Multi-modal remote sensing image registration method based on co-occurrence filtering and histogram of oriented gradients

Abstract

The invention relates to the technical field of multi-modal remote sensing image registration, and aims to solve the problems that a multi-modal remote sensing image imaging sensor and an imaging mechanism are greatly different, and the registration performance and robustness of an image registration method are reduced in different-modal remote sensing image registration tasks. The invention provides a multi-modal remote sensing image registration method based on co-occurrence filtering and a histogram of oriented gradient, and the method comprises the steps: carrying out the rough matching of a reference image and a to-be-registered image, obtaining the spatial relation constraint information, carrying out the adjustment of a rough matching homonymy point pair through a histogram of oriented gradient, improving the matching precision, updating a homography matrix, and achieving the registration of a multi-modal remote sensing image. Resampling the to-be-registered image to complete image registration; the multi-modal remote sensing image registration method can adapt to a multi-modal remote sensing image registration scene and has good robustness on rotation and scale change of the multi-modal remote sensing image, an image registration experiment is carried out on the multi-modal remote sensing image, and an experiment result shows that registration precision and robustness can be kept in different scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal remote sensing image registration, and more specifically, to a multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram. Background Art

[0002] Different modal remote sensing images have their own unique characteristics due to the use of different sensors and technologies. For example, visible light remote sensing images were developed the earliest, the technology is mature, and it conforms to human visual habits and understanding patterns. It has the advantages of low imaging cost, high image resolution, and rich data. However, visible light is a passive imaging, which is easily affected by weather conditions such as clouds and haze, limiting the time and quality of observation and making it impossible to achieve all-weather observation. Synthetic Aperture Radar (SAR) images use the active imaging technology of synthetic aperture radar and have all-weather and all-day observation capabilities. They are widely used in the assessment and statistics of building damage after disasters such as earthquakes, mudslides, and mountain torrents. However, the imaging morphology of SAR images for ground objects is not as intuitive as that of optical images, and the classification of ground objects is relatively difficult and requires processing and interpretation. Infrared remote sensing images mainly use infrared radiation information in the atmospheric window and have special information such as heat distribution and vegetation growth status. For example, in the military field, visible light images and infrared images are fused for precise target strikes. However, atmospheric absorption and scattering will affect the observation quality and spatial resolution of the infrared band, and it is impossible to penetrate the clouds to obtain surface information. Compared with optical images, the spatial resolution of infrared images is relatively low.

[0003] It can be seen that there are great differences in the imaging mechanisms of remote sensing images of different modalities, but there are rich complementary characteristics between their images and they have broad application potential. If multimodal remote sensing images can be fused, analyzed and understood, more information can be mined than single-modal remote sensing images, and problems that cannot be solved by relying solely on single-modal images can be solved. Multimodal remote sensing image matching is the pre-step and key link for multimodal remote sensing image fusion analysis. Traditional remote sensing image registration methods generally only perform image matching once for the reference image and the image to be registered, obtain the mapping relationship between the two images based on the matched pairs of same-name points, and resample the image to be registered to complete the image registration. However, there are obvious differences in the imaging sensors and imaging mechanisms of multimodal remote sensing images. There are not only complex illumination and contrast changes between images, but also different types of noise pollution, which makes it difficult for image registration methods to ensure that the performance and robustness remain robust in different modal image registration scenarios. Summary of the invention

[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram, which can not only adapt to the multimodal remote sensing image registration scenario, but also has good robustness to the rotation and scale changes of the multimodal remote sensing images.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A multimodal remote sensing image registration method based on co-occurrence filtering and directional gradient histogram is proposed. The co-occurrence filters with different parameters are used to smooth the image to be registered and the reference influence, and the co-occurrence filter image pyramid scale space is generated. The initial homography matrix, rotation angle, scale constraint spatial relationship constraint information and directional gradient histogram after rough matching of local feature points extracted on the image pyramid are used to improve the matching accuracy. The image to be registered is resampled to complete the image registration task. The method specifically includes the following steps:

[0007] Step 1. Input the reference image and the image to be registered, and use the co-occurrence filter to smooth the image details of the two images while maintaining the influence edge;

[0008] Step 2. Construct image pyramids for the reference image and the image to be registered, respectively, use co-occurrence filters with different parameters to smooth the "group" image and "layer" image in the pyramid, extract local feature points from the reference image and the image to be registered, and use the non-maximum suppression method to screen the extracted local feature points;

