Image registration method based on improved ORB feature extraction and grid feature matching

Abstract

The invention discloses an image registration method based on improved ORB feature extraction and grid feature matching. The method comprises the following steps: inputting an image to be extracted; using an improved FAST algorithm to extract feature points of the to-be-spliced image; establishing a BRIEF descriptor, and obtaining a BRIEF descriptor with a direction; according to the feature point set of the to-be-spliced image, establishing a Delaunay triangular grid taking the feature points as vertexes; judging the similarity of triangles, and obtaining similar triangles formed by the feature points; matching the similar triangle vertexes by using a Hamming distance to obtain an optimal matching feature point pair; obtaining affine matrix parameters by using an SVD method; according to the image registration method provided by the invention, the image quality is ensured, the feature extraction efficiency and the extraction precision can be improved, wrong matching is reduced, the stability is improved, the real-time performance of video stitching is improved, the waste of memory resources is reduced, the equipment cost is reduced, the working efficiency is greatly improved, and the method is suitable for popularization and application. And a better method is provided for image splicing.

Description

Technical Field

[0001] The invention relates to the technical field of image registration methods, and in particular to an image registration method based on improved ORB feature extraction and grid feature matching. Background Art

[0002] Image registration is the core and key technology in image stitching technology, and the registration accuracy can directly affect the final stitching effect. Image registration refers to the spatial position alignment of two or more images, which is a mapping of images in space and intensity. Accurately finding the position of the overlapping area between two adjacent images, finding a suitable transformation model, and determining the transformation relationship between images are the key to image registration. Among them, image feature information is diverse, and therefore many feature extraction methods have been generated: straight line segments, feature structures, feature points (including corner points, high curvature points, etc.), edges, closed areas, and statistical features, etc. At present, there are many methods for image registration, but due to different application environments and the diversity of images to be stitched, many factors between images need to be considered. With the development of science and technology and society, the real-time and computational efficiency requirements of panoramic image stitching technology are getting higher and higher. Most traditional algorithms use Harris corner operator, SIFT operator and SURF operator. However, the extraction accuracy and efficiency of traditional algorithms are relatively low, the calculation amount is huge, they cannot meet the real-time work requirements and the image quality is poor. The feature extraction and matching of some high-resolution digital images are huge, which seriously affects the stitching efficiency, cannot meet the real-time work requirements and takes up a lot of CPU resources.

[0003] At ICCV in 2011, Ethan Rublee and Vincent Rabaud proposed the ORB algorithm, which is a new binary and efficient local feature descriptor. The ORB algorithm is divided into two steps. First, the improved FAST algorithm is used to detect feature points, and then the improved BRIEF algorithm is used to describe the feature points. Since both the FAST and BRIEF algorithms have the advantage of fast calculation speed, the ORB algorithm can efficiently and quickly complete the extraction of feature points in the feature extraction process. At the same time, the feature extraction of the ORB algorithm is also rotationally invariant and reduces the sensitivity to noise. Further improvements are needed to address the problems of slow extraction speed and poor accuracy caused by the huge number of image feature points extracted by traditional methods. Therefore, the development direction of image stitching technology, in addition to paying attention to the accuracy of image matching, should adopt methods with relatively low calculation amount and high efficiency in extracting features and matching information, greatly reducing the waste of CPU resources and improving the real-time performance of stitching.

[0004] There are several problems with the feature extraction process based on ORB. First, the number of feature points extracted by the FAST algorithm is large, and distribution clustering is prone to occur, so the probability of feature point mismatching will be greatly increased. Second, the FAST feature point detection algorithm can extract a large number of feature points, so when the ORB algorithm calculates the feature values ​​of the local area of ​​the image, the time consumption will increase abnormally, seriously affecting the feature extraction time. Third, the BRIEF feature description algorithm is sensitive to noise, so the integral image is used to compare pixel block pairs. Although it solves the noise sensitivity problem to a certain extent, the problem that the BRIEF algorithm affects feature extraction under uneven illumination has not been well solved. Summary of the invention

[0005] In view of the above technical problems in the related art, the present invention proposes an image registration method based on improved ORB feature extraction and grid feature matching, which can overcome the above shortcomings of the prior art.

