Image registration method based on multi-source homogeneous structural point features

Abstract

The present invention discloses an image registration method based on multi-source homogeneous structure point features. First, the phase information of the image is used to extract the feature information of the image, which is not affected by the grayscale change and radiation difference between heterogeneous images, so as to obtain feature points with strong robustness and high stability. Then, for the reference image and the image to be registered, the improved Sobel operator is used to calculate the gradient amplitude and angle information, and a direction histogram is constructed in the circular neighborhood of the feature point, which overcomes the influence of SAR speckle noise interference and avoids over-reliance on grayscale characteristics. Then, the gradient direction corresponding to the main peak of the gradient histogram is selected as the main direction of the feature point, and the gradient amplitude and angle and the main direction are used to obtain a feature descriptor with a high degree of consistency in a logarithmic polar coordinate system. Finally, a fast sampling consistency algorithm is used to eliminate mismatched points, thereby reducing the root mean square error of the algorithm and improving the feature matching accuracy, making the registration of multi-source rotated images more accurate and rapid.

Description

Technical Field

[0001] The present invention belongs to the technical field of image registration, and in particular relates to an image registration method based on multi-source homogeneous structural point features. Background Art

[0002] Image registration technology is a prerequisite for various remote sensing applications and has been widely used in fields such as drone navigation, environmental monitoring, terrain and mountain mapping, biological diagnostics, and climate prediction. Heterogeneous image registration, which can establish a matching relationship and perform corrective alignment between two or more images of the same scene captured by different sensors at different times and perspectives, is a current research hotspot in image registration technology.

[0003] Currently, the mainstream multi-source remote sensing image registration methods include traditional registration algorithms and deep learning-based registration algorithms. Among them, the traditional multi-source remote sensing image registration algorithms are mainly represented by image registration methods based on regional grayscale information and image registration methods based on feature information.

[0004] Image registration based on regional grayscale information constructs a similarity measurement function according to the grayscale information between images from different sensors, and calculates the extreme points of the reference image and the image to be registered through an optimization algorithm, thereby realizing the calculation of the transformation model parameters.

[0005] Feature information is a more complete and comprehensive description of image information and can exist stably in multi-source remote sensing images. Therefore, image registration methods based on feature information are often used for matching between multi-source remote sensing images and have become the main research direction in this field at home and abroad. Image registration methods based on feature information mainly include feature registration schemes based on intensity, gradient, or phase information. Among them, using intensity or gradient information for feature matching is more common, and using phase information for feature matching is relatively rare. The main commonality of these methods is that they first extract point features, line features, surface features, etc. from different images, and then use the extracted features to generate consistent feature descriptors to complete the rough matching of the two image features. After a series of error elimination algorithms, the feature matching is achieved, and the transformation model of the matching features is calculated, thereby achieving high-precision registration.

[0006] However, due to the significant nonlinear grayscale and radiation intensity differences between multi-source remote sensing images, existing region-based registration algorithms consume a significant amount of time and computer performance due to the need to traverse each pixel in the image one by one during the optimization phase of the search algorithm. As a result, these methods are time-consuming, overly dependent on image grayscale information, and sensitive to rotation changes. In addition, the intensity or gradient information used in existing feature-based image registration methods is sensitive to nonlinear grayscale and radiation intensity differences between heterogeneous images, resulting in unstable extracted feature points, low feature matching accuracy, long time consumption, and difficulty adapting to the registration of multi-source rotated images. Summary of the Invention

[0007] In order to solve the above problems existing in the prior art, the present invention provides an image registration method based on multi-source homogeneous structural point features. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0008] An image registration method based on multi-source homogeneous structural point features, comprising:

[0009] Acquire a first two-dimensional phase congruency overlay image corresponding to the reference image and a second two-dimensional phase congruency overlay image corresponding to the image to be registered;

[0010] Performing SUSAN corner point detection on the first two-dimensional phase consistency overlay image and the second two-dimensional phase consistency overlay image respectively to extract feature point information, and correspondingly obtaining a plurality of first feature points and a plurality of second feature points;