[0009] Step 3. Construct the local feature descriptor of the image, use the Euclidean distance as the matching criterion, perform rough matching on the reference image and the image to be registered, and obtain the rough matching same-name point pairs between the reference image and the image to be registered;

[0010] Step 4. Obtain the spatial relationship constraint information between the reference image and the image to be registered based on the result of the rough matching: initial homography matrix, rotation angle, scale constraint;

[0011] Step 5. Based on the spatial constraint relationship information, the directional gradient histogram is used to improve the matching accuracy of the coarse matching same-name point pairs between the reference image and the image to be registered, and the adjusted same-name point pairs are used to update the mapping relationship homography matrix between the reference image and the image to be registered;

[0012] Step 6. Use the mapping relationship homography matrix between the adjusted reference image and the image to be registered obtained in step 5 to resample the image to be registered to complete the image registration.

[0013] Furthermore, the co-occurrence filter formula is as follows:

[0014]

[0015]

[0016] Where, Jp and Iq are the output and input pixel values, p and q are pixel indices; is the contribution weight of pixel q to the output of pixel p; M(a,b) is a 256×256 dimensional matrix, obtained by dividing the co-occurrence matrix C(a,b) of values ​​a and b by the frequency of values ​​a and b; C(a,b) represents the co-occurrence matrix of values ​​a and b; h(a) and h(b) represent the number of image input pixel values ​​equal to a and equal to b, respectively; d(p,q) is the Euclidean distance between pixel p and pixel q in the image plane.

[0017] Furthermore, the co-occurrence filter window size has the following relationship with the "group" image and "layer" image of the image pyramid:

[0018]

[0019] Wherein, o represents the oth group of the image difference pyramid; i represents the i-th layer image in the oth group of the image difference pyramid; σ s,i represents the parameters of the i-th layer image co-occurrence filter, σ i Represents the filtering parameters of the co-occurrence matrix C(a,b) of value a and value b; round() is the integer function.

[0020] Furthermore, in step 4, the initial homography matrix between the reference image and the image to be registered is obtained using the coarse matching same-name point pairs between the reference image and the image to be registered, and the rotation angle between the reference image and the image to be registered is separated from the initial homography matrix using the singular value decomposition method; the scale constraints of the reference image and the image to be registered are obtained through a voting mechanism using the coarse matching same-name point pairs.

[0021] Furthermore, in step 4, the singular value decomposition method is used to separate the rotation angle between the reference image and the image to be registered from the homography matrix. The specific steps are as follows:

[0022] Step 4.1 Normalize the homography matrix:

[0023] H_normalized=H / H(3,3)

[0024] H: initial homography matrix; H_normalized: normalized homography matrix;

[0025] Step 4.2 Decompose the homography matrix using SVD:

[0026] [U,~,V]=svd(H_normalized(1:2,1:2))

[0027] svd(): singular value decomposition function;

[0028] Step 4.3 Extract the rotation part of the homography matrix: R = U*V';

[0029] Step 4.4 obtains the rotation angle between the reference image and the image to be registered:

[0030] theta=atan2d(R(2,1),R(1,1))

[0031] theta: the rotation angle between the two images, atan2d: inverse tangent function, find the rotation angle.

[0032] Furthermore, in step 4, the scale constraint between the reference image and the image to be registered is obtained through a voting mechanism. The specific process is as follows: according to the coarse matching results of the same-name point pairs between the reference image and the image to be registered, the "group" image of the pyramid where the matching points are located is voted, and the "group" image with the most votes is the "group" in the scale constraint; the same-name point pairs in the "group" are screened, and the "layer" images of the same-name point pairs are voted, and the "layer" image with the most votes is the "layer" in the scale constraint.

[0033] Furthermore, in step 5, the matching accuracy of the coarse matching homonymous points is improved based on the spatial relationship constraint information, as follows:

[0034] Step 5.1 After roughly matching the reference image and the image to be registered to obtain roughly matched same-name point pairs, the gradient features of the same-name point support area of ​​the reference image are extracted and a gradient direction histogram is constructed;

[0035] Step 5.2 extracts the gradient features of the region supporting the features of the same-name points in the image to be registered and constructs a gradient direction histogram. Iteratively adjusts the same-name points in the image to be registered within the spatial relationship constraint range, and measures the similarity between the same-name points in the image to be registered and the reference image until the iteration is completed and the same-name points with the highest structural similarity are obtained.

