Improved image registration algorithm based on SURF-ORB
By extracting ORB feature points in the SURF scale space and combining improved matching and classification algorithms, the problem of speed and real-time performance of the image registration algorithm is solved, and efficient and accurate image registration effect is achieved.
Patent Information
- Application Number
- CN202411260559.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2024-12-17
AI Technical Summary
The existing image registration algorithms are difficult to improve speed and real-time while ensuring accuracy. Especially in cases where scale changes are involved, the algorithm has high complexity and poor real-time performance.
Combining SURF and ORB algorithms, ORB feature points are extracted in the SURF scale space, and feature points are described using binary characters. The nearest neighbor method matching and improved K-means and RANSAC algorithms are used to perform feature point classification and transformation model calculations, and the matching points are filtered to Improve image registration efficiency.
It realizes that while ensuring high accuracy, reduce the calculation amount and improve real-time performance, adapt to situations where real-time performance is high, the algorithm has moderate calculation amount and higher accuracy and efficiency.
Smart Images
Figure CN119152002A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of image processing, and in particular to an improved image registration algorithm based on SURF-ORB. Background Art
[0002] Image registration is a technology that calculates the spatial transformation model based on the correspondence between the features of several small-field-of-view images of the same scene, aligns the overlapping areas, and synthesizes a large-field-of-view image. It is widely used in military reconnaissance, remote sensing technology, medical imaging, satellite imaging and other fields. The ORB algorithm is simple and has a very fast operation speed, but it does not have scale invariance and low accuracy, and is not suitable for occasions with scale changes. The SURF algorithm has scale and rotation invariance, strong anti-noise ability, and good stability. However, the algorithm complexity is high, the operation time is long, and the real-time performance is poor. Therefore, combining SURF with the ORB algorithm to extract ORB feature points in the SURF scale space can make the feature points scale invariant, enhance the accuracy of the algorithm, reduce the amount of calculation, improve the efficiency of the algorithm, and be more suitable for occasions with high real-time requirements. Summary of the invention
[0003] In order to solve the problem of improving image registration speed, reducing calculation amount and improving real-time performance under the condition of ensuring image registration accuracy, the present invention designs an improved image registration algorithm based on SURF-ORB.
[0004] An improved image registration algorithm based on SURF-ORB includes the following steps:
[0005] Step 1: Obtain two images to be registered, preprocess the images, and remove noise interference;
[0006] Step 2: Construct SURF scale space;
[0007] Step 3: Extract ORB feature points;
[0008] Step 4: Determine the main direction of the feature point;
[0009] Step 5: Generate a feature vector of binary characters to describe the feature points;
[0010] Step 6: Match feature points using the nearest neighbor method;
[0011] Step 7: Use the local entropy difference, Euclidean distance, and slope of the line connecting the matching points to select the matching points;
[0012] Step 8: Use the improved K-means algorithm to classify the feature points and classify the feature points in the same area into one category;
[0013] Step 9: Calculate the transformation model using the improved RANSAC algorithm;
[0014] Step 10: Image registration.
[0015] Furthermore, the specific content of step 2 to construct the SURF scale space is:
[0016] The SURF algorithm uses the Hessian matrix to detect feature points. The scale is The Hessian matrix is:
[0017] H ( X , σ ) = [ L xx ( X , σ ) L xy ( X , σ ) L xy ( X , σ ) L yy ( X , σ ) ] (1)
[0018] To simplify the algorithm and improve the operation speed, the SURF algorithm uses a step-by-step box filter. , , replace , , Convolution of integral images creates a scale space.
[0019] Furthermore, the specific content of extracting ORB feature points in step 3 is:
[0020] At each layer of the SURF scale space, construct A circle with center 3 and radius 1. Compared with the 16 points on the circumference, if is the extreme point of at least 9 consecutive points, and then Compare the 26 points in the neighborhood. If it is still an extreme point, then is the ORB corner point.
[0021] To avoid too dense distribution of feature points, the ORB response value of each corner point is calculated in the 8-neighborhood centered on the ORB corner point, and the corner point with the largest ORB response value is retained as the feature point.
[0022] Furthermore, the specific content of step 4 for determining the main direction of the feature point is:
[0023] Construct a neighborhood centered on the feature point, calculate the neighborhood centroid, and the direction of the line connecting the feature point and the centroid is the main direction of the feature point.
[0024] Neighborhood moments: (2)
[0025] Centroid: (3)
[0026] Angle of the main direction of the feature point: (4)
[0027] Furthermore, the specific content of the feature vector of the binary character generated in step 5 is:
[0028] ORB feature vector uses binary string to describe feature points and builds a Neighborhood, randomly selected Points, proceed Test, Generate To make the eigenvector rotationally invariant, The vector of points When a dimensional matrix is multiplied by a rotation matrix, the generated eigenvector is rotationally invariant.
[0029] Furthermore, the specific content of the nearest neighbor method matching feature points in step 6 is:
[0030] The ORB feature vector is a binary string composed of 0 and 1, which is suitable for XOR operation. The result is the Hamming distance. The more similar the feature vectors are, the smaller the Hamming distance is. Select a feature point in image 1 and find its nearest neighbor and second nearest neighbor in image 2. If the ratio of the Hamming distance between the nearest neighbor and the second nearest neighbor of the feature point is less than the threshold, the feature point and its nearest neighbor are matched feature points.
[0031] Furthermore, the specific content of step 7 using the local entropy difference, Euclidean distance, and slope of the matching point connection line to screen the matching points is:
[0032] The matching points are further screened by using the property that the local entropy difference, Euclidean distance, and slope of the matching point connection line of the correct matching point should be approximately equal. First, the local entropy difference of all matching points is calculated, and 20 groups of matching points with the smallest local entropy difference are selected. The Euclidean distance and the slope of the matching point connection line are calculated, and the average is taken as the standard value. If the local entropy difference of the remaining matching points is within the threshold, the difference between the Euclidean distance of this matching point, the slope of the matching point connection line, and the standard value is calculated. The matching point whose difference is within the threshold is the correct matching point.
[0033] Furthermore, the specific content of step 8 using the improved K-means algorithm to classify feature points is:
[0034] Without presetting the number of classifications, feature points are classified according to the actual clustering of feature points in the image. First, the neighborhood density of each feature point is calculated to determine the clustering of feature points in the neighborhood. Feature points with density greater than the threshold are selected as candidate initial class center points. The two candidate initial class center points with the largest density and the farthest distance from each other are selected as the initial class center points for clustering. Among the remaining unclassified candidate initial class center points, the one with the largest density is selected as the new initial class center point for clustering. The above process is repeated until all unclassified candidate initial class center points have been selected. Finally, the class with too few feature points is removed as an outlier set. The remaining classes are the correct feature point groupings.
[0035] Furthermore, the specific content of step 9 using the improved RANSAC algorithm to calculate the optimal transformation model is:
[0036] First, randomly select 4 categories from the N categories of feature points divided in step 8, randomly select 1 feature point from each category, and group the 4 feature points and the corresponding matching points into 1 group, and select M groups in total. The smaller the ratio of the Hamming distance between the nearest neighbor point and the second nearest neighbor point of the matching point, the higher the matching degree of the matching point. Arrange each group of matching points in ascending order according to the sum of their Hamming distance ratios, and select 1 group of matching points in turn to calculate the transformation model. Each calculated transformation model depends on 4 pairs of matching points, so there is a certain error between the transformation model and the inliers. Merge the inliers corresponding to all transformation models into the set Calculation After times, the transformation model with the most inliers is the candidate optimal transformation model. The matching points that meet the error distance threshold with the candidate optimal transformation model are classified as its internal points, and all the internal points of the candidate optimal transformation model are recalculated using the least squares method to obtain a more accurate transformation model.
[0037] Compared with the prior art, the present invention has the following beneficial effects:
[0038] In the feature point extraction stage, ORB feature points are extracted in SURF scale space, and the feature points are described using the feature vector of simple binary characters. The feature points are matched using the nearest neighbor method, and the matching points are screened using the local entropy difference, Euclidean distance, and slope of the matching point connection line. The classic K-means algorithm is improved, and the number of classifications is not preset. The feature points are classified more reasonably according to the actual aggregation of feature points in the image. Then, the efficient improved RANSAC algorithm is used to select matching points from different classes to calculate the transformation model between images and realize image registration. The algorithm has a moderate amount of computation, higher accuracy, and can adapt to places with high real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 A flowchart of an improved image registration algorithm based on SURF-ORB of the present invention;
[0040] Figure 2 is a test image according to an embodiment of the present invention;
[0041] Figure 3 This is a result diagram of SURF+RANSAC algorithm registration of a test image according to an embodiment of the present invention;
[0042] Figure 4 This is a result diagram of ORB+RANSAC algorithm registration of a test image according to an embodiment of the present invention;
[0043] Figure 5 This is a result diagram of the registration of the test image using the algorithm of the present invention in an embodiment of the present invention. DETAILED DESCRIPTION
[0044] The present invention is described in detail below with reference to the accompanying drawings and specific examples.
[0045] like Figure 1 As shown, an improved image registration algorithm based on SURF-ORB includes the following steps:
[0046] Step 1: Obtain two images to be registered, preprocess the images, and remove noise interference.
[0047] Step 2: Construct the SURF scale space.
[0048] The SURF algorithm uses the Hessian matrix to detect feature points. The scale is The Hessian matrix is:
[0049] H ( X , σ ) = [ L xx ( X , σ ) L xy ( X , σ ) L xy ( X , σ ) L yy ( X , σ ) ] (1)
[0050] To simplify the algorithm and improve the operation speed, the SURF algorithm uses a step-by-step box filter. , , replace , , Convolution of integral images creates a scale space.
[0051] Step 3: Extract ORB feature points.
[0052] At each layer of the SURF scale space, construct A circle with center 3 and radius 1. Compared with the 16 points on the circumference, if is the extreme point of at least 9 consecutive points, and then Compare the 26 points in the neighborhood. If it is still an extreme point, then is the ORB corner point.
[0053] To avoid too dense distribution of feature points, the ORB response value of each corner point is calculated in the 8-neighborhood centered on the ORB corner point, and the corner point with the largest ORB response value is retained as the feature point.
[0054] ORB response value:
[0055] (2)
[0056] in For point The gray value of For The grayscale value of any of the 16 points on the circle centered at .
[0057] Step 4: Determine the main direction of the feature points.
[0058] In the neighborhood centered on the feature point The neighborhood centroid is calculated, and the direction of the line connecting the feature point and the centroid is the main direction of the feature point. for size.
[0059] Neighborhood moments: (3)
[0060] Centroid: (4)
[0061] Angle of the main direction of the feature point: (5)
[0062] Step 5: Generate a feature vector of binary characters to describe the feature points.
[0063] ORB feature vector uses binary string to describe feature points, centered on the feature point Randomly select Points, proceed Test, Generate To make the eigenvector rotationally invariant, The vector of points When a dimensional matrix is multiplied by a rotation matrix, the generated eigenvector is rotationally invariant.
[0064] The specific steps to generate a binary character feature vector to describe the feature points are:
[0065] Step 5.1: Feature points The image block P is Test, the present invention is 31.
[0066] (6)
[0067] In the formula , are the grayscale values of points X and Y in the image block P respectively.
[0068] Step 5.2: Feature descriptors are the feature points and image blocks P One obtained by comparing Bit vector.
[0069] (7)
[0070] In the formula .
[0071] Step 5.3: Let this The vector of points The eigenvectors generated by multiplying the eigenvector matrix by the rotation matrix containing the main directions of the feature points are rotationally invariant.
[0072] Using a vector of n points Define a matrix:
[0073] (8)
[0074] The rotation matrix is defined as:
[0075] (9)
[0076] In the formula is the angle of the main direction of the feature point.
[0077] (10)
[0078] Feature descriptors that are rotationally invariant:
[0079] (11)
[0080] Step 6: Match feature points using the nearest neighbor method.
[0081] The ORB feature vector is a binary string composed of 0 and 1, which is suitable for XOR operation. The result is the Hamming distance. The more similar the feature vectors are, the smaller the Hamming distance is. For images Any feature point The characteristic vector of , for In the image Nearest neighbor Next nearest neighbor The feature vector of .
[0082] (12)
[0083] (13)
[0084] like If it is less than the threshold, The ratio threshold in the present invention is 0.65.
[0085] Step 7: Use the local entropy difference, Euclidean distance, and slope of the line connecting the matching points to select the matching points.
[0086] After the feature points are initially matched, there are still some mismatched points, which need to be further screened to improve the accuracy of the algorithm. The local entropy is centered at any point in the image and the neighborhood is The entropy of the image block reflects the global information of the image block. The greater the amount of information contained in the image block, the more prominent the details, and the greater the entropy value. If two feature points are correct matching points, the information content in the neighborhood corresponding to the two feature points should be approximately the same, and the local entropy should be approximately equal.
[0087] The matching points are further screened by using the property that the local entropy difference, Euclidean distance, and slope of the matching point connection line of the correct matching point should be approximately equal. First, the local entropy difference of all matching points is calculated, and 20 groups of matching points with the smallest local entropy difference are selected. The Euclidean distance and the slope of the matching point connection line are calculated, and the average is taken as the standard value. If the local entropy difference of the remaining matching points is within the threshold, the difference between the Euclidean distance of this matching point, the slope of the matching point connection line, and the standard value is calculated. The matching point whose difference is within the threshold is the correct matching point.
[0088] In the characteristic point is the center and its neighborhood is On the image block, the local entropy of this image block is calculated using formula (14).
[0089] (14)
[0090] (15)
[0091] In the formula For point The gray value of .
[0092] set up is a pair of matching points. Formulas (16), (17), and (18) are the calculation formulas for the local entropy difference, Euclidean distance, and slope of the line connecting the matching points.
[0093] (16)
[0094] (17)
[0095] (18)
[0096] The specific steps of screening matching points using the local entropy difference, Euclidean distance, and slope of the matching point connection are as follows:
[0097] Step 7.1: Calculate the local entropy difference of all matching points using formula (16): In the present invention is 5.
[0098] Step 7.2: Take 20 pairs of matching points with the smallest local entropy difference, and use formulas (17) and (18) to calculate the Euclidean distance and the slope of the line connecting the matching points. , Slope of the line connecting the matching points The mean is the standard value .
[0099] Step 7.3: If the remaining matching points The local entropy difference is less than the threshold, of and With standard value and If the difference is less than the threshold, then In the present invention, the local entropy difference threshold is 0.1, the Euclidean distance difference threshold is 0.2, and the matching point connection line slope difference threshold is 0.1.
[0100] Step 8: Use the improved K-means algorithm to classify feature points.
[0101] The present invention does not preset the number of classifications, and classifies feature points according to the actual aggregation of feature points in the image. First, the neighborhood density of each feature point is calculated to determine the aggregation of feature points in the neighborhood. Feature points with a density greater than a threshold are selected as candidate initial class center points. The two candidate initial class center points with the largest density and the farthest distance from each other are selected as the initial class center points for clustering. Among the remaining unclassified candidate initial class center points, the one with the largest density is selected as the new initial class center point for clustering, and the above process is repeated until all unclassified candidate initial class center points have been selected. Finally, the class with too few feature points is removed as an outlier set. The remaining classes are the correct feature point groupings.
[0102] Specific steps for classifying feature points using the improved K-means algorithm:
[0103] Step 8.1: Use formula (19) to calculate the neighborhood density of each feature point and determine the clustering of feature points in the neighborhood.
[0104] (19)
[0105] In the formula For feature points The number of feature points in the circular neighborhood centered at .
[0106] Formula (19) is The reciprocal of the Euclidean distance to the feature points in the neighborhood, representing The degree of closeness with surrounding points. The greater the density of feature points, the more compact the points in the neighborhood are, and the more suitable they are for being the center point of the class.
[0107] Step 8.2: Select feature points with density greater than a threshold as candidate initial class center points, and arrange the candidate initial class center points in descending order according to density values. In the present invention, the density threshold is 50.
[0108] Step 8.3: Select the first point in the sequence ,distance The point with the farthest Euclidean distance is , the feature points in the image are divided into Corresponding class middle.
[0109] Step 8.4: Use formula (20) to recalculate the center points of each category.
[0110] (20)
[0111] Step 8.5: Among the remaining unclassified candidate initial class center points, select the candidate initial class center point with the largest density as the new class center point . Classify the feature points in the image into the nearest class center point Corresponding class middle.
[0112] Step 8.6: Repeat steps 8.3-8.4 until all unclassified candidate initial class center points have been selected.
[0113] Step 8.7: Eliminate the class whose number of feature points is lower than the threshold as an outlier point set. The feature points are divided into N classes. The number threshold in the present invention is 20.
[0114] Image Transformation Model
[0115] The translation, rotation and deformation between two images can be described by affine transformation.
[0116] (twenty two)
[0117] In the formula , for , The corresponding matching points.
[0118] (twenty three)
[0119] Four pairs of matching points are enough to solve the transformation model , match the overlapping areas of the two images.
[0120] Step 9: Calculate the transformation model using the improved RANSAC algorithm.
[0121] First, randomly select 4 categories from the N categories of feature points divided in step 8, randomly select 1 feature point from each category, and group the 4 feature points and the corresponding matching points into 1 group, and extract M groups in total. The smaller the ratio of the Hamming distance between the nearest neighbor and the second nearest neighbor of the matching point, the higher the matching degree of the matching point. Arrange each group of matching points in ascending order according to the sum of their Hamming distance ratios, and select 1 group of matching points in turn to calculate the transformation model. Each calculated transformation model depends on 4 pairs of matching points, so there is a certain error between the transformation model and the inliers. Merge the inliers corresponding to all transformation models into the set Calculation After times, the transformation model with the most inliers is the candidate optimal transformation model. The matching points that meet the error distance threshold with the candidate optimal transformation model are classified as its internal points, and all the internal points of the candidate optimal transformation model are recalculated using the least squares method to obtain a more accurate transformation model.
[0122] Specific steps for calculating the transformation model using the improved RANSAC algorithm:
[0123] Step 9.1: Step 8 divides the feature points into N categories, randomly extracts 4 categories from the N categories each time, randomly extracts 1 feature point from each category, and groups the 4 feature points and the corresponding matching points into 1 group, extracting a total of M groups. .
[0124] Step 9.2: Arrange the M groups of matching points in ascending order according to the sum of the Hamming distance ratios between the nearest neighbor and the second nearest neighbor, and set the initial value of the algorithm: , .
[0125] Step 9.3: Select a group of matching points in turn and calculate the transformation model using formula (23) . Determine the matching point as an internal point and count The corresponding number of interior points , all interior points are merged into middle.
[0126] Step 9.4: ,like , return to step 9.3.
[0127] Step 9.5: The one with the most inliers is the candidate optimal transformation model, judging Whether the matching points in , will meet the conditions Classify In the present invention, is 1.25.
[0128] Step 9.6: Use the least squares method, using All interior points are re-estimated and the transformed model is obtained as the final solution.
[0129] Step 10: Image Registration
[0130] The transformation model is used to calculate the precise position of the matching points in the overlapping area of the two images, so that they correspond one to one, and the overlapping parts are aligned to synthesize a large scene image to complete the image registration.
[0131] Example, Figure 2 For the test image, Figure 3 This is the result of SURF+RANSAC algorithm registration. Figure 4 This is the result of ORB+RANSAC algorithm registration. Figure 5 This is the result of the registration using the algorithm of the present invention.
[0132] Table 1
[0133] Algorithm Name Match Points / Pair Matching accuracy / % Running time / s SURF+RANSAC 73 78.5 9.525 ORB+RANSAC 68 55.2 7.157 Algorithm of the present invention 55 86.8 8.098
[0134] Table 1 shows the experimental results of the registration of the test images by the SURF+RANSAC algorithm, the ORB+RANSAC algorithm and the algorithm of the present invention.
[0135] It can be seen from Table 1 that the algorithm of the present invention reduces the number of matching points by 24.65% compared with SURF+RANSAC, improves the matching accuracy by 8.3%, and reduces the operation time by 1.427s. The algorithm of the present invention reduces the number of matching points by 19.11% compared with ORB+RANSAC, improves the matching accuracy by 31.6%, and increases the operation time by 0.941s. Comprehensive analysis shows that the algorithm of the present invention consumes slightly more time than the ORB+RANSAC algorithm, and the registration accuracy is greatly improved. Compared with the SURF+RANSAC algorithm, the operation time and registration accuracy have great advantages, and the overall efficiency of the algorithm is high.
Claims
1. An improved image registration algorithm based on SURF-ORB, characterized in that: The following steps are involved: Step 1: Obtain two images to be registered, preprocess the images, and remove noise interference; Step 2: Construct SURF scale space; Step 3: Extract ORB feature points; Step 4: Determine the main direction of the feature point; Step 5, generate a feature vector of a binary character to describe the feature point; Step 6: Match feature points using the nearest neighbor method; Step 7, using the local entropy difference, Euclidean distance, and slope of the line connecting the matching points to select the matching points; Step 8: Use the improved K-means algorithm to classify the feature points and classify the feature points in the same area into one category; Step 9: Calculate the transformation model using the improved RANSAC algorithm; Step 10: Image registration.
2. The improved image registration algorithm based on SURF-ORB according to claim 1, characterized in that: In step 2, the SURF scale space is constructed. The specific contents are as follows: The SURF algorithm uses the Hessian matrix to detect feature points. The scale is The Hessian matrix is: (1) To simplify the algorithm and improve the operation speed, the SURF algorithm uses a step-by-step box filter. , , replace , , Convolution of integral images creates a scale space.
