An edge enhancement and distribution equalization remote sensing stereoscopic image feature matching method

Abstract

The application discloses a kind of edge enhancement and distribution equalization's remote sensing stereoscopic image feature matching method, corresponding multi-scale space image of reference image and image to be registered is constructed, and each scale space image is filtered with unbiased ratio UDR edge detector, to obtain the final multi-scale space image;Search the extreme point in the final multi-scale space image, the extreme point obtained is candidate feature point, then using non-maximum suppression algorithm to the candidate feature point is handled uniformly, feature point extraction is completed;With feature point as center, four different concentric circle regions of different sizes are selected from different scale images to construct multi-scale feature descriptor, calculate multi-scale feature descriptor vector, then using nearest neighbor distance ratio algorithm to obtain initial feature matching point pair, then using random sample consensus algorithm to delete the abnormal point in initial feature matching point pair, complete reference image and image to be registered feature point registration.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and particularly relates to a remote sensing stereo image feature matching method with edge enhancement and distribution equalization. BACKGROUND

[0002] With the innovation of sensor imaging technology, the breakthrough of space communication technology and the rapid development of satellite attitude determination and orbit determination technology, contemporary remote sensing technology has entered a brand-new development stage. In China, with the launch of a series of high-resolution satellites such as Mapping Satellite-1, Resource Satellite-3 and High Resolution Satellite, the number and quality of satellite images in China have increased exponentially. Precise geometric positioning of high-resolution space-borne SAR images has become a research hotspot. Due to the limitations of sensor errors, orbit determination errors and atmospheric propagation delay errors, the direct positioning accuracy of space-borne SAR images cannot meet the actual application requirements. Therefore, the present application develops a high-precision feature matching method for high-resolution remote sensing stereo images with edge enhancement and distribution equalization.

[0003] The commonly used SAR image feature matching method is SAR-SIFT algorithm. This algorithm uses ROEWA detector to construct SAR-Harris scale space, has the characteristic of constant false alarm rate and can well reduce the adverse effects of speckle noise in SAR image on feature point extraction. However, the scale space constructed by ROEWA detector can only maintain the characteristic of constant false alarm rate, and cannot achieve good results in retaining edge information of the image. In actual use, edge information is easily ignored, resulting in incorrect feature points, and the extracted feature points are prone to local aggregation. Therefore, in the subsequent feature matching, it is easy to cause loss of accuracy and incorrect feature point matching pairs. SUMMARY

[0004] The present application provides a remote sensing stereo image feature matching method with edge enhancement and distribution equalization, which solves the technical problems of inaccurate edge information positioning and easy local aggregation in the existing SAR-SIFT algorithm.

[0005] The present application can be implemented by the following technical solutions:

[0006] A remote sensing stereo image feature matching method with edge enhancement and distribution equalization, comprising the following steps:

[0007] I. Constructing scale space image

[0008] Constructing multi-scale space images corresponding to the reference image and the image to be registered, and performing edge filtering on each scale space image using the unbiased ratio UDR edge detector to obtain the final multi-scale space image;

[0009] II. Extracting feature points

[0010] The extreme value points in the final multi-scale space image are searched, and the obtained extreme value points are candidate feature points.

[0011] III. Feature point registration

[0012] A multi-scale feature descriptor is constructed by selecting four concentric circles with different sizes from different scale images with the feature points as the center, a multi-scale feature descriptor vector is calculated, an initial feature matching point pair is obtained by using a nearest neighbor distance ratio algorithm, and an abnormal point in the initial feature matching point pair is deleted by using a random sample consensus algorithm, so that feature point registration of the reference image and the image to be registered is completed.

[0013] Further, the unbiased ratio UDR edge detector uses a combination of difference and ratio to perform edge detection. This method has the characteristics of unbiased edge positioning of difference operation and constant false alarm rate under the influence of speckle noise of ratio operation, which ensures the edge detection accuracy as much as possible under the reduction of noise image, and reserves good image information for subsequent feature point extraction. The calculation formula of the unbiased ratio UDR edge detector is as follows

[0014] h UDR (x)=h ⊥ UDR (|x|)h || (y)

[0015] Wherein, the calculation formula of the one-dimensional vertical filter h ⊥ UDR (|x|) is as follows

[0016]

[0017]

[0018]

[0019] The calculation formula of the one-dimensional horizontal filter h || UDR (y) is as follows