[0036] Step 5.3 repeats steps 5.1 to 5.2 to adjust all the same-name points in the image to be registered, improve the matching accuracy of the same-name point pairs between the reference image and the image to be registered, and update the mapping relationship homography matrix between the reference image and the image to be registered.

[0037] In summary, the invention has the following beneficial effects:

[0038] The present invention firstly performs coarse matching on a reference image and an image to be registered to obtain coarse matching same-name point pairs, obtains spatial relationship constraint information between the reference image and the image to be registered from the coarse matching result: initial homography matrix, rotation angle, scale constraint, and uses gradient direction histogram to improve the matching accuracy of coarse matching same-name point pairs between the reference image and the image to be registered under the spatial relationship constraint information; then updates the homography matrix between the reference image and the image to be registered through the same-name point pairs with improved accuracy; finally, uses the updated homography matrix to resample the image to be registered to complete image registration.

[0039] The present invention aims at the problem of unrobust performance and robustness of image registration methods in multimodal remote sensing image registration scenarios, and proposes to first perform rough matching on the reference image and the image to be registered. The rough matching is mainly to obtain the spatial relationship constraint information between the reference image and the image to be registered, and the matching accuracy is low. Then, based on the spatial relationship constraint information, the gradient direction histogram is used to adjust the rough matching same-name point pairs between the reference image and the image to be registered, improve the matching accuracy, update the homography matrix, and complete the image registration by resampling the image to be registered. The experimental results show that the present invention can effectively improve the matching accuracy of the rough matching same-name point pairs between the reference image and the image to be registered to improve the image registration accuracy. It is not only applicable to multimodal remote sensing image registration scenarios such as visible light, infrared, SAR, lidar images, and maps, but also has good robustness to the rotation changes and scale changes between multimodal remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flow chart of the present invention;

[0041] Figure 2 A flow chart of improving the accuracy of the homography matrix between the reference image and the image to be registered according to the present invention;

[0042] Figure 3 The image rough matching result and the image chessboard registration result of Example 1;

[0043] Figure 4 The image matching accuracy is improved and the image chessboard registration result is shown in Example 1;

[0044] Figure 5 The image rough matching result and the image chessboard registration result of Example 2;

[0045] Figure 6 The image matching accuracy of Example 2 is improved and the image chessboard registration result is shown;

[0046] Figure 7 The image rough matching result and the image chessboard registration result of Example 3;

[0047] Figure 8The image matching accuracy is improved and the image chessboard registration result is shown in Example 3;

[0048] Fig. 9 The image rough matching result and the image chessboard registration result of Example 4;

[0049] Fig.10 The image matching accuracy is improved and the image chessboard registration result is shown in Example 4;

[0050] Fig.11 The image rough matching result and the image chessboard registration result of Example 5;

[0051] Fig.12 The image matching accuracy is improved and the image chessboard registration result is shown in Example 5;

[0052] Fig.13 The image rough matching result and the image chessboard registration result of Example 6;

[0053] Fig.14 This is the result after the image matching accuracy is improved and the image chessboard registration result of Example 6. DETAILED DESCRIPTION

[0054] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0055] It should be noted that, for the sake of convenience, the directions described below are consistent with the directions of the drawings themselves, but do not limit the structure of the present invention.

[0056] like Figures 1 to 14 As shown, the present invention discloses a multimodal remote sensing image registration method based on co-occurrence filtering and directional gradient histogram, which smoothes the image to be registered and the reference influence through co-occurrence filters with different parameters, generates a co-occurrence filter image pyramid scale space, uses the initial homography matrix, rotation angle, scale constraint space relationship constraint information and directional gradient histogram extracted from the image pyramid after rough matching of local feature points to improve the matching accuracy, resamples the image to be registered, and completes the image registration task, which specifically includes the following steps:

[0057] Step 1. Input the reference image and the image to be registered, and use the co-occurrence filter to smooth the image details of the two images while maintaining the influence edge; pre-process the reference image and the image to be registered, and complete the construction of the scale space based on the co-occurrence filter on the reference image and the image to be registered.

