Image matching method based on image feature similarity analysis and difference elimination

Abstract

The invention discloses an image matching method based on image feature similarity analysis and difference elimination, and the method comprises the steps: carrying out the gray conversion of an obtained image, carrying out the noise reduction through a specific formula, and adjusting a filtering standard deviation according to a contrast ratio; feature points are extracted and screened by using an SIFT algorithm, and PCA dimension reduction can be carried out; then calculating the Euclidean distance of the feature points, and optimizing the preliminary matching point pairs by using RANSAC (Random Sample Consensus); and finally, analyzing the position and gray difference of the matching point, and correcting the image through affine transformation to realize accurate matching. In each step, various calculation and processing operations, such as gradient information calculation, weighted average and the like, are involved, so that the matching effect is improved. According to the invention, through multi-step processing, image interference factors are effectively dealt with. The method can improve the matching accuracy, enhance the adaptability and robustness, improve the calculation efficiency, accurately correct the image, provide a reliable and efficient solution for image matching in various fields, and promote the development of related technologies.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis and processing, and in particular to an image matching method based on image feature similarity analysis and difference elimination. Background Art

[0002] Image matching technology plays a key role in many fields, such as computer vision, medical image analysis, remote sensing image processing, intelligent security monitoring, etc. With the continuous advancement of science and technology, higher and higher requirements are placed on the accuracy, efficiency and robustness of image matching.

[0003] In the field of computer vision, image matching is the basis for tasks such as target recognition, 3D reconstruction, and image stitching. For example, in autonomous driving technology, vehicles need to accurately identify targets such as road signs, pedestrians, and other vehicles through image matching in order to make correct driving decisions. In medical image analysis, matching medical images of different periods or different modalities (such as CT, MRI, etc.) helps doctors diagnose diseases more accurately, monitor the progression of the disease, and evaluate the effectiveness of treatment.

[0004] However, existing image matching methods face many challenges. On the one hand, images are easily disturbed by various factors during the acquisition process, such as changes in illumination, perspective, occlusion, noise, etc. These factors can cause changes in image features, increasing the difficulty of accurate matching. For example, in outdoor scenes, images taken at different times may have obvious differences in color and brightness features of the same object due to different light intensities and angles; in complex scenes, the target object may be partially occluded, resulting in the loss of some feature information.

[0005] On the other hand, traditional image matching methods often have problems of low efficiency and insufficient accuracy when processing large-scale image data or scenarios with high matching accuracy requirements. Although some matching methods based on feature points can extract key features of images to a certain extent, they are not accurate enough in feature description and similarity measurement, and are prone to false matches. Moreover, for images with large differences, existing methods may not be able to effectively eliminate the differences, thus affecting the matching effect.

[0006] Therefore, there is an urgent need for a new image matching method that can better cope with various interference factors and improve the accuracy and efficiency of matching to meet the ever-evolving scientific and technological needs. The image matching method based on image feature similarity analysis and difference elimination proposed in this invention aims to solve the above problems existing in the prior art and provide a more reliable and efficient solution for the application of image matching technology in various fields. Summary of the invention

[0007] The present invention proposes an image matching method based on image feature similarity analysis and difference elimination to solve the problems mentioned in the above-mentioned prior art.

[0008] In order to achieve the above object, the present invention adopts the following technical solution: an image matching method based on image feature similarity analysis and difference elimination, comprising the following steps:

[0009] Image preprocessing steps:

[0010] S1: Get two images to be matched, and convert them into grayscale images. The grayscale conversion formula is Gray = 0.299R + 0.587G + 0.114B, where Gray is the grayscale value, and R, G, and B are the red, green, and blue color components of the image pixels respectively;

[0011] S2: Perform noise reduction on the grayscale image using a Gaussian filter algorithm with a filter kernel size of (3×3)-(5×5) pixels and a standard deviation of 0.5-1.5. The formula is: (where G(x, y) is the value of the Gaussian filter function at the point (x, y), and σ is the standard deviation);

[0012] Feature extraction steps:

[0013] S3: Use the scale-invariant feature transform (SIFT) algorithm to extract image feature points, detect extreme points in multi-scale space, determine the location and scale information of feature points, and generate feature point descriptors. The descriptors are 128-dimensional and are calculated based on the gradient information of the pixels in the neighborhood of the feature points. The formula is:

[0014]

[0015] (where m(x, y) is the gradient amplitude at the pixel point (x, y), and L(x, y) is the grayscale value of the point). Where θ(x, y) is the gradient direction at the pixel point (x, y);

[0016] S4: Screening the extracted feature points according to the response values ​​of the feature points, retaining the feature points whose response values ​​are greater than a preset threshold, and removing unstable and weak feature points;

[0017] Similarity analysis steps:

[0018] S5: Calculate the Euclidean distance between the feature points of the two images. The formula is: Among them, p and q are the descriptors of two feature points. i ,q i is the th dimension component in the descriptor, and the feature point pairs whose distance is less than the preset similarity threshold are regarded as preliminary matching point pairs;

[0019] S6: Use the random sampling consensus RANSAC algorithm to optimize the preliminary matching point pairs, with the number of iterations being 100-1000. By calculating the number of inliers in the model, the model with the largest number of inliers is selected as the optimal model.