[0020]

[0021] Wherein, h ⊥ UDR (x) and h || UDR(y) respectively represent a one-dimensional vertical filter and a one-dimensional horizontal filter, x and y respectively represent the x coordinate and the y coordinate of the pixel point, the parameter alpha and the parameter beta respectively represent the peak value and the decay rate of the one-dimensional vertical filter, the parameter epsilon (.) represents a unit step function, the parameter rho represents an initial layer scale space (default setting is 2), the parameter eta represents a scale ratio of adjacent two layers (default setting is ), i represents the number of scale space layers, the parameter l || and the parameter sigma || respectively represent the flatness and the decay degree of the bell-shaped function.

[0022] Further, when constructing the multi-scale feature descriptor, the main direction of the feature point is calculated first, then the scale image where the feature point is located, and the upper two layers of scale images adjacent to the scale image, the lower layer of scale image or the upper layer of scale image, the lower two layers of scale image are selected, the coordinate system where the feature point is located is rotated into the main direction of the feature point according to the main direction of the feature point after the scale space is determined, then the concentric circular regions of different sizes are selected with the feature point and the corresponding points of the feature point on other scale images as the center, then the concentric circular regions are divided into a plurality of sub-regions, the gradient histogram of eight directions in each sub-region is counted, and the multi-scale feature descriptor vector is composed.

[0023] Further, the radius of the circular region on the scale image where the feature point is located is the minimum, which is recorded as the minimum circle, and the corresponding radius is recorded as R4, and the radii of other circular regions are R1=0.25R4, R2=0.5R4 and R3=0.73R4 respectively, the minimum circular region of the concentric circular region is set as a separate sub-region, and the annular region corresponding to the remaining concentric circular region is divided into a sub-region every 45 degrees, and a total of 25 sub-regions are obtained.

[0024] Further, when searching for extreme points, the selected neighborhood is set as a 5*5 or 7*7 pixel region.

[0025] The present application has the beneficial technical effects that:

[0026] (1) In view of the problem that the ROEWA cannot accurately obtain the unbiased edge positioning of the SAR image in the existing SAR-SiFT algorithm, the UDR edge detection operator is used to replace the ROEWA operator to construct the multi-scale space of the image gradient, and the pyramid image with constant false alarm rate and unbiased edge positioning is obtained, which has more accurate edge information compared with the ROEWA.

[0027] (2) In view of the problem that the feature points are gathered, redundant and noisy in the spatial distribution in the existing SAR-SiFT algorithm, the non-maximum suppression algorithm is used to remove the redundant feature points on the basis of the feature points extracted in the present application, and the uniformized feature points are obtained for descriptor construction and feature point matching.

[0028] (3) Although the existing I-SAR-SIFT acquires candidate point information of different scales by using concentric circles, the information of the multi-scale feature descriptor is still used for matching in subsequent feature matching, and similar descriptor information still exists, therefore, the number of concentric circles is increased to retain more scale space information on this basis, so as to further improve the uniqueness of the multi-scale feature descriptor. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 It is a general flowchart of the present application;

[0030] Figure 2 It is a space field expansion diagram when screening candidate feature points of the present application;

[0031] Figure 3 It is a process diagram of constructing a multi-scale feature descriptor of the present application;

[0032] Figure 4 It is a visualization result diagram of the present application. DETAILED DESCRIPTION

[0033] The specific embodiments of the present application will be described in detail below in combination with the drawings and preferred embodiments.

[0034] As shown in the drawings, Figure 1 The present application provides a high-precision feature matching method for high-resolution satellite remote sensing stereo images with edge enhancement and distribution equalization. The algorithm is improved on the basis of the existing SAR-SIFT algorithm. The edge unbiased positioning and constant false alarm rate characteristics of the UDR edge extraction are used to reduce the extracted false corner point information and improve the feature point extraction accuracy. In addition, the non-maximum suppression algorithm is used to uniformize the feature points on the basis of the extracted feature points by the algorithm, so as to reduce the local aggregation phenomenon of the feature points. Finally, the descriptor construction strategy that retains multi-scale information is used to improve the uniqueness of the multi-scale feature descriptor when constructing the multi-scale feature descriptor. The SAR image data test shows that the RMSE of the algorithm proposed in the present application is reduced by about one pixel compared with the existing SAR-SIFT algorithm, and there is no local aggregation phenomenon in the matching point distribution in the test data. The specific embodiments are as follows:

[0035] I. Constructing scale space images

[0036] The multi-scale space images corresponding to the reference image and the image to be registered are constructed, and the unbiased ratio UDR edge detector is used to complete the edge filtering of each scale space image, so as to obtain the final multi-scale space image.

