CN115063615A - Repeated texture image matching method based on Delaunay triangulation - Google Patents
Repeated texture image matching method based on Delaunay triangulation Download PDFInfo
- Publication number
- CN115063615A CN115063615A CN202210664988.6A CN202210664988A CN115063615A CN 115063615 A CN115063615 A CN 115063615A CN 202210664988 A CN202210664988 A CN 202210664988A CN 115063615 A CN115063615 A CN 115063615A
- Authority
- CN
- China
- Prior art keywords
- points
- point
- matching
- feature
- seed
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/757—Matching configurations of points or features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a repeated texture image matching method based on Delaunay triangulation, which comprises the following steps: extracting feature points of the two images by adopting an SIFT algorithm to obtain a feature vector of each feature point; calculating Euclidean distances between feature vectors of feature points in the two images, and forming initial matching between the feature points of the two images according to nearest neighbor of the feature vectors; inhibiting and selecting the well-represented initial matching as a seed point pair by using a local non-maximum value; filtering seed point pairs which do not conform to the global topological structure based on Delaunay triangulation; and respectively dividing the two images into a plurality of circular areas by taking the filtered seed point pairs as centers, and fitting affine transformation of local areas in the circular areas by using a PROSAC algorithm to remove wrong matching. When the method is adopted to match images with more similar textures, the accuracy of the matching point pairs is improved. The method has obvious effect on removing the error matching caused by repeated textures.
Description
Technical Field
The invention belongs to the technical field of image matching, and relates to a repeated texture image matching method based on Delaunay triangulation.
Background
Image matching means that corresponding matching points are searched in two pictures with a common area. It is the basis of many visual tasks, such as three-dimensional reconstruction, where the base matrix between two images can be obtained by finding the corresponding points in them, thereby reconstructing the coordinates of these points in three-dimensional space. Image matching can also be applied to image splicing, and corresponding points in two images can be found to be spliced into one image. Image matching also plays a very important role in tasks such as image registration, remote sensing and SLAM, and has therefore led to extensive research.
The existing image matching methods include a simple filter-based method, a global-based method, a local-based method and the like. In the existing image matching method, the ratio test of Lowe is a simple and effective method for removing the error matching, and abnormal values are filtered by setting a threshold value for the ratio of nearest neighbor to next neighbor. A simple filter also has two-way matching detection, where one matching way is obtained when finding the corresponding points of the feature points in the first map in the second map, and conversely another matching way may be obtained when finding the corresponding points of the feature points in the second map in the first map. The bidirectional matching detection reserves point pairs which can be matched in both directions, and removes point pairs which can not be matched in both directions. A simple filter may effectively remove the false matches, but the results obtained by this method of removing the false matches only by the feature vectors of the local feature descriptors are still not very accurate.
The most widely used global-based method is the RANSAC type of robust matcher. Chinese patent CN103400388A utilizes the RANSAC method to eliminate the mismatch, the RANSAC method randomly samples from the assumed set, extracts the least samples for generating the model, then returns the support of the model, and selects the model with the greatest support after reaching the maximum number of iterations. The method is simple, strong in robustness and widely adopted. However, the matching point pairs finally obtained by the global RANSAC-based method are concentrated in texture-rich areas, and the obtained matching number is small. The global RANSAC algorithm also runs longer when there are more mismatches.
Still other approaches eliminate the mismatch by local neighborhood constraints. The GMS algorithm proposed in doi:10.1109/CVPR.2017.302 assumes that adjacent pixels in an image move together, realizes motion statistics based on grids, improves the performance of outlier filtering, greatly reduces the running time of the algorithm at the same time, and can realize real-time application. AdaLAM proposed by ECCV 2020, pages 770-787 integrates the best practices in a large number of manual methods well-established in the field into a coherent framework for rapid and efficient outlier filtering by conventional methods. The method based on the local area can realize the parallelism of each area, greatly reduce the running time and obtain more uniform distribution of the matching points in the image. Local area-based methods such as GMS and AdaLAM can make the distribution of matching point pairs more uniform and denser, and the method of dividing the image into several small areas that do not affect each other can speed up the algorithm. However, in a picture with a large number of repeated textures (such as the appearance of a building), the algorithm based on a local region may cause a mismatch condition of the whole region, as shown in fig. 1, which is also a difficulty in image matching.