[0058] Step 2. Construct image pyramids for the reference image and the image to be registered respectively, use co-occurrence filters with different parameters to smooth the "group" image and "layer" image in the pyramid, extract the gradient features of the support area of ​​the same-name points in the image to be registered and construct the directional gradient histogram. The formula of the co-occurrence filter is as follows:

[0059]

[0060] Where, Jp and Iq are the output and input pixel values, p and q are pixel indices; is the contribution weight of pixel q to the output of pixel p; M(a,b) is a 256×256 dimensional matrix, obtained by dividing the co-occurrence matrix C(a,b) of values ​​a and b by the frequency of values ​​a and b; C(a,b) represents the co-occurrence matrix of values ​​a and b; h(a) and h(b) represent the number of image input pixel values ​​equal to a and equal to b, respectively; d(p,q) is the Euclidean distance between pixel p and pixel q in the image plane.

[0061] The co-occurrence filter window size has the following relationship with the "group" and "layer" images of the image pyramid:

[0062]

[0063] Wherein, o represents the oth group of the image difference pyramid; i represents the i-th layer image in the oth group of the image difference pyramid; σ s,i represents the parameters of the i-th layer image co-occurrence filter; σ i Represents the filtering parameters of the co-occurrence matrix C(a,b) of value a and value b; round() is the integer function.

[0064] Local feature points are extracted from the reference image and the image to be registered respectively, and the extracted local feature points are screened using the non-maximum suppression method; in step 2, the local feature points obtained do not correspond, and the local feature points are corresponded in subsequent steps.

[0065] Step 3. Extract local feature points in the scale space, construct rotation and scale invariant descriptors in the feature point support area, use Euclidean distance as the matching criterion, roughly match the reference image and the image to be registered, and obtain roughly matched pairs of same-name points between the reference image and the image to be registered; the correspondence between the local feature points affected by the reference image and the image to be registered is achieved through step 3, and the corresponding erroneous local feature points are matched with the corresponding correct local feature points in this step.

[0066] Step 4. Obtain the spatial relationship constraint information between the reference image and the image to be registered based on the result of coarse matching: initial homography matrix, rotation angle, scale constraint; use the coarse matching same-name point pairs between the reference image and the image to be registered to obtain the initial homography matrix between the reference image and the image to be registered, and use the singular value decomposition method to separate the rotation angle between the reference image and the image to be registered from the initial homography matrix; use the coarse matching same-name point pairs to obtain the scale constraint of the reference image and the image to be registered through a voting mechanism, and the rotation angle and scale constraint can be calculated simultaneously.

[0067] Use the singular value decomposition method to separate the rotation angle between the reference image and the image to be registered from the homography matrix. The specific steps are as follows:

[0068] Step 4.1 Normalize the homography matrix:

[0069] H_normalized=H / H(3,3)

[0070] H: initial homography matrix; H_normalized: normalized homography matrix.

[0071] Step 4.2 Decompose the homography matrix using SVD:

[0072] [U,~,V]=svd(H_normalized(1:2,1:2))

[0073] svd(): singular value decomposition function;

[0074] Step 4.3 Extract the rotation part of the homography matrix: R = U*V';

[0075] Step 4.4 obtains the rotation angle between the reference image and the image to be registered:

[0076] theta=atan2d(R(2,1),R(1,1))

[0077] theta: the rotation angle between the two images, atan2d: inverse tangent function, find the rotation angle.

[0078] The scale constraint between the reference image and the image to be registered is obtained through a voting mechanism. The specific process is as follows: according to the coarse matching results of the same-name point pairs between the reference image and the image to be registered, the "group" image of the pyramid where the matching points are located is voted, and the "group" image with the most votes is the "group" in the scale constraint; the same-name point pairs in the "group" are screened, and the "layer" images of the same-name point pairs are voted, and the "layer" image with the most votes is the "layer" in the scale constraint.

[0079] Step 5. Based on the spatial constraint relationship information, the directional gradient histogram is used to improve the matching accuracy of the coarse matching of the same-name point pairs between the reference image and the image to be registered. The method is as follows:

[0080] Step 5.1: After roughly matching the reference image and the image to be registered to obtain roughly matched same-name point pairs, the gradient features of the same-name point support area of ​​the reference image are extracted and a gradient direction histogram is constructed.