[0037] The ROEWA edge detector adopted in the prior SAR-SIFT algorithm only considers ensuring the constant false alarm rate of the image, but does not consider the accurate positioning problem of the edge information of the image, while the UDR filter adopts the combination of the ratio operation and the difference operation, and simultaneously has the constant false alarm rate feature in the ratio operation and the edge accurate positioning feature in the difference operation, so that the present application adopts the unbiased ratio UDR edge detector to perform edge filtering on the multi-scale space image, can effectively reduce the false corner points existing in the prior SAR-SIFT algorithm, and extracts high-precision feature points.

[0038] The Gaussian kernel is the only linear kernel for establishing the scale space, the present application adopts the Gaussian function kernel satellite remote sensing stereo image convolution with increasing variance to construct the scale space of the graph to simulate the multi-scale features of the image data. According to the prior SAR-SIFT algorithm, the Gaussian difference pyramid is constructed, then the unbiased ratio UDR edge detector, namely the two-dimensional UDR edge detector, is adopted to perform edge detection on each scale image, the edge filtering images corresponding to different scale spaces are obtained, and the final multi-scale space image is obtained.

[0039] The calculation formula of the two-dimensional UDR edge detector is as follows:

[0040] h UDR (x)=h ⊥ UDR (|x|)h || (y) (1)

[0041] Wherein, the calculation formula of the one-dimensional vertical filter h ⊥ UDR (|x|) is as follows

[0042]

[0043] The calculation formula of the one-dimensional parallel filter h || UDR (y) is as follows

[0044]

[0045] Wherein, the calculation formula of the one-dimensional vertical filter h ⊥ UDR (x) and the one-dimensional parallel filter h || UDR (y) are respectively represented, x and y respectively represent the x coordinate and the y coordinate of the pixel point, the parameter α and the parameter β respectively represent the peak value and the decay rate of the one-dimensional vertical filter, the parameter ε(.) represents the unit step function, the parameter ρ represents the initial layer scale space (default setting is 2), the parameter η represents the scale ratio of adjacent two layers (default setting is ), i represents the scale space layer number, and the parameter l|| and parameter sigma || respectively represent the flatness and the attenuation degree of the clock function.

[0046] After the final multi-scale space image is constructed, the gradient ratio method (GR) is used to calculate the gradient size and gradient direction of each scale space image, so as to be used for subsequent feature point pair description calculation. The method is robust to speckle noise and is specially used for SAR images. In the GR algorithm, the area weighting ratios of the left and right areas in the horizontal direction and the upper and lower areas in the vertical direction of the ROEWA operator are taken logarithm, which are used to calculate the gradient size and gradient direction of the pixel point, and the horizontal and vertical gradient calculation methods are redefined. According to the principle of calculating the gradient according to the GR algorithm, the horizontal rotation angle and the vertical rotation angle of the two-dimensional UDR edge detector are redefined, and the horizontal gradient and the vertical gradient are calculated according to the GR algorithm formula, and the calculation formula is as follows

[0047]

[0048] Wherein, h 0° UDR and h 90° UDR respectively represent the horizontal and vertical rotation angles of the UDR edge filter of formula (1), Gx, alpha and Gy, alpha respectively represent the horizontal and vertical gradients, and the parameter alpha is an exponential weight parameter for calculating the local mean. The gradient size and direction of each pixel point are calculated according to the following formula

[0049]

[0050] Wherein, Gm, alpha and Go, alpha are the gradient size and direction respectively, and Gx, alpha and Gy, alpha are the horizontal and vertical gradients respectively.

[0051] Second, extracting feature points

[0052] The extreme value points obtained by searching the extreme value of each scale space response function R(x, y, sigma n ) are candidate feature points, and the non-maximum suppression algorithm is used to uniformly process the candidate feature points, so as to complete the feature point extraction.