Disclosure of Invention
In order to solve the above technical problems, an object of the present invention is to provide a repeated texture image matching method based on Delaunay triangulation, which adds global constraints to a local region-based method, combines the advantages of the two methods, can obtain dense and uniform matching point pairs, and has an obvious effect on removing false matching caused by repeated textures.
The invention provides a repeated texture image matching method based on Delaunay triangulation, which comprises the following steps:
step 1: extracting feature points of the two images by adopting an SIFT algorithm, and obtaining a feature vector of each feature point;
step 2: calculating Euclidean distance between feature vectors of feature points in the two images, and forming initial matching between the feature points of the two images according to nearest neighbor of the feature vectors;
and step 3: inhibiting and selecting the well-represented initial matching as a seed point pair by using a local non-maximum value;
and 4, step 4: filtering seed point pairs which do not conform to the global topological structure based on Delaunay triangulation;
and 5: and respectively dividing the two images into circular areas by taking the filtered seed point pairs as centers, and fitting affine transformation of local areas in the circular areas by using a PROSAC algorithm so as to remove wrong matching.
In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 1 includes:
step 1.1: respectively carrying out Gaussian blur of different scales on the two images to obtain a series of images with different sizes;
step 1.2: searching extreme points in images with different scales as characteristic points;
step 1.3: and extracting no more than 8000 feature points from each image, and obtaining a 128-dimensional feature vector for describing the feature points according to a gradient histogram in a 4 multiplied by 4 region around the feature points.
In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 2 specifically is: selecting the feature points in the second image for each feature point in the first image to form initial matching, wherein the Euclidean distance between two feature vectors of the two initially matched feature points is minimum, and the method comprises the following steps:
step 2.1: recording a first image G with a distance matrix D 1 Feature vectors of middle feature points and second graph G 2 Euclidean distance of feature vectors of middle feature points, element D in distance matrix D ij Represents G 1 Middle ith characteristic point and G 2 The Euclidean distance of the feature vector of the jth feature point;
step 2.2: for G 1 Finds the minimum value D in the ith row of the distance matrix D ij Then G 1 The ith feature point in (1) and G 2 The j-th feature point in (1) isA pair of matching point pairs;
step 2.3: repeat step 2.2 to find G 1 Form a number of pairs of initially matched points.
In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 3 specifically is:
step 3.1: finding the next nearest neighbor of each feature point in the first image, i.e. the penultimate value D of each row in the distance matrix D ik ；
Step 3.2: giving a proportion test score rt to each initially matched point pair, and generating the nearest neighbor d of the feature vector of each feature point in the first graph found in the initial matching ij And next neighbor d ik As the ratio test score rt of the initial matching point pair:
step 3.3: selected at radius R with local non-maximum suppression 1 The initial matching point pair with the smallest ratio test score rt in the circular local area of (1) is used as the seed point pair.
In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 3.3 specifically is:
step 3.3.1: setting a pair of initially matched point pairs X 1 、X 2 Is rt 1 In the first image by X 1 As a center, R 1 Within a circular local area of radius, if the scale test scores of the remaining initially matched point pairs are all greater than rt 1 Large, then X 1 、X 2 The pair of initially matched point pairs is selected as a seed point pair;
step 3.3.2: the radius R in the two images is calculated according to the fact that the image area is 120 times of the area of the circular local area 1 And R 2 。
In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 4 specifically is:
step 4.1: performing Delaunay triangulation on the seed points of the first image;
step 4.2: connecting corresponding seed points in the second graph according to the connection mode of the seed points obtained by triangulation in the first graph, wherein if the second graph is a triangular grid without intersection points after connection, all the seed point pairs are correctly matched; otherwise, executing step 4.3;
step 4.3: for the seed points in the second graph, the average value of the intersection points of the lines connected with the seed points and other lines is counted, namely the average intersection point number
Wherein N is the number of lines connected with the seed points, and N is the number of intersections of the lines connected with the seed points and other lines;
step 4.3: counting the average intersection point number of each seed point in the second graph in sequence, setting a threshold value, and removing the seed point and the corresponding seed point in the first graph if the average intersection point number of one seed point is greater than the threshold value; the threshold is the average of the average number of intersections of all the seed points.