[0020] Steps to eliminate differences:

[0021] S7: Analyze the position and grayscale differences of the successfully matched feature point pairs, and calculate the mean μ of the position differences x and μ y , where the mean of the position difference in the x and y directions and the mean of the grayscale difference μ g , the formula is in is the x-coordinate of the i-th pair of matching points in the two images, n is the number of matching point pairs, in is the y coordinate of the i-th pair of matching points in the two images, in is the gray value of the first pair of matching points in the two images;

[0022] S8: Correct the image according to the mean of position and grayscale difference, using the affine transformation model, and the transformation matrix is: Among them, a, b, c, d, e, and f are the parameters to be determined. The transformation matrix is ​​solved by the least squares method. The formula is: in are the coordinates of the point in the source image, For the coordinates of the corresponding points in the target image, transform one of the images to eliminate the difference.

[0023] Furthermore, in the image preprocessing step, the standard deviation of the Gaussian filter is adaptively adjusted according to the contrast of the image. If the contrast is low, the standard deviation is appropriately reduced. If the contrast is high, the standard deviation is appropriately increased. The specific adjustment method is: first calculate the grayscale histogram of the image, count the distribution range of the grayscale value, set a contrast threshold T, set it according to the empirical value, such as T = 50, if rabge < T, it is considered that the contrast is low, and the standard deviation is adjusted to If range>T, the contrast is considered high and the standard deviation is adjusted to

[0024] Furthermore, in the feature extraction step, after the feature points are extracted using the SIFT algorithm, the principal component analysis (PCA) algorithm is further used to reduce the dimension of the feature point descriptors. The specific operation is: all the extracted feature point descriptors are combined into a matrix M, and the covariance matrix C = M is calculated. TM, solve the eigenvalues ​​and eigenvectors of the covariance C matrix, sort them from large to small according to the eigenvalue size, select the first k eigenvectors with a value range of 32-64 dimensions to form a projection matrix P, project the feature point descriptor matrix onto the projection matrix P, and obtain the reduced-dimensional feature point descriptor matrix M′=MP, thereby reducing the amount of data.

[0025] Furthermore, in the similarity analysis step, before calculating the Euclidean distance, the feature point descriptor is normalized so that the descriptor has unit length. The specific normalization formula is: Where p′ is the normalized descriptor and p is the original descriptor. The normalization process is as follows: for each feature point descriptor p, calculate its length Then each component in the descriptor is divided by to obtain the normalized descriptor p′.

[0026] Furthermore, in the difference elimination step, when calculating the mean of the position and grayscale differences, the weighted average method is used to assign different weights according to the response value of the feature point. The higher the response value, the greater the weight. The specific calculation method is: let the response value of the i-th feature point be r i , first normalize the response value to get the normalized response value where r min and r max are the minimum and maximum values ​​of all feature point response values, respectively, and then calculate the weight w i =0.5+r′ i ×(1.5-0.5). Weighted mean of position differences and and the weighted mean of the grayscale differences The calculation formulas are Make the difference calculation more consistent with the importance of feature points.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] Improve matching accuracy: This method uses precise image preprocessing to reduce noise and color interference, uses the SIFT algorithm to accurately extract and filter feature points, and combines normalization processing and RANSAC optimization to effectively remove incorrect matching point pairs, thereby improving matching accuracy. In complex scenes, such as when there is occlusion or lighting changes, feature points can still be accurately identified and matched, providing a reliable foundation for subsequent applications.

[0029] Enhanced adaptability and robustness: It can adapt to images with different contrasts, adaptively adjust the Gaussian filter standard deviation according to the contrast, and use weighted average when calculating the difference mean, taking into account the response value of the feature point, so that the method has stronger adaptability and robustness to lighting changes, partial occlusion, etc., ensuring stable operation in various practical application scenarios.

[0030] Improve computing efficiency: The PCA algorithm is used to reduce the dimension of feature point descriptors, reduce the amount of data, and reduce the computational complexity. It has obvious advantages when processing large-scale image data, can quickly complete matching tasks, and improve the real-time performance of the system. It is suitable for application scenarios with high time requirements, such as real-time monitoring and autonomous driving.