[0006] To achieve the above technical objectives, the technical solution of the present invention is implemented as follows:

[0007] Image registration method based on improved ORB feature extraction and grid feature matching;

[0008] The image registration method based on improved ORB feature extraction and grid feature matching comprises the following steps:

[0009] S1, input the image to be extracted;

[0010] S2, using the improved FAST algorithm to extract feature points of the image to be stitched;

[0011] S3, using the idea of ​​BRIEF descriptor to establish BRIEF descriptor, and obtaining BRIEF descriptor with direction;

[0012] S4, establishing a Delaunay triangular mesh with the feature points as vertices according to the feature point set of the image to be stitched extracted in S02;

[0013] S5. Determine the triangle similarity based on the Delaunay triangle mesh established in S04, and obtain similar triangles composed of feature points;

[0014] S6. Use the Hamming distance to pair similar triangle vertices to obtain the best matching feature point pair;

[0015] S7. Use the SVD method to obtain the affine matrix parameters.

[0016] Furthermore, the step S2 uses an improved FAST algorithm to extract feature points of the images to be stitched, specifically comprising the following steps:

[0017] S201, set F 1 and F 2 is the image to be stitched. When detecting feature points of the image to be stitched, a given constant threshold ε is used. d , thus obtaining the point set P of the initial corner points, the total number is N;

[0018] S202. Let (x, y) be any point in the corner point set P. If, with this point as the center, within a 3×3 circular neighborhood, there is any point (x′, y′)∈P, and the Harris response function of this point is:

[0019] Cornerness=Det(M)-k(Trace(M)) 2 ,

[0020] If the response value is greater than that of the (x,y) point, remove the (x,y) point; loop the operation to complete the comparison of all corner points, and form the remaining points in the corner point set P into a set P′;

[0021] S203, setting a threshold value N', when the number of corner points in P' is less than N', all corner points in the set P' are used as available feature points for extraction; and when the number of corner points in P' is greater than N', N' points with the largest response values ​​in the set P' are selected as available feature points for extraction; wherein the value of N' is N / 2;

[0022] S204, adopting the strategy of image region division and adjacent feature point elimination to avoid clustering of feature points and ensure uniform distribution of feature points; evenly dividing the image into regions and setting the total number of feature points in each region to N thd , to ensure the stability of the total number of feature points; then sort the feature points extracted from each image area according to the size of the feature value, retaining N thd The point with the largest eigenvalue, other feature points are eliminated;

[0023] S205, setting the neighbor distance D thd To eliminate local clustering; for different images, set the number of feature points N thd =300 and normalized distance D thd =0.02; when the total number of extracted image block feature points is greater than N thd When extracting features, the features are automatically sorted by feature value, and non-significant feature points are filtered and removed to keep the feature points evenly distributed, thereby improving the efficiency of feature point extraction.

[0024] S206, F 1 and F 2 After the feature point extraction of the stitched image is completed, F 1 and F 2 A set of feature points and Feature points are described using three attribute values: The coordinates of the feature points in the image to be stitched are set to (x, y), and (x, y, σ) is the vector description of the feature points.

[0025] Furthermore, the step S3 adopts the idea of ​​BRIEF descriptor to establish BRIEF descriptor, which specifically includes the following steps:

[0026] S301. Define the moment of the patch as: m pq =∑ x，y x p y p I(x, y), where I(x, y) is the grayscale value of the image, and x, y are the positions relative to the FAST feature points; the radius of the circular neighborhood is r, x, y∈[-r, r];

[0027] S302, calculating the center of gravity of the plaque: The angle between the feature point and the center of gravity is defined as the direction θ of the FAST feature point, then:

[0028] S303, using the integral graph to compare several pairs of pixel blocks, set in a 31×31 pixel neighborhood, randomly select n d The sub-window is set to 5×5, and the pixels and in the window are compared; the center coordinates of the randomly selected pixel blocks are constructed into a matrix S, and then a rotation matrix R is constructed according to the direction θ obtained in S302 to rotate S to the corresponding direction: S θ =RS, get the binary string;

[0029] S304, define the test value τ of the image grayscale difference value as: Where p(x) represents the image spot p after filtering, at x = (u, v) T The pixel intensity at the binary detection τ is uniquely determined by a set of pairs, n d The binary string BRIEF descriptor for the dimension is defined as: Among them, nd is set to 256; the descriptor is a binary string, and the Hamming distance is used to pair the feature points; if the corresponding bit values ​​are the same, the Hamming distance is 0, otherwise it is 1; the more the number of identical bits, the greater the similarity of the descriptors.