In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 5 specifically is:
step 5.1: dividing the two images into circular areas by taking the filtered seed point pairs as centers;
step 5.2: sorting the proportional test scores of the initial matching point pairs in each circular area divided in the step 5.1 from small to large;
step 5.3: the initial matching point pairs in each circular area conform to the central affine transformation, and the formula is as follows:
wherein x, y and x ', y' represent coordinate values of the initially matched point pairs in the first graph and the second graph,
obtaining two equality constraints through a group of initially matched point pairs, and calculating an affine matrix through the two groups of initially matched point pairs;
step 5.4: adopting a PROSAC algorithm to take two groups of initially matched point pairs which are ranked at the top 2 bits in each circular area during the first iteration, randomly selecting two groups of point pairs from the initially matched point pairs which are ranked at the top 3 bits in each circular area during the second iteration, randomly selecting two groups of point pairs from the initially matched point pairs which are ranked at the top 4 bits in the third iteration, sequentially taking points in an increasing sampling space, and respectively adopting the two groups of point pairs selected in each iteration to calculate an affine matrix A corresponding to the current iteration times for each circular area i I represents the number of iterations;
step 5.5: according to r i ＝||A i ·[x,y] T -[x',y'] T I, calculating the errors r of all the initial matching point pairs except two sampling points in each circular area during each iteration i If the error is smaller than the error threshold, the initial matching point pair is an inner point, if the error is larger than the error threshold, the number of the inner points of each circular area of each iteration is recorded, and after 128 iterations, the affine matrix calculated by the iteration with the largest number of the inner points is selected as the affine matrix of the circular area;
step 5.6: for each circular area, the outer points of the affine matrix that do not fit the circular area are removed.
In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 5.1 specifically is:
step 5.1.1: r's' 1 Is R 1 3 times of (1), R' 2 Is R 2 3 times of the total weight of the composition;
step 5.1.2: through the step 3.3, a series of seed point pairs are selected, and R 'is respectively taken as the center of the seed points in the seed point pairs' 1 And R' 2 The first image and the second image are divided into circular regions for the radius.
According to the repeated texture image matching method based on Delaunay triangulation, seed points which do not conform to a global topological structure are filtered through a Delaunay triangulation algorithm, global constraints are added to the method based on a local area, and the advantages of the seed points and the local area are combined. The method has the advantages that the method has an obvious improvement effect on the condition that the matching of the whole area is wrong when the similar texture of the image is more, and the accuracy of the finally obtained matching point pair is improved. Experiments show that the method can finally obtain dense and uniform matching point pairs, and has obvious effect on removing the error matching caused by the repeated texture.
Drawings
FIG. 1 is a flow chart of a repeated texture image matching method based on Delaunay triangulation according to the present invention;
FIG. 2a is a diagram of a triangular network of the first graph;
FIG. 2b is a diagram of the second graph of a triangular network without intersection points;
FIG. 2c is a diagram of the triangular network with intersections of the second graph;
FIG. 3 is a schematic diagram of a local neighborhood partition centered at a seed point.