3. The improved image registration algorithm based on SURF-ORB according to claim 1, characterized in that: In step 3, ORB feature points are extracted. The specific contents are as follows: At each layer of the SURF scale space, construct A circle with center 3 and radius 1. Compared with the 16 points on the circumference, if is the extreme point of at least 9 consecutive points, and then Compare the 26 points in the neighborhood. If it is still an extreme point, then is the ORB corner point; To avoid too dense distribution of feature points, the ORB response value of each corner point is calculated in the 8-neighborhood centered on the ORB corner point, and the corner point with the largest ORB response value is retained as the feature point.
4. The improved image registration algorithm based on SURF-ORB according to claim 1, characterized in that: In step 4, the main direction of the feature point is determined, and the specific content is: Construct a neighborhood centered on the feature point, calculate the neighborhood centroid, and the direction of the line connecting the feature point and the centroid is the main direction of the feature point; Neighborhood moments: (2) Centroid: (3) Angle of the main direction of the feature point: (4).
5. The improved image registration algorithm based on SURF-ORB according to claim 1, characterized in that: In step 5, the feature vector of the binary character is generated, and the specific content is: ORB feature vector uses binary string to describe feature points and builds a Neighborhood, randomly selected Points, proceed Test, form To make the eigenvector rotationally invariant, When the matrix composed of the vectors of the points is multiplied by the rotation matrix, the generated eigenvector is rotationally invariant.
6. The improved image registration algorithm based on SURF-ORB according to claim 1, characterized in that: In step 6, the nearest neighbor method matches feature points. The specific content is: The ORB feature vector is a binary string composed of 0 and 1, which is suitable for XOR operation. The result is the Hamming distance. The more similar the feature vectors are, the smaller the Hamming distance is. Select a feature point in image 1 and find its nearest neighbor and second nearest neighbor in image 2. If the ratio of the Hamming distance between the nearest neighbor and the second nearest neighbor of the feature point is less than the threshold, the feature point and its nearest neighbor are matched feature points.
7. The improved image registration algorithm based on SURF-ORB according to claim 1, characterized in that: In step 7, the local entropy difference, Euclidean distance, and slope of the line connecting the matching points are used to select the matching points. The specific contents are as follows: The matching points are further screened by using the property that the local entropy difference, Euclidean distance, and slope of the matching point connection line of the correct matching point should be approximately equal. First, the local entropy difference of all matching points is calculated, and 20 groups of matching points with the smallest local entropy difference are selected. The Euclidean distance and the slope of the matching point connection line are calculated, and the average is taken as the standard value. If the local entropy difference of the remaining matching points is within the threshold, the difference between the Euclidean distance of this matching point, the slope of the matching point connection line, and the standard value is calculated. The matching point whose difference is within the threshold is the correct matching point.
8. The improved image registration algorithm based on SURF-ORB according to claim 1, characterized in that: The improved K-means algorithm in step 8 classifies the feature points. The specific contents are as follows: Without presetting the number of classifications, feature points are classified according to the actual clustering of feature points in the image. First, the neighborhood density of each feature point is calculated to determine the clustering of feature points in the neighborhood. Feature points with density greater than the threshold are selected as candidate initial class center points. The two candidate initial class center points with the largest density and the farthest distance from each other are selected as the initial class center points for clustering. Among the remaining unclassified candidate initial class center points, the one with the largest density is selected as the new initial class center point for clustering. The above process is repeated until all unclassified candidate initial class center points have been selected. Finally, the class with too few feature points is removed as an outlier set. The remaining class is the correct feature point classification.
9. The improved image registration algorithm based on SURF-ORB according to claim 1, characterized in that: The improved RANSAC algorithm in step 9 calculates the transformation model, and the specific content is: First, randomly select 4 categories from the N categories of feature points divided in step 8, randomly select 1 feature point from each category, and group the 4 feature points and the corresponding matching points into 1 group, and extract M groups in total. The smaller the ratio of the Hamming distance between the nearest neighbor and the second nearest neighbor of the matching point, the higher the matching degree of the matching point. Arrange each group of matching points in ascending order according to the sum of their Hamming distance ratios, and select 1 group of matching points in turn to calculate the transformation model. Each calculated transformation model depends on 4 pairs of matching points, so there is a certain error between the transformation model and the inliers. Merge the inliers corresponding to all transformation models into the set Calculation After times, the transformation model with the most inliers is the candidate optimal transformation model. The matching points that meet the error distance threshold with the candidate optimal transformation model are classified as its internal points, and all the internal points of the candidate optimal transformation model are recalculated using the least squares method to obtain a more accurate solution.