[0030] Furthermore, the method for establishing the Delaunay triangular mesh in S4 specifically comprises the following steps:

[0031] S401, for the feature point set f 1 and f 2, sort and split all the points in the set (recursive processing), and divide the point set step by step until the number of feature points in the subset is no more than 3, and all subsets do not intersect with each other;

[0032] S402, constructing a Delaunay sub-triangulation network using the subsets after recursive segmentation;

[0033] S403, gradually merge the subsets from bottom to top to generate a Delaunay triangulated mesh of the entire point set.

[0034] Furthermore, the method for determining the triangle similarity in S5 specifically comprises the following steps:

[0035] S501, suppose the triangles in the reference image and the image to be registered are ΔABC and ΔA′B′C′, where A, B, C and A', B', C' are corresponding point pairs; the triangle similarity is I, and the similarities between the three angles are I A ,I B ,I C , the value of angle A is α, the value of angle A' is α', then the similarity between angle A and angle A' is: Where σ = α / 6;

[0036] S502, respectively calculate the similarity I of the other two pairs of interior angles B and I C ; The similarity between triangles ΔABC and ΔA′B′C′ is Set the similarity threshold ω, which is 0.75 in this method; when I is greater than the threshold ω, it is determined that triangles ΔABC and ΔA′B′C′ are similar.

[0037] Beneficial effects of the present invention: The present invention designs and develops an image registration method based on improved ORB feature extraction and grid feature matching. The extracted feature points have scale invariance and rotation invariance, and can be combined with the geometric information of the feature points for feature extraction, thereby reducing the sensitivity to noise, effectively improving the disadvantage of slow image feature extraction speed, and having good stability. This method is different from the traditional method of matching point pairs by only considering the geometric correlation between point pairs. By constructing a Delaunay triangle, the similarity of the triangle is judged, and the wrong matching points are eliminated. The present invention can improve the efficiency and accuracy of video splicing, increase the calculation speed, etc., and solve the problems of poor real-time video splicing effect and high memory resource usage. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 It is a flow chart of image registration according to the image registration method based on improved ORB feature extraction and grid feature matching described in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0041] It should be understood that, in the description of the embodiments of the present invention, the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the embodiments of the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the embodiments of the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present invention, the meaning of "several" is two or more, unless otherwise clearly and specifically defined.

[0042] like Figure 1As shown, according to the image registration method based on improved ORB feature extraction and grid feature matching described in the embodiment of the present invention, after studying the key technologies and principles of traditional image feature extraction and matching methods, the ORB feature extraction method is improved to address the problems of slow feature point extraction speed, low accuracy, and excessive memory resource usage. The ORB algorithm is mainly integrated with scale clustering analysis to achieve the purpose of improving feature extraction efficiency and extraction accuracy, and better stability, and solve the problems of high redundancy, easy mismatching, and missed matching of traditional image matching methods. In addition, after the improved ORB algorithm, grid feature matching is used instead of the traditional matching method to complete efficient and accurate feature matching.

[0043] The present invention is described in detail below in conjunction with the accompanying drawings. There are several problems with the feature extraction process based on ORB. First, the number of feature points extracted by the FAST algorithm is large, and distribution clustering is prone to occur, so the probability of feature point mismatching will be greatly increased. Second, the FAST feature point detection algorithm can extract a large number of feature points, so when the ORB algorithm calculates the feature values ​​of the local area of ​​the image, the time consumption will be abnormally increased, which seriously affects the feature extraction time. Third, the BRIEF feature description algorithm is more sensitive to noise, so an integral graph is used to compare pixel block pairs. Although the noise sensitivity problem is solved to a certain extent, the problem that the BRIEF algorithm affects feature extraction under the condition of uneven illumination has not been well solved.

[0044] In view of the above problems, the FAST feature point detection algorithm is improved. M is defined as the second-order moment proposed by the Harris operator to describe the gradient distribution in the neighborhood of a pixel point. The Harris corner point response function is: Cornerness = Det(M)-k(Trace(M)) 2 By combining the circular template with the Harris response value of the FAST corner point, the clustering phenomenon of corner point distribution is eliminated, and the corner points with unstable edge responses are eliminated, so as to improve the accuracy and speed of feature extraction and enhance the robustness of subsequent matching.