Detailed Description
The image matching method based on the global situation is long in calculation time, and the obtained matching point pairs are distributed unevenly in the image. The image matching based on the local area can widely cover each area of the image, and abundant matching point pairs are obtained. However, when a large number of similar textures are included in an image, a region is easily mismatched by only a local relationship. In order to obtain a large number of uniform matching point pairs and simultaneously ensure the matching accuracy, the invention combines global and local constraints and optimizes the seed points through Delaunay triangulation according to the invariance of a scene structure. Firstly, Delaunay triangulation is carried out on the seed points in the first graph to form a triangular mesh, and then corresponding points in the second graph are connected according to the connection mode of the seed points in the first graph. If the matching point pairs are all correct, a triangular mesh without intersection points can still be formed after connection, and otherwise, an error exists. According to the method, the seed points are filtered by setting the threshold value for the number of the average intersection points, so that the condition of region mismatching caused by a large number of similar textures is effectively reduced, and the finally obtained matching is more accurate.
The repetitive texture image matching method based on Delaunay triangulation according to the present invention will be described in detail with reference to the accompanying drawings.
As shown in fig. 1, the repeated texture image matching method based on Delaunay triangulation of the present invention includes the following steps:
step 1: extracting feature points of the two images by adopting an SIFT algorithm, and obtaining a feature vector of each feature point, wherein the step 1 comprises the following steps:
step 1.1: respectively carrying out Gaussian blur of different scales on the two images to obtain a series of images with different sizes;
step 1.2: searching extreme points in images with different scales as characteristic points;
step 1.3: each image extracts no more than 8000 feature points, and obtains 128-dimensional feature vectors describing the feature points from gradient histograms (each histogram has 8 directions) in 4 × 4 regions around the feature points.
Step 2: calculating Euclidean distances between feature vectors of feature points in the two images, and forming initial matching between the feature points of the two images according to nearest neighbor of the feature vectors;
the step 2 specifically comprises the following steps: selecting the feature points in the second image for each feature point in the first image to form initial matching, wherein the Euclidean distance between two feature vectors of the two initially matched feature points is minimum, and the method comprises the following steps:
step 2.1: recording a first image G with a distance matrix D 1 Feature vector of middle feature point and second graph G 2 Euclidean distance of feature vectors of middle feature points, element D in distance matrix D ij Represents G 1 Middle ith characteristic point and G 2 The Euclidean distance of the feature vector of the jth feature point;
step 2.2: for G 1 Finds the minimum value D in the ith row of the distance matrix D ij Then G 1 The ith character ofSymbol point and G 2 The jth feature point in (a) is a pair of matching point pairs;
step 2.3: and (3) repeating the step 2.2 to find out the matching points of all the characteristic points in the G1, and forming a plurality of initially matched point pairs.
The initial match obtained in step 2 is found based on the nearest neighbor of the feature vector, and the hypothetical match formed only by the local descriptor is quite inaccurate, so that the hypothetical match needs to be further filtered.
And step 3: inhibiting and selecting the well-represented initial matching as a seed point pair by using a local non-maximum value;
in order to retain the advantages of high calculation speed and uniform distribution of obtained seed points based on a local method, a method based on a local area is adopted firstly. When the neighborhood is divided, a local non-maximum value is adopted to inhibit and select a seed point, and then the neighborhood is divided near the seed point. Local non-maximum suppression is performed by first assigning a score to each initially matched pair of points, where the value of the scale test is used as the score for the matched pair of points. The step 3 specifically comprises the following steps:
step 3.1: finding the next nearest neighbor of each feature point in the first image, i.e. the penultimate value D of each row in the distance matrix D ik ；
Step 3.2: giving a proportion test score rt to each initially matched point pair, and generating the nearest neighbor d of the feature vector of each feature point in the first graph found in the initial matching ij And next neighbor d ik As the ratio of the initial matching point pair, the ratio of (i) to (ii) is used as the ratio test score rt:
step 3.3: selected at radius R with local non-maximum suppression 1 The initial matching point pair with the smallest proportion test score rt in the circular local area is used as a seed point pair, and specifically comprises the following steps:
step 3.3.1: setting a pair of initially matched point pairs X 1 、X 2 Is rt 1 In the first placeBy X in the image 1 As a center, R 1 Within a circular local area of radius, if the scale test scores of the remaining initially matched point pairs are all greater than rt 1 Large, then X 1 、X 2 The pair of initially matched point pairs is selected as a seed point pair;
step 3.3.2: the radius R in the two images is calculated according to the condition that the image area is 120 times of the area of the circular local area 1 And R 2 。
And 4, step 4: filtering seed point pairs which do not conform to the global topological structure based on Delaunay triangulation;
in step 3, a series of seed point pairs are selected first, and next, the error matching elimination is performed in the local neighborhood. The method can enable each neighborhood to operate in parallel, greatly shortens the running time, and enables the obtained final matching to be distributed more uniformly in the image. But only with local neighborhood based methods can produce a mismatch of the entire region in a heavily textured image. To improve the situation, a constraint based on a global topology is added, and the seed points are filtered.