[0031] Achieve precise correction and optimize matching results: By analyzing the position and grayscale differences of matching point pairs, the affine transformation model is used to correct the image, eliminate the differences, and make the matched images more accurately aligned. The matching results are further optimized to meet the application requirements with high matching accuracy, such as medical image analysis, high-precision mapping and other fields.

[0032] Providing comprehensive and effective image matching solutions: Comprehensively considering multiple key links in the image matching process, from preprocessing to feature extraction, similarity analysis to difference elimination, each step works together to form a complete method system, providing comprehensive and effective solutions for image matching problems in different fields, and promoting the widespread application and development of image matching technology in multiple fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic block diagram of an image matching system for image feature similarity analysis and difference elimination proposed by the present invention;

[0034] Figure 2 The present invention provides a schematic block diagram of an image matching method for image feature similarity analysis and difference elimination. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0037] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0038] Reference Figure 1-2 :An image matching method based on image feature similarity analysis and difference elimination, comprising the following steps:

[0039] Image preprocessing steps:

[0040] S1: Get two images to be matched, convert the images into grayscale images, and the grayscale conversion formula is Gray = 0.299R + 0.587G + 0.114B (where Gray is the grayscale value, R, G, and B are the red, green, and blue color components of the image pixel points, respectively), so as to reduce the interference of color information on feature extraction;

[0041] S2: Perform noise reduction on the grayscale image using a Gaussian filter algorithm with a filter kernel size of (3×3)-(5×5) pixels and a standard deviation of 0.5-1.5. The formula is: (where G(x, y) is the value of the Gaussian filter function at the point (x, y), and σ is the standard deviation), which removes noise points in the image and improves image quality.

[0042] Feature extraction steps:

[0043] S3: Use the scale-invariant feature transform (SIFT) algorithm to extract image feature points, detect extreme points in multi-scale space, determine the location and scale information of feature points, and generate feature point descriptors. The descriptors are 128 dimensions and are calculated based on the gradient information of the pixels in the neighborhood of the feature points. The formula is:

[0044] (where m(x, y) is the gradient amplitude at the pixel point (x, y), and L(x, y) is the grayscale value of the point). (where θ(x, y) is the gradient direction at the pixel point (x, y));

[0045] S4: Screen the extracted feature points according to their response values, retain feature points whose response values ​​are greater than a preset threshold (the threshold is set according to the distribution of image features and ranges from 0.01 to 0.1), remove unstable and weak feature points, and improve the quality and representativeness of the feature points.

[0046] Similarity analysis steps:

[0047] S5: Calculate the Euclidean distance between the feature points of the two images. The formula is: (where p and q are the descriptors of two feature points, p i ,q i is the dimensional component in the descriptor), and the feature point pairs whose distance is less than the preset similarity threshold (set according to the similarity of image features, ranging from 0.5 to 1.5) are regarded as preliminary matching point pairs;

[0048] S6: The random sampling consensus (RANSAC) algorithm is used to optimize the preliminary matching point pairs, with the number of iterations being 100-1000 times. The model with the largest number of inliers is selected as the optimal model by calculating the number of inliers in the model (inliers are points that meet the model assumptions), and the incorrect matching point pairs are removed to improve the matching accuracy.

[0049] Steps to eliminate differences:

[0050] S7: For the successfully matched feature point pairs, analyze their position and grayscale differences and calculate the mean μ of the position differences x and μ y (mean of position differences in x and y directions) and mean of grayscale differences μ g , the formula is (in is the x-coordinate of the i-th pair of matching points in the two images, and n is the number of matching point pairs). (in is the y coordinate of the i-th pair of matching points in the two images), (in is the gray value of the first pair of matching points in the two images);

[0051] S8: Correct the image according to the mean of position and grayscale difference, using the affine transformation model, and the transformation matrix is: (where a, b, c, d, e, and f are the parameters to be determined), the transformation matrix is ​​solved by the least squares method, and the formula is: (in are the coordinates of the point in the source image, is the coordinate of the corresponding point in the target image), transform one of the images to eliminate the difference and achieve accurate matching.