[0045] The specific implementation process of the improved ORB algorithm is as follows:

[0046] 1. Let F 1 and F 2 is the image to be stitched. When detecting feature points of the image to be stitched, a given constant threshold ε is used. d , thus obtaining the point set P of the initial corner points, with a total number of N.

[0047] 2. Let (x, y) be any point in the corner point set P. If, with this point as the center, within a 3×3 circular neighborhood, there is any point (x′, y′)∈P, and the Harris response value of this point is Cornerness=Det(M)-k(Trace(M)) 2 If the response value is greater than that of the (x, y) point, remove the (x, y) point; repeat the operation to complete the comparison of all corner points, and form the remaining points in the corner point set P into a set P′.

[0048] 3. Set the threshold N'. When the number of corner points in P' is less than N', all corner points in the set P' are used as available feature points for extraction. When the number of corner points in P' is greater than N', select the N' points with the largest response value in P' as available feature points for extraction. The value of N' can generally be N / 2.

[0049] 4. Use the strategy of image region division and adjacent feature point elimination to avoid clustering of feature points and ensure uniform distribution of feature points. Divide the image into regions evenly and set the total number of feature points in each region to N. thd , to ensure the stability of the total number of feature points. Then the feature points extracted from each image area are sorted according to the size of the feature value, and N thd The point with the largest eigenvalue is selected, and the other feature points are eliminated.

[0050] 5. Since each region retains N thd There are 3 significant feature points, so the distribution of image feature points is uniform. However, feature points in local areas of the image may still cluster, so the mismatch rate will increase during image matching. Therefore, the nearest neighbor distance D is set. thd To eliminate local clustering. For different images, set the number of feature points N thd =300 and normalized distance D thd =0.02. The number of feature points is related to factors such as image extraction speed, and the threshold can also be modified appropriately; the neighbor distance threshold is set to ensure successful extraction when the image texture is rich or the feature points are close. thd When extracting features, the features are automatically sorted by the size of the eigenvalues, and non-significant feature points are filtered and removed to keep the feature points evenly distributed, effectively improving the efficiency of feature point extraction and avoiding the problem of increased time consumption.

[0051] 6. F 1 and F 2 After the feature point extraction of the stitched image is completed, F 1 and F 2 A set of feature points and Feature points are described using three attribute values: The coordinates of the feature points in the image to be stitched are set to (x, y), and (x, y, σ) is the vector description of the feature points.

[0052] 7. After obtaining the feature points, the descriptor is established using the idea of ​​BRIEF descriptor.

[0053] To solve the problem that the ORB algorithm does not have rotation invariance, a direction is added to the BRIEF descriptor and rotated to obtain a directed BRIEF. First, the moment of the patch is defined as:

[0054] m pq =∑ x,y x p y p I(x,y) (1)

[0055] Where I(x,y) is the grayscale value of the image, x,y is the position relative to the FAST feature point, the circular neighborhood radius is r, and x,y∈[-r,r].

[0056] Then calculate the center of gravity of the patch:

[0057]

[0058] The angle between the feature point and the center of gravity is defined as the direction θ of the FAST feature point:

[0059]

[0060] In order to solve the noise impact caused by uneven illumination and improve real-time performance, an integral image is used to compare several pairs of pixel blocks. In a 31×31 pixel neighborhood, nd pixel blocks are randomly selected, and the subwindow is set to 5×5. The sum of pixels in the window is compared.

[0061] The center coordinates of the randomly selected pixel block are constructed into a matrix S, and then a rotation matrix R is constructed according to the direction θ obtained by formula (3) to rotate S to the corresponding direction:

[0062]

[0063] S θ =RS (5)

[0064] Get the binary string and define the test value τ of the image grayscale difference value as:

[0065]

[0066] In formula (6), p(x) represents the image spot p after filtering, at x = (u, v) TThe pixel intensity at . A binary detection τ is uniquely determined by a set of pairs, so the descriptor of the nd-dimensional binary string BRIEF is defined as:

[0067]

[0068] In this method, n d Set to 256. This descriptor is a binary string that uses the Hamming distance to pair feature points. If the corresponding bit values ​​are the same, the Hamming distance is 0, otherwise it is 1. The more identical bits there are, the greater the similarity of the descriptors.