When the same scene is shot from two different angles, the topological structure of the scene is stable and cannot be changed along with the change of the shooting angle. For example, the relative positions of points are unchanged no matter which direction the human face structure is shot from, and ears are always on two sides of eyes and cannot run to the middle of the eyes. Therefore, we can further constrain the seed points by the global topological distribution. The invention adopts Delaunay triangulation to judge the topological distribution of the seed points in the two graphs.
The triangulation is to connect point sets on a plane by using a closed line segment to form a triangular net, the connected graphs have no intersecting edges, and the collection of the triangle is a convex hull of a scattered point set. There are many kinds of results obtained by triangulation of a point set on a plane, the more symmetrical a dissected triangle is, the better the subsequent image processing effect is, in order to make the obtained result unique and make the dissected triangle more symmetrical, a Delaunay triangulation, which is a special triangulation, is used herein. The Delaunay triangulation meets the characteristics of a hollow circle, any four points cannot be in a common circle, and the formed triangular mesh is unique. The step 4 specifically comprises the following steps:
step 4.1: performing Delaunay triangulation on the seed points of the first image to obtain a triangular network shown in FIG. 2 a;
step 4.2: connecting corresponding seed points in the second graph according to the connection mode of the seed points obtained by triangulation in the first graph, wherein if the second graph is a triangular mesh without intersection points after connection as shown in FIG. 2b, all the seed point pairs are correctly matched; if the line segment cross condition exists after the connection, the point with matching error is shown as figure 2c, otherwise, step 4.3 is executed;
step 4.3: for the seed points in the second graph, the average value of the intersection points of the lines connected with the seed points and other lines is counted, namely the average intersection point number
Wherein N is the number of lines connected with the seed points, and N is the number of intersections of the lines connected with the seed points and other lines;
step 4.3: counting the average intersection point number of each seed point in the second graph in sequence, setting a threshold value, and removing the seed point and the corresponding seed point in the first graph if the average intersection point number of one seed point is greater than the threshold value; the threshold is the average of the average number of intersections of all the seed points.
Through the step 4, the global constraint is combined with the error matching filtering based on the local region, and the wrong seed points can be preliminarily filtered, so that the condition that the subsequent whole neighborhood is matched wrongly is reduced.
And 5: dividing the two images into circular areas respectively by taking the filtered seed point pairs as centers, and fitting affine transformation of local areas in the circular areas by using a PROSAC algorithm so as to remove wrong matching, wherein the step 5 specifically comprises the following steps:
step 5.1: dividing the two images into circular areas by taking the filtered seed point pairs as centers respectively, specifically:
step 5.1.1: r's' 1 Is R 1 3 times of (1), R' 2 Is R 2 3 times of the total weight of the composition;
step 5.1.2: through the step 3.3, a series of seed point pairs are selected, and R 'is respectively taken as the center of the seed points in the seed point pairs' 1 And R' 2 The first image and the second image are divided into circular regions for the radius, as shown in fig. 3.