[0081] Step 5.2 extracts the gradient features of the feature support area of ​​the same-name points in the image to be registered and constructs the gradient direction histogram of the feature support area of ​​the same-name points in the image to be registered. Iteratively adjusts the same-name points in the image to be registered within the spatial relationship constraint range, and measures the similarity between the same-name points in the image to be registered and the reference image until the iteration is completed to obtain the same-name points with the highest structural similarity.

[0082] Step 5.3 repeats steps 5.1 to 5.2 to adjust all the same-name points in the image to be registered, improve the matching accuracy of the same-name point pairs between the reference image and the image to be registered, and update the mapping relationship homography matrix between the reference image and the image to be registered.

[0083] The adjusted same-name point pairs are used to update the homography matrix of the mapping relationship between the reference image and the image to be registered.

[0084] Step 6. Use the mapping relationship homography matrix between the adjusted reference image and the image to be registered obtained in step 5 to resample the image to be registered to complete the image registration.

[0085] The present invention constructs scale space based on co-occurrence filters on a reference image and an image to be registered respectively; extracts local feature points on the scale space, and constructs rotation and scale invariant descriptors in a feature point support region; performs coarse matching on the reference image and the image to be registered, and obtains spatial relationship constraint information between the two images, such as an initial homography matrix, rotation constraints, and scale constraints; constructs directional gradient histograms for coarse matching same-name points of the reference image and the image to be registered respectively; adjusts same-name point pairs using the directional gradient histograms of same-name point pairs based on the spatial relationship constraint information and using a similarity criterion to improve the matching accuracy of the same-name point pairs; calculates a mapping relationship homography matrix between the reference image and the image to be registered according to the adjusted same-name point pairs, and resamples the image to be registered to complete the image registration task.

[0086] like Figures 3 to 14 As shown, the present invention carried out an image registration experiment on multimodal remote sensing images. The experimental results show that the present invention can maintain good registration accuracy and robustness in visible light, infrared, lidar, SAR, map and other scenarios.

[0087] from Figure 3 and Figure 4It can be seen that after the present invention improves the matching accuracy of the reference image and the image to be registered, the image registration accuracy is significantly improved. Figure 3 and Figure 4 Shown in red circle on the right.

[0088] from Figures 3 to 10 It can be seen that after the present invention improves the matching accuracy of the reference image and the image to be registered, the image registration accuracy is significantly improved. Figure 3 and Figure 4 , Figure 5 and Figure 6 , Figure 7 and Figure 8 , Fig. 9 and Fig.10 Shown in red circle in the right image.

[0089] Fig.11 and Fig.12 The images are rotated 90 degrees from Fig.11 and Fig.12 It can be seen that there is a 90-degree rotation change between the reference image and the image to be registered, but the method proposed in the present invention can still perform high-precision image registration, and when the coarse matching accuracy is already very good, when the gradient direction histogram is used to adjust the same-name points between the reference image and the image to be registered, it will not affect the coarse matching image registration accuracy.

[0090] Fig.13 and Fig.14 There is also a rotation change of nearly 90 degrees between the reference image and the image to be registered. It can be seen from the chessboard registration result on the right that the present invention can not only cope with large-angle image registration scenes but also improve the matching accuracy between the reference image and the image to be registered through the gradient direction histogram, thereby further improving the image registration accuracy, as shown by the red circle in the figure on the right.

[0091] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram, characterized in that: The co-occurrence filters with different parameters are used to smooth the image to be registered and the reference influence, and the co-occurrence filter image pyramid scale space is generated. The initial homography matrix, rotation angle, scale constraint spatial relationship constraint information, and directional gradient histogram after rough matching of local feature points extracted on the image pyramid are used to improve the matching accuracy. The image to be registered is resampled to complete the image registration task. The specific steps include: Step 1. Input the reference image and the image to be registered, and use the co-occurrence filter to smooth the image details of the two images while maintaining the influence edge; Step 2. Construct image pyramids for the reference image and the image to be registered, respectively. Use co-occurrence filters with different parameters to smooth the "group" image and "layer" image in the pyramid. Extract local feature points from the reference image and the image to be registered, respectively. Use the non-maximum suppression method to screen the extracted local feature points. Step 3. Construct the local feature descriptor of the image, use the Euclidean distance as the matching criterion, perform rough matching on the reference image and the image to be registered, and obtain the rough matching same-name point pairs between the reference image and the image to be registered; Step 4. Obtain the spatial relationship constraint information between the reference image and the image to be registered based on the result of the rough matching: initial homography matrix, rotation angle, scale constraint; Step 5. Based on the spatial constraint relationship information, the directional gradient histogram is used to improve the matching accuracy of the coarse matching same-name point pairs between the reference image and the image to be registered, and the adjusted same-name point pairs are used to update the mapping relationship homography matrix between the reference image and the image to be registered; Step 6. Use the mapping relationship homography matrix between the adjusted reference image and the image to be registered obtained in step 5 to resample the image to be registered to complete the image registration.

