H can be solved using 8 linear equations listed for 4 pairs of feature points. After obtaining the homography matrix H, the transformation error of the remaining matching points is calculated based on this parameter model. That is, the distance between the transformed coordinates of the matching point on the reference base image and its corresponding matching point, and the distance between the transformed coordinates of the matching point on the image to be matched and its corresponding matching point. The sum of the two distances is called the transformation error, and the formula is as follows:

[0159]

[0160] In the formula, d i For transformation error, These are the corresponding matching points on the two images.

[0161] Points with a transformation error less than the threshold b (b is 3 times the minimum distance of the matching point set) are counted as inliers. In this example, when the proportion of inliers is greater than 70%, the parameter model is considered to be a better model. The parameter and the inlier set are saved.

[0162] In this example, the number of iterations is set to 500. The calculation continues until the next step. When a parametric model with a better interior point ratio exists, it replaces the optimal model from the previous iteration, and this model is recorded as the optimal model. For fitting the final optimal model using the least bisector method after applying the RANSAC algorithm, the following can be used: Figure 4 As shown, the parameters and interior point set are saved until the iteration ends.

[0163] The interior point set calculated using the optimal parameter model after the above iterations is fitted using the least squares method to obtain the final optimal model parameters and matching point set, thus completing the purification and error removal of matching points in this example.

[0164] This invention improves the efficiency and accuracy of matching corresponding points of Gaofen-2 to a certain extent. Using the invented matching method, the accuracy error between the final Gaofen-2 image to be matched and the reference base map in this example is generally controlled within one pixel, providing a reliable foundation for subsequent remote sensing image processing.

[0165] Example 2

[0166] Reference Figure 5 The diagram illustrates a structural schematic of a corresponding point matching device based on Gaofen-2 satellite remote sensing imagery provided by an embodiment of the present invention. Figure 5 As shown, the device may include the following modules:

[0167] The image processing and acquisition module 510 is used to perform grayscale processing on the remote sensing images of the Gaofen-2 satellite and the reference base map that matches them to obtain the processed image;

[0168] The feature point set acquisition module 520 is used to perform convolution operation on the processed image using a Gaussian function to obtain the image scale space corresponding to the processed image, and to perform extreme value detection on the image scale space to obtain the feature point set.

[0169] The feature vector acquisition module 530 is used to process the feature point set according to the distribution characteristics of the gradient direction of the neighboring pixels of the feature points to obtain the SIFT feature vector.

[0170] The matching feature point acquisition module 540 is used to acquire matching feature points based on the calculated ratio of the nearest neighbor distance to the second nearest neighbor distance between the Euclidean distances of the SIFT feature vectors and a set threshold.

[0171] The corresponding point determination module 550 is used to determine the corresponding points of the Gaofen-2 satellite remote sensing image based on the matching feature points.

[0172] Optionally, the feature point set acquisition module includes:

[0173] The image generation unit is used to perform convolution operations on the processed image with Gaussian kernel functions under different scale factors to form a Gaussian pyramid and generate images at different scales.

[0174] The Difference-of-Gaussian Pyramid (DOG) creation unit is used to obtain the DOG pyramid function by subtracting the image functions of adjacent scale spaces in the Gaussian pyramid, and to create the Difference-of-Gaussian pyramid based on the DOG pyramid function.

[0175] An extreme feature point determination unit is used to determine the extreme feature points corresponding to the processed image based on the difference-of-Gaussian pyramid.

[0176] The feature point set acquisition unit is used to perform interpolation and curve fitting processing on the extreme feature points to obtain the feature point set.

[0177] Optionally, the feature vector acquisition module includes:

[0178] The feature point determination method acquisition unit is used to acquire the feature point determination direction of the feature point set based on the distribution characteristics of the gradient direction of the neighboring pixels of the feature point;

[0179] The target image feature determination unit is used to determine the target image features in the feature point set based on the direction and vector representation of the feature points.

[0180] The feature vector generation unit is used to generate SIFT feature vectors with rotation invariance and scale invariance based on the target image features.

[0181] Optionally, the corresponding point determination module includes:

[0182] The ordered sequence acquisition unit is used to sort the initially matched point set according to the similarity value from smallest to largest, so that the initially matched point set is formed into an ordered sequence;

[0183] The point set determination unit is used to determine the homography matrix and solve for the point set based on the ordered sequence.

[0184] The corresponding point determination unit is used to determine the corresponding points of the Gaofen-2 satellite remote sensing image by solving the point set based on the homography matrix.

[0185] The specific embodiments described in this application are intended to enable those skilled in the art to gain a more comprehensive understanding of this application, but do not limit this application in any way. Therefore, those skilled in the art should understand that modifications or equivalent substitutions can still be made to this application; and all technical solutions and improvements that do not depart from the spirit and technical essence of this application should be covered within the scope of protection of this patent application.

[0186] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A same-name point matching method based on Gao Fen-2 satellite remote sensing images, characterized in that, The method includes: The Gaofen-2 satellite remote sensing image and its matching reference base map are processed into grayscale to obtain the processed image; The processed image is convolved using a Gaussian function to obtain the image scale space corresponding to the processed image, and extreme value detection is performed on the image scale space to obtain the feature point set. The feature point set is processed based on the distribution characteristics of the gradient direction of the neighboring pixels of the feature point to obtain the SIFT feature vector. Based on the calculated ratio of the nearest neighbor distance to the second nearest neighbor distance between the Euclidean distances of the SIFT feature vectors and a set threshold, matching feature points are obtained; Based on the matching feature points, determine the corresponding points of the Gaofen-2 satellite remote sensing image; The process involves performing a convolution operation on the processed image using a Gaussian function to obtain the image scale space corresponding to the processed image, and then performing extremum detection on the image scale space to obtain a feature point set, including: The processed image is convolved with Gaussian kernel functions of different scale factors to form a Gaussian pyramid and generate images of different scales. The DOG pyramid function is obtained by subtracting the image functions of adjacent scale spaces in the Gaussian pyramid, and a Gaussian difference pyramid is created based on the DOG pyramid function. Based on the difference-of-Gaussian pyramid, determine the extreme feature points corresponding to the processed image; The extreme feature points are interpolated and curve-fitted to obtain the feature point set. The step of processing the feature point set according to the distribution characteristics of the gradient direction of the neighboring pixels of the feature points to obtain the SIFT feature vector includes: Based on the distribution characteristics of the gradient direction of the neighboring pixels of the feature point, the direction of the feature point in the feature point set is determined. Based on the feature points, determine the direction and vector representation, and determine the target image features in the feature point set; Based on the target image features, a SIFT feature vector with rotation invariance and scale invariance is generated, specifically as follows: A circular neighborhood is created centered on the feature point, and the circular neighborhood is divided into multiple concentric circular sub-regions; the gradient direction distribution of each concentric circular sub-region is statistically analyzed to generate a SIFT feature vector with rotation invariance and reduced dimension.

2. The method of claim 1, wherein, The step of determining the corresponding points of the Gaofen-2 satellite remote sensing image based on the matching feature points includes: Sort the initially matched point set in ascending order of similarity value, so that the initially matched point set becomes an ordered sequence; Based on the ordered sequence, determine the solution point set of the homography matrix; Based on the homography matrix, the point set is solved to determine the corresponding points of the Gaofen-2 satellite remote sensing image.

3. A same-name point matching device based on Gao Fen II satellite remote sensing images, characterized in that, The apparatus comprising, using the method of claim 1, includes: The image acquisition module is used to perform grayscale processing on the remote sensing images from the Gaofen-2 satellite and their matching reference base maps to obtain the processed images. The feature point set acquisition module is used to perform convolution operation on the processed image using a Gaussian function to obtain the image scale space corresponding to the processed image, and to perform extreme value detection on the image scale space to obtain the feature point set. The feature vector acquisition module is used to process the feature point set according to the distribution characteristics of the gradient direction of the neighboring pixels of the feature points to obtain SIFT feature vectors. The matching feature point acquisition module is used to acquire matching feature points based on the calculated ratio of the nearest neighbor distance to the second nearest neighbor distance between the Euclidean distances of the SIFT feature vectors and a set threshold. The corresponding point determination module is used to determine the corresponding points of the Gaofen-2 satellite remote sensing image based on the matching feature points.

4. The apparatus according to claim 3, characterized in that, The feature point set acquisition module includes: The image generation unit is used to perform convolution operations on the processed image with Gaussian kernel functions under different scale factors to form a Gaussian pyramid and generate images at different scales. The Difference-of-Gaussian Pyramid (DOG) creation unit is used to obtain the DOG pyramid function by subtracting the image functions of adjacent scale spaces in the Gaussian pyramid, and to create the Difference-of-Gaussian pyramid based on the DOG pyramid function. An extreme feature point determination unit is used to determine the extreme feature points corresponding to the processed image based on the difference-of-Gaussian pyramid. The feature point set acquisition unit is used to perform interpolation and curve fitting processing on the extreme feature points to obtain the feature point set.

5. The apparatus according to claim 3, characterized in that, The feature vector acquisition module includes: The feature point determination method acquisition unit is used to acquire the feature point determination direction of the feature point set based on the distribution characteristics of the gradient direction of the neighboring pixels of the feature point; The target image feature determination unit is used to determine the target image features in the feature point set based on the direction and vector representation of the feature points. The feature vector generation unit is used to generate SIFT feature vectors with rotation invariance and scale invariance based on the target image features.

6. The apparatus according to claim 3, characterized in that, The corresponding point determination module includes: The ordered sequence acquisition unit is used to sort the initially matched point set according to the similarity value from smallest to largest, so that the initially matched point set is formed into an ordered sequence; The point set determination unit is used to determine the homography matrix and solve for the point set based on the ordered sequence. The corresponding point determination unit is used to determine the corresponding points of the Gaofen-2 satellite remote sensing image by solving the point set based on the homography matrix.

Images