Step 5.2: sorting the proportional test scores of the initial matching point pairs in each circular area divided in the step 5.1 from small to large;
step 5.3: the initial matching point pairs in each circular area conform to the central affine transformation, and the formula is as follows:
wherein x, y and x ', y' represent coordinate values of the initially matched point pairs in the first graph and the second graph,
obtaining two equality constraints through a group of initially matched point pairs, and calculating an affine matrix through the two groups of initially matched point pairs;
step 5.4: adopting a PROSAC algorithm to take two groups of initially matched point pairs which are ranked at the top 2 bits in each circular area during the first iteration, randomly selecting two groups of point pairs from the initially matched point pairs which are ranked at the top 3 bits in each circular area during the second iteration, randomly selecting two groups of point pairs from the initially matched point pairs which are ranked at the top 4 bits in the third iteration, sequentially taking points in an increasing sampling space, and respectively adopting the two groups of point pairs selected in each iteration to calculate an affine matrix A corresponding to the current iteration times for each circular area i I represents the number of iterations;
step 5.5: according to r i ＝||A i ·[x,y] T -[x',y'] T I calculates all the rest initial points except two sampling points in each circular area during each iterationError r of matching point pair i Recording the number of the inner points of each circular area of each iteration, and selecting an affine matrix calculated by the iteration with the largest number of the inner points as an affine matrix of the circular area after 128 iterations, wherein the initial matching point pair with the error smaller than the error threshold is an inner point, and the initial matching point pair with the error larger than the error threshold is an outer point;
step 5.6: for each circular area, the outer points of the affine matrix that do not fit the circular area are removed.
The similar texture is difficult in image matching, global constraint is added into PROSAC based on a local region, and Delaunay triangulation is applied to filter seed points which do not conform to a global topological structure, so that the formation of region mismatching is avoided. Through the improvement, the condition that the matching of the whole area is wrong when the similar texture of the image is more is improved, and the accuracy of the finally obtained matching point pair is improved. The method can finally obtain dense and uniform matching point pairs, and has obvious effect on removing the error matching caused by the repeated texture.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the scope of the present invention, which is defined by the appended claims.
Claims (8)
1. A repeated texture image matching method based on Delaunay triangulation is characterized by comprising the following steps:
step 1: extracting feature points of the two images by adopting an SIFT algorithm, and obtaining a feature vector of each feature point;
step 2: calculating Euclidean distance between feature vectors of feature points in the two images, and forming initial matching between the feature points of the two images according to nearest neighbor of the feature vectors;
and step 3: inhibiting and selecting the well-represented initial matching as a seed point pair by using a local non-maximum value;
and 4, step 4: filtering seed point pairs which do not conform to the global topological structure based on Delaunay triangulation;
and 5: and respectively dividing the two images into circular areas by taking the filtered seed point pairs as centers, and fitting affine transformation of local areas in the circular areas by using a PROSAC algorithm so as to remove wrong matching.
2. The method for repetitive texture image matching based on Delaunay triangulation as claimed in claim 1, wherein the step 1 comprises:
step 1.1: respectively carrying out Gaussian blur of different scales on the two images to obtain a series of images with different sizes;
step 1.2: searching extreme points in images with different scales as characteristic points;
step 1.3: and extracting no more than 8000 feature points from each image, and obtaining a 128-dimensional feature vector for describing the feature points according to a gradient histogram in a 4 multiplied by 4 region around the feature points.
3. The method for matching repeated texture images based on Delaunay triangulation as claimed in claim 1, wherein the step 2 is specifically: selecting the feature points in the second image for each feature point in the first image to form initial matching, wherein the Euclidean distance between two feature vectors of the two initially matched feature points is minimum, and the method comprises the following steps:
step 2.1: recording a first image G with a distance matrix D 1 Feature vectors of middle feature points and second graph G 2 Euclidean distance of feature vectors of middle feature points, element D in distance matrix D ij Represents G 1 Middle ith characteristic point and G 2 The Euclidean distance of the feature vector of the j-th feature point;
step 2.2: for G 1 Finds the minimum value D in the ith row of the distance matrix D ij Then G 1 The ith feature point in (1) and G 2 The jth feature point in (a) is a pair of matching point pairs;
step 2.3: repeat step 2.2 to find G 1 Form a number of pairs of initially matched points.