2. The multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram according to claim 1, characterized in that: The co-occurrence filter formula is as follows: Where, Jp and Iq are the output and input pixel values, p and q are pixel indices; is the contribution weight of pixel q to the output of pixel p; M(a,b) is a 256×256 dimensional matrix, obtained by dividing the co-occurrence matrix C(a,b) of values ​​a and b by the frequency of values ​​a and b; C(a,b) represents the co-occurrence matrix of values ​​a and b; h(a) and h(b) represent the number of image input pixel values ​​equal to a and equal to b, respectively; d(p,q) is the Euclidean distance between pixel p and pixel q in the image plane.

3. The multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram according to claim 1, characterized in that: The co-occurrence filter window size has the following relationship with the "group" image and "layer" image of the image pyramid: Wherein, o represents the oth group of the image difference pyramid; i represents the i-th layer image in the oth group of the image difference pyramid; σ s,i represents the parameters of the i-th layer image co-occurrence filter, σ i Represents the filtering parameters of the co-occurrence matrix C(a,b) of value a and value b; round() is the integer function.

4. The multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram according to claim 1, characterized in that: In step 4, the initial homography matrix between the reference image and the image to be registered is obtained by using the coarse matching same-name point pairs between the reference image and the image to be registered, and the rotation angle between the reference image and the image to be registered is separated from the initial homography matrix by using the singular value decomposition method; the scale constraint of the reference image and the image to be registered is obtained by using the coarse matching same-name point pairs through a voting mechanism.

5. The multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram according to claim 4, characterized in that: In step 4, the rotation angle between the reference image and the image to be registered is separated from the homography matrix using the singular value decomposition method. The specific steps are as follows: Step 4.1 Normalize the homography matrix: H_normalized=H / H(3,3) H: initial homography matrix; H_normalized: normalized homography matrix; Step 4.2 Decompose the homography matrix using SVD: [U,~,V]=svd(H_normalized(1:2,1:2)) svd(): singular value decomposition function; Step 4.3 Extract the rotation part of the homography matrix: R = U*V'; Step 4.4 obtains the rotation angle between the reference image and the image to be registered: theta=atan2d(R(2,1),R(1,1)) theta: the rotation angle between the two images, atan2d: inverse tangent function, find the rotation angle.

6. The multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram according to claim 4, characterized in that: In step 4, the scale constraint between the reference image and the image to be registered is obtained through a voting mechanism, and the specific process is as follows: according to the coarse matching result of the same-name point pairs between the reference image and the image to be registered, the "group" image of the pyramid where the matching points are located is voted, and the "group" image with the most votes is the "group" in the scale constraint; the same-name point pairs in the "group" are screened, and the "layer" images of the same-name point pairs are voted, and the "layer" image with the most votes is the "layer" in the scale constraint.

7. The multimodal remote sensing image registration method based on co-occurrence filtering and oriented gradient histogram according to claim 1, characterized in that: In step 5, the matching accuracy of the roughly matched homonymous points is improved based on the spatial relationship constraint information, and the method is as follows: Step 5.1 After roughly matching the reference image and the image to be registered to obtain roughly matched same-name point pairs, the gradient features of the same-name point support area of ​​the reference image are extracted and a gradient direction histogram is constructed; Step 5.2 extracts the gradient features of the region supporting the features of the same-name points in the image to be registered and constructs a gradient direction histogram. Iteratively adjusts the same-name points in the image to be registered within the spatial relationship constraint range, and measures the similarity between the same-name points in the image to be registered and the reference image until the iteration is completed and the same-name points with the highest structural similarity are obtained. Step 5.3 repeats steps 5.1 to 5.2 to adjust all the same-name points in the image to be registered, improve the matching accuracy of the same-name point pairs between the reference image and the image to be registered, and update the mapping relationship homography matrix between the reference image and the image to be registered.

Images