[0053] The Harris-Laplace algorithm is used to extract candidate feature points by searching the extreme value of each scale space response function R(x, y, sigma n ), and the matrix corresponding to the Harris-Laplace algorithm and the response function formula are as follows:

[0054]

[0055] Wherein, M(x, y, sigma n) represents the Harris-Laplace matrix based on the two-dimensional UDR edge detector, R(x, y, σ n ) is the response function of M(x, y, σ n ), σ n represents the nth scale space, represents the two-dimensional Gaussian function with a standard deviation of , Gx, a and Gy, a are the values of the horizontal and vertical gradients calculated in formula (4), and d is an arbitrary parameter, which is 0.04 by default.

[0056] In the existing SAR-SIFT algorithm, candidate feature points are extracted by searching the extreme value of the response function R(x, y, σ n ) of each scale space. The algorithm only compares the response values of the 3*3 spatial neighborhood of the current layer to detect whether the current point is the extreme value point of the 3*3 spatial neighborhood. When extracting feature points, it is noticed that a large number of feature points are gathered in the area with a larger gray value, and the interval between the feature points is within two pixels. These feature points with close spatial positions may cause false matching in the subsequent feature point matching algorithm, thereby reducing the matching accuracy. Therefore, the extreme value point detection of the present application expands the comparison range of candidate feature points to solve the above problems. As shown in formula (5), a 5*5 or 7*7 spatial neighborhood is selected in the same scale space for comparison. If it is an extreme value point, it is retained, and if it is not, the candidate feature point is removed. By expanding the field comparison range, the extraction of feature points with close spatial positions can be limited, the gathering of feature points in the local area can be reduced, the distinguishability of the corner points can be increased, and better feature point information can be provided for the subsequent feature point matching. Figure 2

[0057] Although the spatial distribution of the feature points obtained after the above operation is more uniform compared with the existing SAR-SIFT, a large number of feature points obtained through the final matching are still gathered and redundant. Therefore, the present application selects an adaptive non-maximum suppression method to retain the feature points with the most relevance. The non-maximum suppression algorithm is used to suppress the feature points within a certain radius that are below the empirically determined threshold, and is a direct method for suppressing unnecessary corner point information. However, this method can only avoid the gathering of feature points within the window, but cannot guarantee the uniform distribution of the global feature points. The present application adds the non-maximum suppression algorithm to the feature detection, uniformly processes the initial feature points obtained by expanding the field comparison range in the previous step, and only constructs the multi-scale feature descriptor for the uniformly processed feature points in the subsequent processing.

[0058] III. Feature point registration

[0059] ​The multi-scale feature descriptor is constructed by taking the feature point as the center and selecting four concentric circular regions with different sizes from different scale images, a multi-scale feature descriptor vector is calculated, and then the nearest neighbor distance ratio algorithm is used to obtain initial feature matching point pairs, and the random sample consensus algorithm is used to delete abnormal points in the initial feature matching point pairs to complete the feature point registration of the reference image and the image to be registered.

[0060] The multi-scale feature descriptor construction flow chart is shown in Figure 3 Firstly, the main direction of the feature point is calculated, as shown in Figure 3 (a) and (b), a circular neighborhood with a radius of 6σ (σ is the scale space where the feature point is located) is divided with the feature point as the center, and each pixel point in the circular neighborhood has its own gradient size and direction, as shown in Figure 3 (c), then the gradient size and direction of each pixel point in the neighborhood are counted using a histogram, the gradient histogram divides the 0-360 degree direction range into 36 columns, each column is 10 degrees, and the peak value direction of the histogram represents the main direction of the feature point;

[0061] Secondly, the multi-scale feature descriptor vector is calculated, when constructing the multi-scale feature descriptor, the scale image where the feature point is located, and the two upper scale images, the lower scale image or the upper scale image, the lower two scale images adjacent to the size image are selected to determine the scale space, then the coordinate system where the feature point is located is rotated according to the main direction of the feature point, then the feature point and the corresponding point of the feature point on other scale images are taken as the center, different size concentric circular regions are selected, then the concentric circular regions are divided into multiple sub-regions, the gradient histogram of eight directions in each sub-region is counted, and the multi-scale feature descriptor vector is composed, thereby the multi-scale descriptor is constructed.