4. The method for repeated texture image matching based on Delaunay triangulation as recited in claim 3, wherein the step 3 is specifically:
step 3.1: finding the next nearest neighbor of each feature point in the first image, i.e. the penultimate value D of each row in the distance matrix D ik ；
Step 3.2: giving a proportion test score rt to each initially matched point pair, and generating the nearest neighbor d of the feature vector of each feature point in the first graph found in the initial matching ij And next neighbor d ik As the ratio of the initial matching point pair, the ratio of (i) to (ii) is used as the ratio test score rt:
step 3.3: selected at radius R with local non-maximum suppression 1 The initial matching point pair with the smallest ratio test score rt in the circular local area of (1) is used as the seed point pair.
5. The method for matching repeated texture images based on Delaunay triangulation as claimed in claim 1, wherein said step 3.3 is specifically:
step 3.3.1: setting a pair of initially matched point pairs X 1 、X 2 Is rt 1 In the first image by X 1 Is centered on R 1 Within a circular local area of radius, if the scale test scores of the remaining initially matched point pairs are all greater than rt 1 Large, then X 1 、X 2 The pair of initially matched point pairs is selected as a seed point pair;
step 3.3.2: the radius R in the two images is calculated according to the fact that the image area is 120 times of the area of the circular local area 1 And R 2 。
6. The method for matching repeated texture images based on Delaunay triangulation as claimed in claim 1, wherein the step 4 is specifically:
step 4.1: performing Delaunay triangulation on the seed points of the first image;
step 4.2: connecting corresponding seed points in the second graph according to the connection mode of the seed points obtained by triangulation in the first graph, wherein if the second graph is a triangular grid without intersection points after connection, all the seed point pairs are correctly matched; otherwise, executing step 4.3;
step 4.3: for the seed points in the second graph, the average value of the intersection points of the lines connected with the seed points and other lines is counted, namely the average intersection point number
Wherein N is the number of lines connected with the seed points, and N is the number of intersections of the lines connected with the seed points and other lines;
step 4.3: counting the average intersection point number of each seed point in the second graph in sequence, setting a threshold value, and removing the seed point and the corresponding seed point in the first graph if the average intersection point number of one seed point is greater than the threshold value; the threshold is the average value of the average number of the intersection points of all the seed points.
7. The method for repeated texture image matching based on Delaunay triangulation as recited in claim 4, wherein the step 5 specifically comprises:
step 5.1: dividing the two images into circular areas by taking the filtered seed point pairs as centers;
step 5.2: sorting the proportional test scores of the initial matching point pairs in each circular area divided in the step 5.1 from small to large;
step 5.3: the initial matching point pairs in each circular area conform to the central affine transformation, and the formula is as follows:
wherein x isY and x ', y' represent coordinate values of the initially matched pair of points in the first and second graphs,
obtaining two equality constraints through a group of initially matched point pairs, and calculating an affine matrix through the two groups of initially matched point pairs;
step 5.4: adopting a PROSAC algorithm to take two groups of initially matched point pairs which are ranked at the top 2 bits in each circular area during the first iteration, randomly selecting two groups of point pairs from the initially matched point pairs which are ranked at the top 3 bits in each circular area during the second iteration, randomly selecting two groups of point pairs from the initially matched point pairs which are ranked at the top 4 bits in the third iteration, sequentially taking points in an increasing sampling space, and respectively adopting the two groups of point pairs selected in each iteration to calculate an affine matrix A corresponding to the current iteration times for each circular area i I represents the number of iterations;
step 5.5: according to r i ＝||A i ·[x,y] T -[x',y'] T I calculates the error r of all initial matching point pairs except two sampling points in each circular area during each iteration i Recording the number of the inner points of each circular area of each iteration, and selecting an affine matrix calculated by the iteration with the largest number of the inner points as an affine matrix of the circular area after 128 iterations, wherein the initial matching point pair with the error smaller than the error threshold is an inner point, and the initial matching point pair with the error larger than the error threshold is an outer point;
step 5.6: for each circular area, the outer points of the affine matrix that do not fit the circular area are removed.