[0011] The improved Sobel operator is used to calculate the gradient amplitude and angle of the reference image and the image to be registered respectively;

[0012] Constructing a gradient consistency direction histogram based on the circular neighborhood of the first feature point and the second feature point, and correspondingly obtaining a first direction histogram and a second direction histogram;

[0013] Taking the direction information corresponding to the main peak of the first direction histogram and the second direction histogram as the main direction of the feature point, and combining the gradient amplitude angle to construct a feature descriptor with rotation invariance in a logarithmic polar coordinate system, thereby obtaining a first feature descriptor and a second feature descriptor respectively;

[0014] A similarity measurement is performed on the first feature descriptor and the second feature descriptor, and mismatched points are eliminated to obtain a registration result between the image to be registered and the reference image.

[0015] Beneficial effects of the present invention:

[0016] The present invention proposes an image registration method which first uses the phase information of the image to extract the feature information of the image, which is not affected by the grayscale changes and radiation differences between heterogeneous images, thereby obtaining feature points with strong robustness and high stability; then, for the reference image and the image to be registered, the improved Sobel operator is used to calculate the gradient amplitude and angle information, and a direction histogram is constructed in the circular neighborhood of the feature point, which overcomes the influence of SAR speckle noise interference and avoids over-reliance on grayscale characteristics; then, the gradient direction corresponding to the main peak of the gradient histogram is selected as the main direction of the feature point, and the gradient amplitude and angle and the main direction are used to obtain a feature descriptor with a high degree of consistency in the logarithmic polar coordinate system; finally, a fast sampling consistency algorithm is used to eliminate mismatched points, which reduces the root mean square error of the algorithm and improves the feature matching accuracy, making the registration of multi-source rotated images more accurate and rapid.

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of an image registration method based on multi-source homogeneous structural point features provided by an embodiment of the present invention;

[0019] Figure 2 1 is another flow chart of an image registration method based on multi-source homogeneous structural point features provided by an embodiment of the present invention;

[0020] Figure 3 Schematic diagram of a circular template constructed in the SUSAN corner detection algorithm provided by an embodiment of the present invention;

[0021] Figure 4 is the registration result of the visible light image and SAR image in simulation experiment 1;

[0022] Figure 5 is the registration result of the visible light image and infrared image in simulation experiment 2;

[0023] Figure 6 It is the registration result of SAR image and SAR image in simulation experiment 3. DETAILED DESCRIPTION

[0024] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0025] Example 1

[0026] Please see the joint Figure 1-2 , Figure 1 This is a flow chart of an image registration method based on multi-source homogeneous structural point features provided by an embodiment of the present invention. Figure 2This is another flowchart of an image registration method based on multi-source homogeneous structural point features provided by an embodiment of the present invention.

[0027] The method provided by the present invention utilizes the phase information of the image to extract the feature information of the reference image and the image to be registered, which is not affected by the grayscale changes and radiation differences between the heterogeneous images, thereby obtaining feature points with strong robustness and high stability; the gradient amplitude and gradient angle information of each reference image and the image to be registered are calculated, and then a direction histogram is constructed in the circular neighborhood of the feature point, and the gradient direction corresponding to the main peak of the gradient histogram is selected as the main direction of the feature point; the gradient amplitude angle and main direction are used to obtain feature descriptors in a logarithmic polar coordinate system; and finally, a fast sampling consistency algorithm is used to eliminate mismatched points.

[0028] It can be understood that, in this embodiment, both the reference image and the image to be registered are remote sensing images, and the image to be registered is a real-time image.

[0029] Specifically, the image registration method based on multi-source homogeneous structural point features provided in this embodiment includes the following steps:

[0030] Step 1: Obtain a first two-dimensional phase congruency overlay image corresponding to the reference image and a second two-dimensional phase congruency overlay image corresponding to the image to be registered.

[0031] Because phase congruency information is less affected by grayscale differences and is highly robust to image radiation distortion and noise, the extracted feature points are more robust. Based on this, this embodiment designs an image feature detection algorithm based on phase congruency theory. Using a Log-Gabor filter, the phase congruency (PC) information at different angles and scales corresponding to the pixel points at each position in the reference image and the image to be registered is obtained and superimposed to obtain a first two-dimensional phase congruency overlay image and a second two-dimensional phase congruency overlay image.

[0032] Specifically, the Log-Gabor wavelet has good positioning performance in both the spatial domain and the frequency domain, and can be used to extract multi-directional and multi-scale edge information from the input image. The function of the two-dimensional Log-Gabor (2D-LGF) filter in the frequency domain is defined as follows:

[0033]

[0034] Among them, r and θ represent the polar diameter and polar angle in the logarithmic polar coordinate system, s and o represent the scale coefficient and angle coefficient of the filter, f0 represents the center frequency of the filter, and θ (s,o) represents the direction of the filter, σ r and σ θare the radial bandwidth and angular bandwidth of the filter.

[0035] Since this function does not have a DC component, it is not affected by the grayscale of the image. The spatial domain filter corresponding to the 2D-LGF can be obtained by inverse Fourier transform, and its expression is as follows:

[0036] LG(x,y)=LG even (x,y)+i·LG odd (x,y);

[0037] Among them, LG even (x,y) and LG odd (x, y) are the even-symmetric wavelet and odd-symmetric wavelet expressions after decomposition of the Log-Gabor filter, respectively. Both filters have good directionality and locality.

[0038] By convolving the input image with the two spatial domain filters mentioned above, we can obtain its Log-Gabor response components at different angles and scales. The specific definitions are as follows:

[0039] e so (x,y)=I(x,y)*LG even (x,y);

[0040] o so (x,y)=I(x,y)*LG odd (x,y);

[0041] Among them, e so (x,y) is the scale response component, o so (x,y) is the angle response component. Therefore, the amplitude component A of the corresponding image at the scale and direction can be obtained. so (x,y) and phase component φ so (x,y), which is defined as follows:

[0042]

[0043]

[0044] In addition, considering that images usually contain noise interference, a noise threshold T0 is introduced to obtain more stable image features. The two-dimensional phase consistency PC information of each pixel in the image is calculated by the following formula:

[0045]

[0046] Among them, w0(x,y) is the frequency diffusion weight factor, ξ is a minimum error constant to prevent the denominator from being zero, T0 is the noise suppression factor, Δφ so(x,y) is the phase deviation measured in scale and direction. so (x,y)Δφ so The expression for (x,y) is as follows:

[0047]

[0048] Where,

[0049]

[0050]

[0051] E(x,y) represents the local energy of the signal, and its expression is:

[0052]

[0053] The two-dimensional phase consistency PC information in different directions is added together to obtain a two-dimensional phase consistency overlay map, which is expressed as:

[0054]

[0055] The above operation is performed on the reference image to obtain a first two-dimensional phase congruency overlay map; correspondingly, the above operation is performed on the image to be registered, that is, the real-time image, to obtain a second two-dimensional phase congruency overlay map.

[0056] Step 2: Perform SUSAN corner point detection on the first two-dimensional phase consistency overlay image and the second two-dimensional phase consistency overlay image respectively to extract feature point information, and obtain a number of first feature points and a number of second feature points accordingly.

[0057] The SUSAN (Smallest Univalue Segment Assimilating Nucleus) operator is a highly efficient edge and angle detection operator. It not only detects image edges but also efficiently detects corner information, with good noise robustness. Furthermore, SUSAN does not involve differential operations and exhibits good illumination, rotation, and scale invariance. Therefore, this embodiment uses SUSAN corner detection to extract feature point information.

[0058] The basic idea of ​​the SUSAN corner detection algorithm is to place a circular template with a certain radius on the image. If there is an area on the template such that the grayscale value of each pixel in the corresponding image is the same (or close) to the grayscale value of the center of the circle, then this area is defined as a kernel value similarity area, or USAN. The central pixel of the circular template is called the kernel, and the number of pixels with the same grayscale value as the kernel is defined as the area of ​​the template. The USAN area is obtained after the USAN template is calculated. The smaller the USAN area, the greater the possibility that the current point is a feature point (corner point). In other words, the output image enhances the feature points, and the degree of enhancement for two-dimensional features (such as corner points) is greater than that for one-dimensional features (straight line edges).

[0059] In this embodiment, SUSAN corner point detection is performed on the first two-dimensional phase consistency overlay image and the second two-dimensional phase consistency overlay image to extract feature point information, and a plurality of first feature points and a plurality of second feature points are obtained accordingly. The steps of SUSAN corner point detection to extract feature points are as follows:

[0060] 1) Construct a circular template with the pixel to be calculated as the center of the circle.

[0061] Specifically, the circular template is usually selected in a 7×7 window, such as Figure 3 As shown, a circular template with a radius of 3.5 and 37 pixels is formed;

[0062] 2) Scan the pixels in the circular template and compare the pixels of each point with the core point; the comparison function is:

[0063]

[0064] in, is the position of the core point in the two-dimensional image, is the position of all pixels in the circular template except the core point. t represents the brightness interpolation threshold. The smaller t is, the more pixels are extracted, and the number of false positives also increases.

[0065] 3) Calculate the area of ​​the USAN region: Add the grayscale value comparison function of each pixel in the circular template and the core point pixel to obtain the area of ​​the USAN region, which is expressed as:

[0066]

[0067] in, Represents the core point in the two-dimensional image The area of ​​the USAN region at .

[0068] 4) Compare the area of ​​the USAN region with a given threshold to obtain the corner response function of the image, and use non-maximum suppression to perform SUSAN corner detection to obtain several feature points.

[0069] Specifically, Compare with the given geometric threshold g and finally get the corner response function of the image Then use non-maximum suppression to find corner points.

[0070]

[0071] Among them, the larger the threshold g value is, the flatter the features are and the more numerous they are; on the contrary, the smaller the threshold g value is, the sharper the features are and the fewer they are.

[0072] Regarding the value of the threshold g, optionally, as an implementation method, if corner points are to be extracted, then g = n max / 2; if you want to extract the edge of the image, g=3n max / 4. Here n max yes The maximum value of .

[0073] The present invention utilizes the phase information of an image to extract feature information of the image, which is not affected by grayscale changes and radiation differences between heterogeneous images, thereby obtaining feature points with strong robustness and high stability.

[0074] Step 3: Use the improved Sobel operator to calculate the gradient amplitude and angle of the reference image and the image to be registered respectively.

[0075] The Sobel operator is a discrete differential gradient operator that combines Gaussian smoothing and differential derivation to estimate the approximate gradient of an image. It performs neighborhood convolution on each pixel in the image using two 3×3 directional templates (horizontal and vertical), which detect horizontal and vertical edges, respectively. The Sobel operator also smooths noise and provides accurate edge information. Gradient information is unaffected by grayscale variations between heterogeneous images, making the calculated main direction more robust.

[0076] Based on this, this embodiment uses the Sobel operator to calculate the gradient magnitude and angle of the reference image and the image to be registered. Specifically, this embodiment uses multiple directional templates based on the Sobel operator to perform neighborhood convolution with each pixel in the image to obtain gradient information in different directions, thereby determining the calculation formulas for the gradient magnitude and gradient angle. The multiple directional templates include at least three directional templates, such as horizontal, vertical, and diagonal templates.

[0077] It should be noted that the traditional Sobel operator estimates the approximate value of the image gradient using templates in both the horizontal and vertical directions. However, in reality, the image gradient direction is unknown. The traditional Sobel operator does not consider the influence of diagonal directions on the image, resulting in deviations in the calculation results. Therefore, this embodiment improves the Sobel operator by adding templates in the 45° and 135° diagonal directions to the original two-directional templates to obtain the gradient image. This allows edge detection in multiple directions, thus addressing the limitation of only horizontal and vertical directions to a certain extent and making the calculation results more accurate.

[0078] Optionally, in this embodiment, the expressions of the horizontal template h1, the vertical template h2, the 45° diagonal template h3, and the 135° diagonal template h4 are respectively:

[0079]

[0080]

[0081] In addition, this embodiment can also select two diagonal direction templates at other angles based on the horizontal and vertical directions to obtain the gradient image.

[0082] It is understandable that the present embodiment can further improve the accuracy of calculation by continuing to increase the number of templates. However, considering the calculation efficiency and the limitation of the edge range, the templates in the above four directions are preferably used for calculation.

[0083] Specifically, let the original image be f(x,y), and the first-order derivatives of each pixel in the image in the horizontal, vertical, and two diagonal directions are D1(x,y), D2(x,y), D3(x,y), and D4(x,y), respectively. We can get:

[0084]

[0085]

[0086]

[0087]

[0088] The calculation formula of the gradient amplitude is defined as:

[0089]

[0090] It should be noted that this embodiment still only considers the horizontal and vertical templates when calculating the gradient argument, so the calculation formula for the gradient argument is:

[0091]

[0092] Step 4: Construct a gradient consistency direction histogram based on the circular neighborhood of the first feature point and the second feature point, and obtain a first direction histogram and a second direction histogram accordingly.

[0093] Specifically, first, the coordinates of the first feature point and the second feature point are respectively used as the center point, and a circular neighborhood with a certain radius is selected as the local area. Starting from zero degrees, the entire neighborhood is divided at a certain interval to obtain several small fan-shaped areas.

[0094] Then, the sum of the gradient amplitudes of each fan-shaped small area is counted, and the gradient consistency direction histograms are constructed respectively, and the first direction histogram and the second direction histogram are obtained correspondingly.

[0095] Among them, the ordinate of the gradient consistency direction histogram is the gradient amplitude after Gaussian weighting, and the abscissa is the direction information obtained after gradient amplitude angle conversion.

[0096] Step 5: Take the direction information corresponding to the main peak of the first direction histogram and the second direction histogram as the main direction of the feature point, and construct a descriptor with rotation invariance in the logarithmic polar coordinate system in combination with the gradient amplitude angle, and obtain the first descriptor and the second descriptor accordingly.

[0097] First, a descriptor support domain is constructed with each first feature point and second feature point as the center, and a circular neighborhood with a radius of R is taken on the corresponding first direction histogram and second direction histogram, and a logarithmic polar coordinate system is established based on the main direction.

[0098] Then, the circular neighborhood is divided into a set of concentric circles with {r1 = 0.25R, r2 = 0.5R, r3 = 0.75R}, the inner circle is kept unchanged, and the remaining three outer circles are divided into eight-quadrant grids to obtain 25 small areas.

[0099] Next, the sum of the gradient amplitudes of each small area is counted and represented in the form of a histogram; the dimension of the histogram is 8.

[0100] Finally, the histogram statistics of all 25 small regions are combined to obtain the first and second feature descriptors. The first and second descriptors are 25×8 dimensional, that is, each feature point descriptor has 200 dimensions, and the total feature descriptor dimension is N0×200, where N0 is the number of feature points.

[0101] Step 6: Perform similarity measurement on the first feature descriptor and the second feature descriptor and remove mismatched points to obtain the registration result of the image to be registered and the reference image.

[0102] First, the Euclidean distance is used to measure the similarity between the first feature descriptor and the second feature descriptor to obtain matching point pairs.

[0103] The specific process of using Euclidean distance to measure similarity can be implemented by referring to existing related technologies, and will not be described in detail in this embodiment.

[0104] Then, the fast sampling consistency algorithm is used to remove the mismatched points in the matching point pairs to obtain the registration result.

[0105] The Fast Sample Consensus (FSC) algorithm first requires a set of observations as input. It then selects a parameterized model for this set of observations and chooses parameters with high confidence for the model. The input data is divided into inliers and outliers, and the most appropriate model is calculated by repeatedly selecting a random subset of the data. First, a suitable model is selected for the inliers, and all unknown parameters of the model are calculated. This model is then used to test the outliers. If the data from a particular outlier also fits the model, the outlier is converted to an inlier. This process continues until a sufficient number of outliers have been converted to inliers, indicating that the selected model is appropriate. Finally, all inliers are used to estimate the model and perform error analysis to assess its accuracy. This process is repeated n times, and the model with the largest number of inliers and the highest accuracy is selected as the optimal model.

[0106] At this point, the matching between the image to be registered and the reference image is completed, and the registration result is obtained.

[0107] The image registration method proposed in the present invention first uses the phase information of the image to extract the feature information of the image, which is not affected by the grayscale changes and radiation differences between heterogeneous images, thereby obtaining feature points with strong robustness and high stability; then, for the reference image and the image to be registered, the improved Sobel operator is used to calculate the gradient amplitude and angle information, and a direction histogram is constructed in the circular neighborhood of the feature point, which overcomes the influence of SAR speckle noise interference and avoids over-reliance on grayscale characteristics; then, the gradient direction corresponding to the main peak of the gradient histogram is selected as the main direction of the feature point, and the gradient amplitude and angle and the main direction are used to obtain a feature descriptor with a high degree of consistency in the logarithmic polar coordinate system, finally, a fast sampling consistency algorithm is used to eliminate mismatched points, which reduces the root mean square error of the algorithm and improves the feature matching accuracy, making the registration of multi-source rotated images more accurate and rapid.

[0108] Example 2

[0109] The following simulation experiment is conducted on the method provided in the above embodiment 1 to verify the feasibility and beneficial effects of the present invention.

[0110] This simulation experiment is all 90-degree rotation image registration.

[0111] Experiment 1: Register the visible light image and SAR image. The results are as follows: Figure 4 As shown, Figure 4 Figure a is the connection diagram of the registration result, figure b is the fusion diagram, and figure c is the checkerboard splicing diagram.

[0112] Experiment 2: Register the visible light image and infrared image. The results are as follows: Figure 5 As shown, Figure 5 Figure a is the connection diagram of the registration result, figure b is the fusion diagram, and figure c is the checkerboard splicing diagram.

[0113] Experiment 3: SAR image and SAR image registration, the results are as follows Figure 6 As shown, Figure 6 Figure a is the connection diagram of the registration result, figure b is the fusion diagram, and figure c is the checkerboard splicing diagram.

[0114] In addition, this embodiment also simulates and compares the three pairs of images in the above three experiments using the existing radiation-variation insensitive feature transform algorithm (RIFT) and the homogeneous structure point feature algorithm of the present invention. The root mean square error (RMSE) and running time are shown in Table 1 below.

[0115] Table 1

[0116] RIFT The present invention RMSE / pixel Time / t RMSE / pixel Time / t Images from Experiment 1 1.4016 20.2269 1.3739 12.4353 Experiment 2 images 1.2747 21.2878 1.2206 10.9670 Image from Experiment 3 1.3755 7.6173 1.3641 5.7407

[0117] As can be seen from Table 1, compared with the existing RIFT algorithm, the homogeneous structure point feature algorithm of the present invention has a smaller root mean square error, better registration effect, and shorter running time, thereby verifying the feasibility and effectiveness of the algorithm proposed in the present invention.

[0118] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. An image registration method based on multi-source homogeneous structural point features, characterized in that: include: Acquire a first two-dimensional phase congruency overlay image corresponding to the reference image and a second two-dimensional phase congruency overlay image corresponding to the image to be registered; Performing SUSAN corner point detection on the first two-dimensional phase consistency overlay image and the second two-dimensional phase consistency overlay image respectively to extract feature point information, and correspondingly obtaining a plurality of first feature points and a plurality of second feature points; The improved Sobel operator is used to calculate the gradient amplitude and angle of the reference image and the image to be registered, including: Based on the Sobel operator, multiple directional templates are used to perform neighborhood convolution with each pixel in the image to obtain gradient information in different directions, and a calculation formula for the gradient amplitude and the gradient angle is determined; wherein the multiple directional templates include at least three directional templates; Constructing a gradient consistency directional histogram based on the circular neighborhood of the first feature point and the second feature point, and correspondingly obtaining a first directional histogram and a second directional histogram, including: Taking the coordinates of the first feature point and the second feature point as the center point, respectively, a circular neighborhood with a certain radius is selected as the local area, and the entire neighborhood is divided at a certain interval starting from zero degrees to obtain a number of small sector-shaped areas; Counting the sum of the gradient amplitudes of each sector-shaped small area, constructing gradient consistency direction histograms respectively, and obtaining a first direction histogram and a second direction histogram respectively; wherein the ordinate of the gradient consistency direction histogram is the gradient amplitude after Gaussian weighting, and the abscissa is the direction information obtained after gradient amplitude angle conversion; Taking the direction information corresponding to the main peak of the first directional histogram and the second directional histogram as the main direction of the feature point, and combining the gradient amplitude angle to construct a feature descriptor with rotation invariance in a logarithmic polar coordinate system, the first feature descriptor and the second feature descriptor are obtained accordingly, including: Construct the descriptor support domain with each first feature point and second feature point as the center, and take the radius as The circular neighborhood of size is used to establish a logarithmic polar coordinate system based on the main direction; The circular neighborhood is Divide it into a set of concentric circles, keep the inner ring unchanged, and divide the other three outer rings into eight quadrants to obtain 25 small areas; Count the sum of the gradient amplitudes of each small area and express it in the form of a histogram; The histogram statistics of all 25 small regions are combined to obtain the first feature descriptor and the second feature descriptor respectively; A similarity measurement is performed on the first feature descriptor and the second feature descriptor, and mismatched points are eliminated to obtain a registration result between the image to be registered and the reference image.

2. The image registration method based on multi-source homogeneous structural point features according to claim 1, characterized in that: Acquiring a first two-dimensional phase congruency overlay map corresponding to the reference image and a second two-dimensional phase congruency overlay map corresponding to the image to be registered, including: The phase consistency PC information at different angles and scales corresponding to the pixel points at each position in the reference image and the image to be registered is obtained through the Log-Gabor filter, and superimposed to obtain the first two-dimensional phase consistency superposition map and the second two-dimensional phase consistency superposition map.

3. The image registration method based on multi-source homogeneous structural point features according to claim 1, characterized in that: SUSAN corner point detection is performed on the first two-dimensional phase consistency overlay image and the second two-dimensional phase consistency overlay image respectively to extract feature point information, and a plurality of first feature points and a plurality of second feature points are correspondingly obtained, including: The following processing is performed on the first two-dimensional phase consistency overlay image and the second two-dimensional phase consistency overlay image, respectively, to obtain a plurality of first characteristic points and a plurality of second characteristic points accordingly: Construct a circular template with the pixel to be calculated as the center of the circle; Scanning pixel points in the circular template and comparing the pixels of each point with the core point; Add the gray value comparison function of each pixel in the circular template and the core point pixel to obtain the area of ​​the USAN region; The area of ​​the USAN region is compared with a given threshold to obtain the corner response function of the image, and non-maximum suppression is used to perform SUSAN corner detection to obtain several feature points.

4. The image registration method based on multi-source homogeneous structural point features according to claim 1, characterized in that: The multiple direction templates include a horizontal direction template, a vertical direction template, a 45° diagonal direction template and a 135° diagonal direction template.

5. The image registration method based on multi-source homogeneous structural point features according to claim 4, characterized in that: The expressions of the horizontal template, vertical template, 45° diagonal template and 135° diagonal template are respectively: ； ； ； ； in, 、 、 、 They represent the horizontal template, vertical template, 45° diagonal template, and 135° diagonal template respectively.

6. The image registration method based on multi-source homogeneous structural point features according to claim 5, characterized in that: The calculation formulas for the gradient amplitude and the gradient angle are: ； ； in, represents the gradient amplitude, 、 、 、 Represent the first-order derivatives of each pixel in the image in the horizontal, vertical, 45° diagonal, and 135° diagonal directions, respectively, and ； ； ； ； Represents the input reference image and the image to be registered; represents the gradient argument.

7. The image registration method based on multi-source homogeneous structural point features according to claim 1, characterized in that: Performing similarity measurement on the first feature descriptor and the second feature descriptor and eliminating mismatching points to obtain a registration result between the image to be registered and the reference image, including: Using Euclidean distance to measure the similarity between the first feature descriptor and the second feature descriptor to obtain a matching point pair; A fast sampling consistency algorithm is used to eliminate mismatched points in the matching point pairs to obtain a registration result.

Images