[0062] As shown in Figure 3 (d1), in order to improve the uniqueness of the multi-scale feature descriptor, the present application takes the scale space where the feature point is currently located as the basis, respectively selects three different scale space information as the auxiliary, constructs four different scale regions of the multi-scale feature descriptor, wherein the D1 region retains the image gradient size and direction in the scale space where the feature point is currently located, the range is a circular region with the feature point as the center, and the corresponding radius is R1, the D2-D3 region retains the image gradient size and direction information of different scales, the range is different circular regions with the corresponding points of the feature point in other scale space images as the center, and the corresponding radius is R2-R4, the size of the circular region is as shown in Figure 3(d2) as shown, R4 = 12σ, R3 = 0.73R4, R2 = 0.5R4, R1 = 0.25R4, which together constitute concentric circular regions, after determining the scale space, the coordinate system in which the feature point is located is rotated to the main direction of the feature point, after retaining the information of different scale spaces, in order to more detailedly describe the feature information in the neighborhood of the feature point, it is divided into 25 sub-regions, wherein the innermost circle, i.e. the smallest circle R1, is not angle-divided and is a separate sub-region, and the remaining three circular ring regions are divided into a sub-region every 45 degrees. Figure 3 (c) is similar, the gradient histogram of the scale space in each sub-region in eight directions (45°, 90°, …, 360°) is counted, and the counted 200-dimensional gradient information is the multi-scale feature descriptor vector of the feature point.

[0063] After obtaining the multi-scale feature descriptor vector, a two-step method is used to obtain correct matching point pairs, namely a nearest neighbor distance ratio, i.e. NNDR, which is used to find an initial matching point pair, and a random sample consensus algorithm, i.e. RANSAC, which is used to remove outliers in the initial matching point pair, wherein the NNDR algorithm is based on the Euclidean distance representation, and the nearest distance obtained after sorting is compared with the next nearest distance. If a feature point satisfies:

[0064] d n <d sn ×d ratio

[0065] then it is considered to be a preliminary matching pair; wherein d n represents the nearest neighbor distance, d sn represents the next nearest neighbor distance, and d ratio represents a matching threshold, in the present application, d ratio = 0.8.

[0066] The matching points obtained by the NNDR algorithm are initial matching pairs, and usually still have many false matching pairs, therefore, the RANSAC algorithm is used for subsequent outlier removal, usually an RANSAC sample is extracted from the initial matching pairs, a transformation matrix M is calculated according to the RANSAC sample, a consistent set satisfying the current transformation matrix is calculated according to the sample set and the transformation matrix M and an error measurement function, and the number of point pairs in the consistent set is recorded. Whether the current consistent set is the optimal (maximum) consistent set is judged according to the number of elements in the current consistent set, if yes, the current optimal consistent set is updated, the current error probability p is updated, if p is still greater than the allowed minimum error probability, the above operation is repeated for iteration, until the current error probability p is less than the minimum error probability, which is set to 0.01 in the present application, and the matching point pairs in the final optimal consistent set are the final correct matching point pairs, which are used for the registration of the reference image and the image to be registered.

[0067] To verify the feasibility of the feature matching method of the present application, we carried out the following experiment:

[0068] For the existing SAR-SIFT algorithm, the number of image feature points is limited to 4000, the first layer scale is set to p1=2, and the constant of the adjacent two scale spaces is The number of scale spaces is set to 5, and the arbitrary parameter d is set to 0.04. The parameter settings of the method proposed in the present application are the same as the existing SAR-SIFT parameter settings, wherein the parameter settings of the unbiased ratio UDR edge detector are adjusted according to the image, and the initial values are set to a=2, b=1, l=1.5, and s=1.5. After the feature points are homogenized by the method of the present application, the number of retained feature points is 1 / 4 of the initial number of feature points, the threshold value d of the NNDR algorithm in feature matching is ratio =0.8, and the specific test data are shown in Table 1.

[0069] Table 1 Detailed information of test data

[0070]

[0071] The quantitative indicators of the present application are as follows:

[0072] Precision evaluation index: the root mean square error RMSE is used as the evaluation of the deviation between the predicted value and the positioning value. Ten pairs of uniformly distributed matching points are manually selected to iteratively estimate the affine model H of the image in each test set. The positioning value is obtained according to the affine transformation model, and therefore, the RMSE is defined as:

[0073]

[0074] (x i ′,y i ′,1)=H·(x i ,y i ,1)

[0075] wherein N is the total number of matching feature points, (x i ,y i ) is the predicted value representing the inlier coordinates of the secondary image after RANSAC outlier screening, and (x i ′,y i ′) is the positioning value representing the positioning coordinates of the secondary image obtained according to the affine model H obtained after the uniform matching points are manually selected.

[0076] Correct matching rate CMR: the correct matching rate is used to evaluate the robustness of the matching algorithm. The CMR is defined as wherein N corrN represents the number of matching pairs calculated by using the affine model H. nndr N represents the number of matching pairs calculated by using the affine model H.

[0077] Feature point repetition rate: the repetition rate is defined as the ratio of the number of repeated feature points in two images to the total number of detected points, and the repetition rate is defined as follows: N represents the number of matching pairs calculated by using the affine model H. repeat N represents the number of matching pairs calculated by using the affine model H.

[0078] Through two sets of experiments, qualitative and quantitative comparisons are made, as shown in Table 2, we find that the method proposed in the present application has higher quantitative indicators, and through Figure 4 The qualitative results show that the matching pairs obtained by the present application do not have the problem of local aggregation.

[0079] Table 2 quantitative analysis

[0080]

[0081] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A remote sensing stereo image feature matching method with edge enhancement and distribution equalization, characterized in that... Includes the following steps: I. Constructing Scale-Space Images Construct multi-scale spatial images corresponding to the reference image and the image to be registered, and use the unbiased ratio UDR edge detector to perform edge filtering on the spatial images at each scale to obtain the final multi-scale spatial image; II. Feature Point Extraction The extreme points in the final multi-scale spatial image are searched, and the obtained extreme points are candidate feature points. Then, the non-maximum suppression algorithm is used to homogenize the candidate feature points to complete the feature point extraction. III. Feature Point Registration Using the feature points as the center, four concentric circular regions of different sizes are selected from images of different scales to construct multi-scale feature descriptors. The multi-scale feature descriptor vectors are calculated. Then, the nearest neighbor distance ratio algorithm is used to obtain the initial feature matching point pairs. Finally, the random sampling consensus algorithm is used to delete the outliers in the initial feature matching point pairs, thus completing the feature point registration between the reference image and the image to be registered. The unbiased ratio UDR edge detector uses a combination of differential and ratio methods for edge detection. This method has the characteristics of unbiased edge localization by differential operation and constant false alarm rate under speckle noise by ratio operation. It ensures edge detection accuracy while minimizing noisy images, thus preserving better image information for subsequent feature point extraction. The calculation formula for the unbiased ratio UDR edge detector is as follows: Among them, the one-dimensional vertical filter The calculation formula is as follows; One-dimensional parallel filter The calculation formula is as follows; in, and Let x and y represent a one-dimensional vertical filter and a one-dimensional parallel filter, respectively. Let x and y represent the x and y coordinates of the pixel, respectively. Let α and β represent the peak value and attenuation rate of the one-dimensional vertical filter, respectively. This represents the unit step function, where ρ represents the initial layer scale space, η represents the scale ratio between two adjacent layers, and i represents the number of scale space layers. and parameters These represent the flatness and decay of the bell-shaped function, respectively. When constructing a multi-scale feature descriptor, the principal direction of the feature point is first calculated. Then, the scale image where the feature point is located, as well as the two upper scale images and the lower scale image adjacent to the scale image where the feature point is located, or the upper scale image and the two lower scale images are selected. After determining the scale space, the coordinate system of the feature point is rotated to the principal direction of the feature point according to the principal direction of the feature point. Then, with the feature point and its corresponding point on other scale images as the center, concentric circle regions of different sizes are selected. Then, the concentric circle regions are divided into multiple sub-regions, and the gradient histograms of eight directions in each sub-region are calculated to form the multi-scale feature descriptor vector.

2. The remote sensing stereo image feature matching method for edge enhancement and distribution equalization according to claim 1, characterized in that: The circular region on the scale image where the feature point is located has the smallest radius, which is denoted as the smallest circle, and the corresponding radius is denoted as R1. The radii of the other circular regions are R4=4R1, R2=0.5R4, and R3=0.73R4, respectively. The smallest circular region of the concentric circular regions is set as a separate sub-region, and the annular regions corresponding to the other concentric circular regions are divided into sub-regions every 45 degrees, for a total of 25 sub-regions.

3. The remote sensing stereo image feature matching method for edge enhancement and distribution equalization according to claim 1, characterized in that: When performing extreme point search, the selected neighborhood is set to a 5*5 or 7*7 pixel area.

Images