[0069] The improved ORB algorithm obtains the BRIEF descriptor with direction according to the above formula.

[0070] 8. Feature matching based on Delaunay triangulated mesh

[0071] a) Using the previously accurately extracted feature point set f 1 and f 2 , divide the convex hull formed by the feature points, and construct the image feature information into a Delaunay triangle mesh. Plane triangulation set f 1 and f 2 The convex hull is constructed by recursively partitioning the point set, step by step until the subset contains only three discrete points that can form a triangle, and then merged step by step from bottom to top to form the final triangulated network.

[0072] 1) For the feature point set f 1 and f 2 , sort and split all the points in the set (recursive processing), and divide the point set level by level until the number of feature points in the subset is no more than 3, and all subsets do not intersect with each other.

[0073] 2) Use the recursively segmented subsets to construct a Delaunay sub-triangulation network.

[0074] 3) Merge subsets step by step from bottom to top to generate the Delaunay triangulated mesh of the entire point set.

[0075] Since the Delaunay triangulation network is unique, the feature point set f 1 and f 2 Has a unique Delaunay triangulation.

[0076] b) Determine the similarity of triangles, extract triangle pairs with greater similarity, and obtain accurate matching point pairs. The method for determining triangle similarity is as follows:

[0077] Suppose the triangles in the reference image and the image to be registered are ΔABC and ΔA′B′C′, where A, B, C and A', B', C' are corresponding point pairs. The triangle similarity is I, and the similarities between the three angles are I A ,I B ,I C , the value of angle A is α, the value of angle A' is α', then the similarity between angle A and angle A' is:

[0078]

[0079] Where σ = α / 6.

[0080] Then calculate the similarity I of the other two pairs of interior angles B and I C The similarity between triangles ΔABC and ΔA′B′C′ is Set the similarity threshold ω, which is 0.75 in this method. When I is greater than the threshold ω, it is determined that triangles ΔABC and ΔA′B′C′ are similar.

[0081] After obtaining similar triangle pairs, the triangle vertex pairs of the similar triangle pairs are paired using the Hamming distance to obtain the best matching feature point pairs, and finally the SVD (singular value decomposition) method is used to obtain the affine matrix parameters.

[0082] In summary, the image registration method based on improved ORB feature extraction and grid feature matching proposed in the present invention can improve feature extraction efficiency and extraction accuracy while ensuring image quality, reduce false matching and improve stability, which not only improves the real-time performance of video stitching, but also reduces the waste of memory resources, reduces equipment costs, greatly improves work efficiency, and provides a better method for image stitching.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An image registration method based on improved ORB feature extraction and grid feature matching, characterized in that: The following steps are involved: S1, input the image to be extracted; S2, using the improved FAST algorithm to extract feature points of the image to be stitched; S3, using the idea of ​​BRIEF descriptor to establish BRIEF descriptor, and obtaining BRIEF descriptor with direction; S4, establishing a Delaunay triangular mesh with the feature points as vertices according to the feature point set of the image to be stitched extracted in S02; S5. Determine the triangle similarity based on the Delaunay triangle mesh established in S04, and obtain similar triangles composed of feature points; S6. Use the Hamming distance to pair similar triangle vertices to obtain the best matching feature point pair; S7. Use the SVD method to obtain the affine matrix parameters.