8. The method for matching repeated texture images based on Delaunay triangulation as claimed in claim 7, wherein said step 5.1 is specifically:
step 5.1.1: r's' 1 Is R 1 3 times of (1), R' 2 Is R 2 3 times of the total weight of the composition;
step 5.1.2: a series of pairs of seed points are selected, via step 3.3, with ones of the pairs of seed pointsWith seed point as center, each R' 1 And R' 2 The first image and the second image are divided into circular regions for the radius.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210664988.6A CN115063615A (en) | 2022-06-13 | 2022-06-13 | Repeated texture image matching method based on Delaunay triangulation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210664988.6A CN115063615A (en) | 2022-06-13 | 2022-06-13 | Repeated texture image matching method based on Delaunay triangulation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115063615A true CN115063615A (en) | 2022-09-16 |
Family
ID=83199914
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210664988.6A Pending CN115063615A (en) | 2022-06-13 | 2022-06-13 | Repeated texture image matching method based on Delaunay triangulation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115063615A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117036623A (en) * | 2023-10-08 | 2023-11-10 | 长春理工大学 | Matching point screening method based on triangulation |
-
2022
- 2022-06-13 CN CN202210664988.6A patent/CN115063615A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117036623A (en) * | 2023-10-08 | 2023-11-10 | 长春理工大学 | Matching point screening method based on triangulation |
| CN117036623B (en) * | 2023-10-08 | 2023-12-15 | 长春理工大学 | Matching point screening method based on triangulation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107563381B (en) | Multi-feature fusion target detection method based on full convolution network | |
| CN106778604B (en) | Pedestrian re-identification method based on matching convolutional neural network | |
| CN106023298B (en) | Point cloud Rigid Registration method based on local Poisson curve reestablishing | |
| CN111882593A (en) | Point cloud registration model and method combining attention mechanism and three-dimensional graph convolution network | |
| Choy et al. | High-dimensional convolutional networks for geometric pattern recognition | |
| CN111625667A (en) | Three-dimensional model cross-domain retrieval method and system based on complex background image | |
| Yue et al. | Robust loop closure detection based on bag of superpoints and graph verification | |
| CN110827304B (en) | Traditional Chinese medicine tongue image positioning method and system based on deep convolution network and level set method | |
| CN108615229B (en) | Collision detection optimization method based on curvature point clustering and decision tree | |
| CN113610905B (en) | Deep learning remote sensing image registration method based on sub-image matching and application | |
| CN111582091A (en) | Pedestrian identification method based on multi-branch convolutional neural network | |
| CN112946679A (en) | Unmanned aerial vehicle surveying and mapping jelly effect detection method and system based on artificial intelligence | |
| CN115063615A (en) | Repeated texture image matching method based on Delaunay triangulation | |
| Yin et al. | A new apple segmentation and recognition method based on modified fuzzy C-means and hough transform | |
| CN114689038A (en) | Fruit detection positioning and orchard map construction method based on machine vision | |
| CN114495170A (en) | Pedestrian re-identification method and system based on local self-attention inhibition | |
| CN108108700B (en) | Pig feature region identification method based on chord axis transformation | |
| CN114782455B (en) | Cotton row center line image extraction method for agricultural machine embedded equipment | |
| CN112614161A (en) | Three-dimensional object tracking method based on edge confidence | |
| CN110472495B (en) | Deep learning face recognition method based on graphic reasoning global features | |
| Liu et al. | Deep learning of directional truncated signed distance function for robust 3D object recognition | |
| CN114494586B (en) | Lattice projection deep learning network broadleaf branch and leaf separation and skeleton reconstruction method | |
| Geng et al. | A novel color image segmentation algorithm based on JSEG and Normalized Cuts | |
| Hou et al. | Forest: A Lightweight Semantic Image Descriptor for Robust Visual Place Recognition | |
| Jiang et al. | Superpixel segmentation based gradient maps on RGB-D dataset |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |