Image line feature matching method considering significant visual angle difference of image

Abstract

This invention relates to an image line feature matching method that takes into account significant viewpoint differences in images, belonging to the field of image matching technology. First, the invention fuses the pixel-level directional gradient histogram features and image grayscale information features of the image, and forms a feature description grid based on the fused features, increasing the information content and dimensionality of the image features. Then, line segments in the image are divided into multiple discrete points, and the discrete point descriptors on the lines are aggregated to overcome the requirements of fixed line segment lengths and consistency of line segment support domain content. Finally, line feature matching is performed using the constructed descriptors on both sides of the line features and the feature description grid. During matching, the matching results are verified through geometric constraints of corresponding points, effectively solving the problems of line segment length differences across viewpoints and differences in image content within the single-sided support domain of corresponding lines, thus improving the matching accuracy.

Description

Technical Field

[0001] This invention relates to an image line feature matching method that takes into account significant differences in image perspective, and belongs to the field of image matching technology. Background Technology

[0002] In recent years, with the continuous development of Earth observation technology and the Internet, photogrammetry has evolved from a single-platform approach to a multi-platform, cross-view collaborative processing approach encompassing air, space, ground, and network. Finding corresponding features between cross-view images to accurately acquire geospatial information has become an important image processing method. Among the features commonly used for image matching, point features have poor globality and are prone to losing scene structural information; area feature matching is computationally complex and inefficient; line features fall between the two, containing a suitable amount of stable mid-level information (contour and semantic information), and can fully reflect the spatial characteristics of the target. Utilizing line features for matching can enhance the expression of target structural relationships in cross-view images, providing a new approach for multi-platform, cross-view image processing.

[0003] Based on the different matching primitives, existing image line feature matching methods can be divided into three categories: line group-based matching, point-line feature-based matching, and single-segment matching. Line group-based matching methods require grouping two or more lines according to certain rules and using affine invariant features within the group to build descriptors for overall matching. For example, "line-intersection-line" (LJL) structures and V-shaped structure line groups can be constructed respectively. Connecting the line feature intersections and intensity maxima points forms local affine invariant regions, and SIFT descriptions are calculated in these local regions to complete the matching. Alternatively, V-shaped line groups with only one single line segment in the neighborhood can be considered unstructured line groups, and the spatial structure consistency of line groups with the same name can be used to obtain correct matches. Line feature matching methods based on point-line features construct geometric invariants using the positional relationships between points and lines within the support domain. Line feature matching pairs are then deduced from the matching results between point features and these geometric invariants. For example, descriptor performance can be enhanced by constructing affine invariants of "one line + two points" and projection invariants of "one line + four points." Alternatively, coplanar line-point projection invariants can be constructed based on the number of features, reducing the descriptor's dependence on texture information and demonstrating good performance against grayscale differences in the neighborhood of line features caused by viewpoint changes. However, the construction process of invariants places certain requirements on the accuracy of feature points. Single-line-segment matching methods mainly achieve matching by solving for the overlapping portion of straight lines in the reference image and the image to be matched. For images with viewpoint changes, prior knowledge or numerous constraints are usually required to compensate for the instability of single-line-segment information. For example, establishing transformation relationships between images using POS data and obtaining corresponding line pairs using the transformed images eliminates geometric deformation between cross-view images, resulting in a high matching accuracy. However, this method is not applicable to images without prior information.

[0004] Furthermore, with the rapid development of deep learning, Convolutional Neural Networks (CNNs), with their powerful deep feature extraction capabilities, have been widely applied in the field of feature matching. SOLD2 is the first network to utilize deep learning for line feature detection and description, expanding image line feature matching methods and achieving good matching results on images with viewpoint changes. However, most of the aforementioned line feature matching methods rely on highly repetitive common features within the support domain on both sides of the line segment. These features are greatly affected by the line segment length and image geometric deformation, making it difficult to reflect the common attributes of corresponding line features across viewpoints, thus affecting the final matching accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide an image line feature matching method that takes into account significant differences in image perspective, so as to solve the problem of low matching accuracy in existing image line feature matching methods.

[0006] To address the aforementioned technical problems, this invention provides an image line feature matching method that takes into account significant differences in image perspective. This matching method includes the following steps:

[0007] 1) Obtain the pixel-level histogram of oriented gradients (HAR) features of the image pair to be matched, and fuse the obtained pixel-level HAR features with the grayscale features of the image to form a feature description grid. The feature value of each grid is the fused feature. The image pair to be matched includes a reference image and an image to be registered.

[0008] 2) Divide the line segments in the image pair to be matched to obtain the corresponding discrete point set. Calculate the descriptors on both sides of each discrete point according to the feature description grid. Aggregate the descriptors on both sides of each discrete point on the line segment to obtain the line feature descriptors on both sides.

[0009] 3) Perform line feature matching based on the descriptors on both sides of the obtained line feature.

[0010] This invention first fuses the pixel-level directional gradient histogram features and image grayscale information features of an image, and then forms a feature description grid to describe the image based on the fused features. This not only increases the information content of the image features but also increases the expression of dimensional features. Next, line segments in the image are divided into multiple discrete points. By aggregating the descriptors of discrete points on the lines, the requirement for fixed length of line segments is overcome. At the same time, left and right regions are divided by lines to form left and right side descriptions of discrete points on the lines, overcoming the requirement for consistency of the content of the support domain of line segments. Finally, line feature matching is performed using the constructed descriptors on both sides of the line features and the feature description grid, achieving robust matching of cross-view images and improving matching accuracy.

[0011] Furthermore, in step 3), when performing line feature matching, known corresponding points in the image pair to be matched are first grouped for matching to obtain corresponding candidate line feature matching pairs. The candidate line feature matching pairs are then determined based on the descriptors on both sides of the obtained line features. The process for determining candidate line feature matching pairs is as follows:

[0012] Circular search windows with radius R are set in the reference image and the image to be registered, respectively, with the corresponding points as the center. Line features located within the window are grouped together for matching.

[0013] In the reference image and the image to be registered, the line features of the same group of corresponding points are obtained respectively. The distance between the descriptors on both sides of the line features in the group is calculated to obtain the distance matrix in each group.

[0014] Select the line pairs corresponding to the minimum values ​​in the row and column of each matrix as candidate feature matching pairs.

[0015] This invention solves the problem of high computational load and many mismatches in the matching process caused by the large number and complex changes of line features by using the information of corresponding points to narrow the search range of line features, thereby further improving matching efficiency and matching accuracy.

[0016] Furthermore, the method also includes filtering the one-to-many and many-to-many matching results to select the line feature matching pairs with the smallest distance.

[0017] This invention sorts non-one-to-one matching results according to descriptor distance, retains the line feature matching pairs with the smallest distance, effectively removes redundant matching, and further improves matching efficiency and accuracy.

[0018] Furthermore, it also includes verifying the one-to-one matching results. The verification process is as follows:

[0019] Search for multiple pairs of corresponding points in the baseline image that are closest to the corresponding line segment in the line feature matching pair to be verified. Construct a positive topological descriptor based on the distribution relationship of each corresponding point in each pair on the left and right sides of the corresponding line segment in the line feature matching pair to be verified.

[0020] Search for multiple pairs of corresponding points in the image to be registered that are closest to the corresponding line segment in the line feature matching pair to be verified. Construct a reverse topological descriptor based on the distribution relationship of each corresponding point in each pair on the left and right sides of the corresponding line segment in the line feature matching pair to be verified.

[0021] The similarity of the line feature matching pairs to be verified is determined based on the forward and reverse topological descriptors. If the similarity is not less than the set threshold, the line feature matching pairs to be verified meet the verification and are considered successfully matched. Otherwise, the verification is not met and they are considered unsuccessfully matched.

[0022] This invention utilizes corresponding points to construct forward and reverse topological descriptors, thereby calculating the similarity of line feature matching pairs to be verified, and verifying the matching results based on the similarity, thus improving the accuracy of line feature matching.

[0023] Furthermore, the threshold value is set to 0.8n, where n is the number of identical point pairs searched.

[0024] Furthermore, the formula for calculating the similarity is as follows:

[0025]

[0026] avgsim{l i ,l′ k} represents the line feature matching pair {l} to be verified. i ,l′ k The similarity of} (Γ ik ,Γ′ ik ) is a positive topological descriptor, (OΓ ik ,OΓ′ ik ) is the reverse topology descriptor, and n is the number of identical point pairs to search.

[0027] Furthermore, the characteristic is that, in step 1), fusion refers to performing a convolution operation after numerically superimposing the grayscale and single-pixel directional gradient histogram features pixel by pixel to obtain an image feature grid.

[0028] This invention comprehensively considers the correlation between spatial structure information and grayscale information, and fully integrates pixel-level HOG features and grayscale features. On the one hand, it performs pixel-by-pixel numerical superposition operations on grayscale and single-pixel directional gradient histogram features to generate new feature maps, increasing the amount of image feature information and improving the ability to express fine structures; on the other hand, through convolution operations, it explores the deep-level correlation between the two types of features, further improving the matching accuracy.

[0029] Furthermore, the convolution operation uses a superpoint network.

[0030] Furthermore, the process of determining the descriptors of discrete points in step 2) is as follows:

[0031] A. Divide the line segments in the image pair to be matched according to a set interval to obtain a discrete set of points to describe the line. Use the endpoint with the smaller x-coordinate as the starting point of the line segment to determine the direction of the line segment.

[0032] B. Determine the grid point on the line segment that is closest to the discrete point in the grid where the discrete point is located as the first grid point, and determine whether the first grid point is on the left or right side of the line segment along the direction of the line segment;

[0033] C. Map the obtained first grid point to the opposite side of the line segment where the discrete point is located to obtain the opposite mapped point, and assign the interpolated feature value of the opposite mapped point to the first grid point as the feature value of the first grid point;

[0034] D. Calculate the descriptor of the discrete point based on the difference between the horizontal and vertical coordinate axes of the discrete point and the grid point, as well as the characteristic values ​​of each grid point in the grid. The first grid point uses the assigned characteristic value.

[0035] Furthermore, the formula for calculating the descriptor in step D is as follows:

[0036]

[0037] For discrete points The left-hand descriptor; Δx and Δy are discrete points. The difference between the x and y coordinates of the point in the top-left grid of the same grid; v 00 v 01 v 10 and v 11 Discrete points The characteristic values ​​of the top-left, top-right, bottom-left, and bottom-right grid points of the given grid.

[0038] This invention divides line segments into discrete points and overcomes the requirement for fixed line segment length by aggregating discrete point descriptors on the line. At the same time, it divides left and right regions with lines to form left and right side descriptions of discrete points on the line, overcoming the requirement for consistency of line segment support domain content and further improving matching accuracy. Attached Figure Description

[0039] Figure 1 This is a flowchart of the image line feature matching method of the present invention that takes into account significant differences in image perspective;

[0040] Figure 2 This is a schematic diagram of the grid point mapping used in this invention;

[0041] Figure 3a This is a schematic diagram of the circular search window for the reference image in an embodiment of the present invention;

[0042] Figure 3b This is a schematic diagram of the circular search window for the image to be matched in an embodiment of the present invention;

[0043] Figure 3c This is a schematic diagram illustrating the principle of line feature grouping and matching in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the topology description verification principle used in this invention. Detailed Implementation

[0045] For cross-view imagery, this invention proposes a line feature matching method that takes into account the unilateral neighborhood of line segments and length variations. By constructing line descriptors with the equidistant point information contained in the line features, line features that meet the geometric constraints of corresponding points are grouped for matching. By comparing the topological consistency of corresponding line pairs, erroneous matches are eliminated, effectively solving the problems of line segment length differences in cross-view imagery and differences in image content within the unilateral support domain of corresponding lines.