2. The image registration method based on improved ORB feature extraction and grid feature matching according to claim 1, characterized in that: The step S2 uses the improved FAST algorithm to extract feature points of the images to be stitched, and specifically includes the following steps: S201, let F1 and F2 be the images to be stitched, and when detecting feature points of the images to be stitched, use a given constant threshold ε d , thus obtaining the point set P of the initial corner points, the total number is N; S202. Let (x, y) be any point in the corner point set P. If, with this point as the center, within a 3×3 circular neighborhood, there is any point (x′, y′)∈P, and the Harris response function of this point is: Cornerness=Det(M)-k(Trace(M)) 2 , If the response value is greater than that of the (x,y) point, remove the (x,y) point; loop the operation to complete the comparison of all corner points, and form the remaining points in the corner point set P into a set P′; S203, setting a threshold value N', when the number of corner points in P' is less than N', all corner points in the set P' are used as available feature points for extraction; and when the number of corner points in P' is greater than N', N' points with the largest response values ​​in the set P' are selected as available feature points for extraction; wherein the value of N' is N / 2; S204, adopting the strategy of image region division and adjacent feature point elimination to avoid clustering of feature points and ensure uniform distribution of feature points; evenly dividing the image into regions and setting the total number of feature points in each region to N thd , to ensure the stability of the total number of feature points; then sort the feature points extracted from each image area according to the size of the feature value, retaining N thd The point with the largest eigenvalue, other feature points are eliminated; S205, setting the neighbor distance D thd To eliminate local clustering; for different images, set the number of feature points N thd =300 and normalized distance D thd =0.02; when the total number of extracted image block feature points is greater than N thd When extracting features, the features are automatically sorted by feature value, and non-significant feature points are filtered and removed to keep the feature points evenly distributed, thereby improving the efficiency of feature point extraction. S206: Extract feature points of the images to be spliced ​​F1 and F2, and obtain feature point sets of F1 and F2 respectively. and Feature points are described using three attribute values: The coordinates of the feature points in the image to be stitched are set to (x, y), and (x, y, σ) is the vector description of the feature points.

3. The image registration method based on improved ORB feature extraction and grid feature matching according to claim 1, characterized in that: The step S3 uses the idea of ​​BRIEF descriptor to establish the BRIEF descriptor, which specifically includes the following steps: S301. Define the moment of the patch as: m pq =∑ x,y x p y p I(x,y), where I(x,y) is the grayscale value of the image, and x,y are the positions relative to the FAST feature points; the radius of the circular neighborhood is r, x,y∈[-r,r]; S302, calculating the center of gravity of the plaque: The angle between the feature point and the center of gravity is defined as the direction θ of the FAST feature point, then: S303, using the integral graph to compare several pairs of pixel blocks, set in a 31×31 pixel neighborhood, randomly select n d The sub-window is set to 5×5, and the pixels and in the window are compared; the center coordinates of the randomly selected pixel blocks are constructed into a matrix S, and then a rotation matrix R is constructed according to the direction θ obtained in S302 to rotate S to the corresponding direction: S θ =RS, get the binary string; S304, define the test value τ of the image grayscale difference value as: Where p(x) represents the image spot p after filtering, at x = (u, v) T The pixel intensity at the binary detection τ is uniquely determined by a set of pairs, n d The binary string BRIEF descriptor for the dimension is defined as: Among them, nd is set to 256; the descriptor is a binary string, and the Hamming distance is used to pair the feature points; if the corresponding bit values ​​are the same, the Hamming distance is 0, otherwise it is 1; the more the number of identical bits, the greater the similarity of the descriptors.

4. The image registration method based on improved ORB feature extraction and grid feature matching according to claim 2, characterized in that: The method for establishing the Delaunay triangular mesh in S4 specifically comprises the following steps: S401, for feature point sets f1 and f2, sort and split all points in the sets (recursive processing), and divide the point sets step by step until the number of feature points in the subsets is no more than 3, and all subsets do not intersect with each other; S402, constructing a Delaunay sub-triangulation network using the subsets after recursive segmentation; S403, gradually merge the subsets from bottom to top to generate a Delaunay triangulated mesh of the entire point set.

5. The image registration method based on improved ORB feature extraction and grid feature matching according to claim 4, characterized in that: The method for determining the triangle similarity in S5 specifically comprises the following steps: S501, suppose the triangles in the reference image and the image to be registered are ΔABC and ΔA′B′C′, where A, B, C and A', B', C' are corresponding point pairs; the triangle similarity is I, and the similarities between the three angles are I A ,I B ,I C , the value of angle A is α, the value of angle A' is α', then the similarity between angle A and angle A' is: Where σ = α / 6; S502, respectively calculate the similarity I of the other two pairs of interior angles B and I C ; The similarity between triangles ΔABC and ΔA′B′C′ is Set the similarity threshold ω, which is 0.75 in this method; when I is greater than the threshold ω, it is determined that triangles ΔABC and ΔA′B′C′ are similar.

Images