[0046] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0047] This invention first obtains the pixel-level histogram of oriented gradients (HARQs) features of the image pairs to be matched. The obtained HARQ features are then fused with the grayscale features of the images to form a feature description grid. Next, the line segments in the image pairs to be matched are divided at equal intervals to obtain corresponding discrete point sets. Descriptors on both sides of each discrete point are calculated based on the feature description grid. The descriptors on both sides of each discrete point on the line segment are aggregated to obtain the descriptors on both sides of the line feature. Finally, the matching regions are constrained using known corresponding points for group matching. The matching results are verified by comparing the topological consistency of the line pairs, achieving robust matching of cross-view images. The implementation process of this method is as follows: Figure 1 As shown below, the specific implementation process of this procedure will be explained in detail.

[0048] 1. The obtained pixel-level oriented gradient histogram features are fused with the grayscale features of the image to form a corresponding feature grid.

[0049] The Histogram of Oriented Gradient (HOG) feature is a descriptor that statistically analyzes gradient information within local grid cells of an image. This descriptor needs to be determined based on the magnitude and direction of the gradient. The magnitude and direction of the gradient at pixel (x, y) can be calculated based on the grayscale values ​​of neighboring pixels. The specific calculation formula is as follows:

[0050]

[0051]

[0052] The gradient direction [0°, 180°] is divided into 9 directional intervals. Using a sliding window approach, 2×2 units of size 4×4 are grouped into blocks. The image directional gradient histogram is calculated on a block-by-block basis and normalized using the L2 norm, resulting in a 2×2×9 = 36-dimensional feature vector. Based on this, the stride of the sliding window is reduced to a single pixel, which can form a 36-dimensional pixel-level HOG feature description, effectively enhancing the ability to express edge structure information.

[0053] The calculation of gradient histograms neglects some detailed features, resulting in poor performance when dealing with minute edges. Therefore, this invention comprehensively considers the correlation between spatial structure information and grayscale information, fully fusing pixel-level HOG features and grayscale features. Feature fusion includes parallel and serial methods. The former increases the amount of image information without changing the dimensionality, while the latter increases the feature dimensions of the image while preserving information within each dimension. This invention addresses these two aspects by optimizing the combination of different feature tensors of the same image. The specific process is as follows:

[0054] (1) Feature map information integration.

[0055] To fully utilize the complementarity of different feature sets, it is necessary to superimpose and integrate the feature maps at the information level, that is, to fuse the two types of features sequentially. During fusion, the original image dimensions are kept unchanged, and the grayscale and single-pixel directional gradient histogram features are numerically superimposed pixel by pixel to generate a new feature map, increasing the amount of image feature information and improving the ability to express fine structures.

[0056] (2) Feature map dimension reconstruction.

[0057] The two sets of features, after simple superposition, lack information exchange. In order to further explore the deep correlation between different features, this invention uses a superpoint network to perform convolution operation on the integrated feature map to obtain a 256-dimensional image feature grid of size 64×64, so that the new features have advanced semantic recognition capabilities at the dimensional level and are more discriminative.

[0058] 2. Construct descriptors on both sides of the line feature.

[0059] Traditional line feature description methods typically establish a support domain based on the line and construct descriptors using the point-line structure or grayscale information within the support domain. In cross-view images, line features are prone to occlusion and breakage, leading to differences in the support domain range of the same line; while changes in viewpoint cause differences in image content within the support domain on one side of the same line.

[0060] To address this issue, this invention draws upon the method of segmenting sentences using processors of the same length to effectively solve the problem of inconsistent sentence lengths in natural language processing. This invention divides line segments in an image into discrete points and overcomes the requirement for fixed line segment lengths by aggregating descriptors of these discrete points. Simultaneously, it divides left and right regions along the lines, forming left and right side descriptions of the discrete points on the lines respectively, overcoming the requirement for consistency in the content of the line segment's support domain. The specific steps are as follows:

[0061] (1) Establish a discrete point set.

[0062] Line segments in the image are divided into discrete points. During the division, the line features are segmented at equal intervals according to a set time, uniformly obtaining effective discrete points to describe the line, reducing information redundancy while ensuring the saliency of the line features. To obtain an ordered set of discrete points, the direction of the line features needs to be determined in advance. This invention regards the endpoint with the smaller x-coordinate value as the starting point of the line segment, and the direction from the starting point to the other endpoint of the line segment as the line direction. With the starting point as the origin, the line segment is divided according to a fixed step size s to obtain a uniform set of discrete points.

[0063] (2) Mark the grid points.

[0064] In three-dimensional geometry, the relative relationships between two vectors can be inferred from the Z-direction of the normal vector obtained by the cross product of the two vectors. Assume... They are line segments l i The starting point and the ending point, Indicate l i The kth discrete point, G k =(x G ,y G )express The grid in which it is located and with The nearest grid point, defined Let A be a three-dimensional vector pointing from point A to point B. Then the vector... and The magnitude of the projection of the normal vector onto the Z-axis is:

[0065]

[0066] like Decision vector In vector To the left of, and then determine the grid point G. k On segment l i On the left, marked as Otherwise, record as

[0067] (3) Construct the left and right descriptors of the point.

[0068] When calculating the one-sided descriptor of discrete points on the line (taking the left side as an example), the grid points marked as R in the feature grid are... Map to the opposite side of the line, and map the left side to the point. interpolated eigenvalues Assigned (like Figure 2 As shown), the left descriptor of the line points in each layer of the feature grid is calculated according to formula (4), and finally a 256-dimensional left point descriptor is formed; the right descriptor is obtained in the same way.

[0069]

[0070] For discrete points The left-hand descriptor; Δx and Δy are discrete points. The difference between the x and y coordinates of the point in the top-left grid of the same grid; v 00 v 01 v 10 and v 11 Discrete points The eigenvalues ​​of the top-left, top-right, bottom-left, and bottom-right grid points of the given grid. Figure 2 In the example

[0071] When calculating the right-side descriptor, the grid point marked as L is mapped to the right side, and the interpolated feature value of the right-side mapped point is assigned to the left-side grid point. Then, the feature value of the assigned left-side grid point is used to calculate according to formula (4), where Δx and Δy in the formula are the discrete points. The difference between the horizontal and vertical coordinates of the point in the lower right grid of the grid.

[0072] (4) Aggregation point descriptor.

[0073] This invention extends the number of encoding layers on the Transformer model, using the extended Transformer model as the point feature aggregation network employed in this invention. This network utilizes a 12-layer encoder-decoder to extract the deep characteristics of multiple sets of feature vectors. Each layer consists of two sub-layers: a multi-head self-attention mechanism (MSA) and a multi-layer perceptron mechanism (MLP). By adding residual links and normalization layers between layers, the effective information utilization rate is improved and noise interference is reduced. The feature aggregation method adopted in this invention can comprehensively consider global information to obtain the optimal integrated vector of multiple features. Therefore, this invention uses the left and right point descriptors of pixel-level HOG features as the basic input of this network, aggregating them to obtain the left and right descriptors Des of line features. L and Des R This gives it advanced semantic information.

[0074] 3. Perform group matching based on the same point constraint, and verify the matching results.

[0075] Line features, as fundamental features, are numerous and complex in images, leading to a large computational load in the matching process and a tendency to produce false matches. To address this, this invention combines corresponding points to narrow the search range of image line features and implements group matching and verification to effectively improve matching efficiency and accuracy.

[0076] 1) Line feature grouping matching.

[0077] Circular search windows with radius R are set in the reference image and the image to be matched, respectively, centered on the corresponding points m and m′. Figure 3a and Figure 3b As shown, using the distance between points and lines as a metric, line features within a window are grouped together. For line features in the same group, the distance between the left-side line descriptors of the two images is calculated according to formula (5), forming an LL distance matrix (e.g., ...). Figure 3c The matrix represents the line features of the reference image and the image to be matched, respectively. The descriptor distance of the line pair (l2, l′1) is the minimum value of its row and column; therefore, (l2, l′1) is considered a candidate matching line pair. Similarly, the distances of the one-sided descriptors of the same group of line features are calculated in the manner of "left-right, right-left, right-right," forming RL, LR, and RR distance matrices to filter candidate matching pairs. "Left-left" refers to the left descriptor of the reference image line feature minus the left descriptor of the image to be matched; "left-right" refers to the left descriptor of the reference image line feature minus the right descriptor of the image to be matched; "right-left" refers to the right descriptor of the reference image line feature minus the left descriptor of the image to be matched; and "right-right" refers to the right descriptor of the reference image line feature minus the right descriptor of the image to be matched.

[0078]

[0079] In reality, the distribution of corresponding point pairs in cross-view images is not uniform and may be concentrated in a small region with significant texture. This makes it difficult to match line features far from corresponding points. To solve this problem, after group matching, the unmatched line features are treated as a group and residual matching is performed to supplement the group matching results.

[0080] (2) Verification of matching results.

[0081] The neighborhood overlap of points with similar distances is large, and the same line segment may belong to multiple groups, resulting in "one-to-many" and "many-to-many" matching results. To address this, the present invention sorts non-"one-to-one" matching results according to descriptor distance and retains only the line feature matching pairs with the smallest distance to effectively remove redundant matching.

[0082] To enhance the accuracy of "one-to-one" matching, this invention filters and judges the distribution of corresponding points on the left and right sides of the line pair, constructs a topological description for each matching line pair, and calculates descriptor similarity to verify the matching results. For example... Figure 4 As shown, the matching pairs to be verified (l) i -l′ k For example, first, search for l in the reference image. i The nearest n (n=10) pairs of points with the same name {(m1-m′1),(m2-m′2),…,(m i10 -m′ i10 Then determine {m1, m2, ..., m} respectively. i10} in l iand {m′1,m′2,…,m′ i10} at l′ k The distribution relationship between the left and right sides: if a point is located on the left side of the line, the corresponding element in the topological description is denoted as L, and the one on the right side is denoted as R, forming the positive topological description Γ of the line pair. ik and Γ′ ik .

[0083] Because of matching errors in corresponding point pairs, and because the nearest neighbor points of corresponding line features differ across images, a reverse topological descriptor is added to reduce the impact of mismatched points on the verification results. This involves selecting l′ from the images to be registered. k The nearest n pairs of points with the same name {(m1-m′1),(m2-m′2),…,(m k10 -m′ k10 )}, determine the relationship between points and lines, and form a reverse topological descriptor OΓ ik and OΓ′ ik Calculate the similarity between the positive and negative descriptors of the line pairs to be verified according to formula (6), and retain line pairs with a similarity of not less than 0.8n.

[0084]

[0085] in:

[0086] Through the above process, for cross-view images, this invention constructs line descriptors by utilizing the equidistant point information contained in the line features, groups and matches line features that satisfy the geometric constraints of corresponding points, and eliminates erroneous matches by comparing the topological consistency of corresponding line pairs. This effectively solves the problems of line segment length differences in cross-view images and differences in image content within the single-side support domain of corresponding lines, and can achieve robust matching of cross-view images.

Claims

1. An image line feature matching method that takes into account significant differences in image perspective, characterized in that, The matching method includes the following steps: 1) Obtain the pixel-level histogram of oriented gradients (HAR) features of the image pair to be matched, and fuse the obtained pixel-level HAR features with the grayscale features of the image to form a feature description grid. The feature value of each grid is the fused feature. The image pair to be matched includes a reference image and an image to be registered. 2) Divide the line segments in the image pair to be matched to obtain the corresponding discrete point set. Calculate the descriptors on both sides of each discrete point based on the feature description grid. Aggregate the descriptors on both sides of each discrete point on the line segment to obtain the line feature descriptors on both sides. The process of determining the descriptors of discrete points is as follows: A. Divide the line segments in the image pair to be matched according to a set interval to obtain a discrete set of points to describe the line. Use the endpoint with the smaller x-coordinate as the starting point of the line segment to determine the direction of the line segment. B. Determine the grid point on the line segment that is closest to the discrete point in the grid where the discrete point is located as the first grid point, and determine whether the first grid point is on the left or right side of the line segment along the direction of the line segment; C. Map the obtained first grid point to the opposite side of the line segment where the discrete point is located to obtain the opposite mapped point, and assign the interpolated feature value of the opposite mapped point to the first grid point as the feature value of the first grid point; D. Calculate the descriptor of the discrete point based on the difference between the horizontal and vertical coordinate axes of the discrete point and the grid point, as well as the characteristic values ​​of each grid point in the grid. The first grid point uses the assigned characteristic value. 3) Perform line feature matching based on the descriptors on both sides of the obtained line feature.

2. The image line feature matching method considering significant image viewpoint differences according to claim 1, characterized in that, In step 3), when performing line feature matching, the known corresponding points in the image pair to be matched are first grouped for matching to obtain corresponding candidate line feature matching pairs. Line feature matching is then performed based on the obtained candidate line feature matching pairs. The process for determining the candidate line feature matching pairs is as follows: Circular search windows with radius R are set in the reference image and the image to be registered, respectively, with the corresponding points as the center. Line features located within the window are grouped together for matching. In the reference image and the image to be registered, the line features of the same group of corresponding points are obtained respectively. The distance between the descriptors on both sides of the line features in the group is calculated to obtain the distance matrix in each group. Select the line pairs corresponding to the minimum values ​​in the row and column of each matrix as candidate feature matching pairs.

3. The image line feature matching method considering significant image perspective differences according to claim 2, characterized in that, The method also includes filtering the one-to-many and many-to-many matching results and selecting the line feature matching pairs with the smallest distance.

4. The image line feature matching method considering significant image viewpoint differences according to claim 3, characterized in that, It also includes verifying the one-to-one matching results. The verification process is as follows: Search for multiple pairs of corresponding points in the baseline image that are closest to the corresponding line segment in the line feature matching pair to be verified. Construct a positive topological descriptor based on the distribution relationship of each corresponding point in each pair on the left and right sides of the corresponding line segment in the line feature matching pair to be verified. Search for multiple pairs of corresponding points in the image to be registered that are closest to the corresponding line segment in the line feature matching pair to be verified. Construct a reverse topological descriptor based on the distribution relationship of each corresponding point in each pair on the left and right sides of the corresponding line segment in the line feature matching pair to be verified. The similarity of the line feature matching pairs to be verified is determined based on the forward and reverse topological descriptors. If the similarity is not less than the set threshold, the line feature matching pairs to be verified meet the verification and are considered successfully matched. Otherwise, the verification is not met and they are considered unsuccessfully matched.

5. The image line feature matching method considering significant image viewpoint differences according to claim 4, characterized in that, The threshold value is set to 0.8n, where n is the number of identical point pairs searched.

6. The image line feature matching method considering significant image viewpoint differences according to claim 4, characterized in that, The formula for calculating the similarity is: Line feature matching pairs to be verified similarity, , Forward topology descriptor, is the reverse topology descriptor, and n is the number of identical point pairs to search.

7. The image line feature matching method considering significant image viewpoint differences according to any one of claims 1-4, characterized in that, In step 1), fusion refers to performing a convolution operation after numerically superimposing the grayscale and single-pixel directional gradient histogram features pixel by pixel to obtain an image feature grid.

8. The image line feature matching method considering significant image viewpoint differences according to claim 7, characterized in that, The convolution operation described uses a superpoint network.

9. The image line feature matching method considering significant image viewpoint differences according to any one of claims 1-4, characterized in that, The acquisition of pixel-level directional gradient histogram features includes: calculating the gradient magnitude and direction of each pixel, dividing the gradient direction [0°, 180°] into 9 directional intervals, forming blocks of 2×2 units of size 4×4 using a sliding window, statistically analyzing the image directional gradient histogram on a block-by-block basis, and normalizing it using the L2 norm.

10. The image line feature matching method considering significant image viewpoint differences according to claim 1, characterized in that, The formula for calculating the descriptor in step D is as follows: For discrete points The left-hand descriptor; , Discrete points The difference between the x and y coordinates of the point in the top-left grid of the grid; , , and Discrete points The characteristic values ​​of the top-left, top-right, bottom-left, and bottom-right grid points of the given grid.

Images