[0052] In the present invention, in the image preprocessing step, the standard deviation of the Gaussian filter can also be adaptively adjusted according to the contrast of the image. If the contrast is low (by calculating the grayscale histogram of the image and counting the grayscale value distribution range, if the grayscale values ​​are mainly concentrated in a narrow interval, it is judged that the contrast is low), the standard deviation is appropriately reduced (in the range of 0.3-0.8) to enhance the image details; if the contrast is high (the grayscale value distribution range is wide), the standard deviation is appropriately increased (in the range of 1.2-2.0). The specific adjustment method is: first calculate the grayscale histogram of the image, count the grayscale value distribution range range, set a contrast threshold T (set according to the empirical value, such as T = 50), if range < T, it is considered that the contrast is low, and the standard deviation is adjusted to If range>T, the contrast is considered high and the standard deviation is adjusted to

[0053] In the present invention, in the feature extraction step, after the feature points are extracted using the SIFT algorithm, the principal component analysis (PCA) algorithm can be further used to reduce the dimension of the feature point descriptors. The specific operation is: all the extracted feature point descriptors are combined into a matrix M, and the covariance matrix C of the matrix is ​​calculated as T M, solve the eigenvalues ​​and eigenvectors of the covariance C matrix, sort them from large to small according to the eigenvalue size, select the first k (k ranges from 32-64 dimensions) eigenvectors to form the projection matrix P, project the feature point descriptor matrix onto the projection matrix P, and obtain the reduced dimensionality feature point descriptor matrix M′=MP, thereby reducing the amount of data, improving computational efficiency, and maintaining the representativeness of the features.

[0054] In the present invention, in the similarity analysis step, before calculating the Euclidean distance, the feature point descriptor may be normalized so that the descriptor has a unit length. The specific normalization formula is: (where p′ is the normalized descriptor and p is the original descriptor) The normalization process is: for each feature point descriptor p, calculate its length Then each component in the descriptor is divided by to obtain the normalized descriptor p′, which improves the accuracy of similarity judgment.

[0055] In the present invention, in the difference elimination step, when calculating the mean of the position and grayscale differences, a weighted average method can be used to assign different weights according to the response value of the feature point. The higher the response value, the greater the weight (the weight range is 0.5-1.5). The specific calculation method is: let the response value of the i-th feature point be r i , first normalize the response value to get the normalized response value where r min and r max are the minimum and maximum values ​​of all feature point response values ​​respectively), and then calculate the weight w i =0.5+r′ i ×(1.5-0.5). Weighted mean of position differences and and the weighted mean of the grayscale differences The calculation formulas are Make the difference calculation more consistent with the importance of feature points and improve the accuracy of correction.

[0056] In actual application scenarios, such as the field of autonomous driving, the camera continuously collects road images while the vehicle is driving. First, the system enters the image preprocessing step, and the system instantly converts the acquired color road image into a grayscale image, and quickly analyzes the image contrast. If the vehicle is traveling through a tunnel with alternating light and dark, the image preprocessing module determines that the contrast is low, and then adaptively adjusts the Gaussian filter standard deviation to 0.5, effectively removing the noise caused by light changes, laying a solid foundation for subsequent accurate matching.

[0057] When extracting features, the SIFT algorithm is busy "searching" in the multi-scale space of complex road scenes, accurately locating key feature points such as road signs and other vehicle outlines. Assuming that the vehicle ahead is partially blocked by a billboard, the feature point response value is screened to decisively eliminate those feature points affected by the blockage and with poor stability, and only retain the strong response feature points. If the amount of data is too large, the PCA algorithm quickly intervenes to reduce the dimension of the feature point descriptor to ensure that the matching process can be quickly promoted even with limited on-board computing resources.

[0058] When it comes to similarity analysis, facing a large number of feature point descriptors, they are first normalized in an orderly manner. For example, in a congested city, the image features of many vehicles and pedestrians are mixed, and the normalized descriptors make the Euclidean distance calculation more accurate. After hundreds of iterations, the RANSAC algorithm "eliminates the false and retains the true" in the image information of the busy traffic, screens out wrong matches, and provides strong support for the vehicle to accurately identify surrounding objects and make correct driving decisions, ensuring safe and smooth driving.

[0059] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An image matching method based on image feature similarity analysis and difference elimination, characterized in that: The following steps are involved: Image preprocessing steps: S1: Get two images to be matched, and convert them into grayscale images. The grayscale conversion formula is Gray = 0.299R + 0.587G + 0.114B, where Gray is the grayscale value, and R, G, and B are the red, green, and blue color components of the image pixels respectively; S2: Perform noise reduction on the grayscale image using a Gaussian filter algorithm with a filter kernel size of (3×3)-(5×5) pixels and a standard deviation of 0.5-1.

5. The formula is: Where G(x, y) is the value of the Gaussian filter function at the point (x, y), and σ is the standard deviation; Feature extraction steps: S3: Use the scale-invariant feature transform (SIFT) algorithm to extract image feature points, detect extreme points in multi-scale space, determine the location and scale information of feature points, and generate feature point descriptors. The descriptors are 128-dimensional and are calculated based on the gradient information of the pixels in the neighborhood of the feature points. The formula is: , where m(x, y) is the gradient amplitude at the pixel point (x, y), L(x, y) is the gray value of the point, Where θ(x, y) is the gradient direction at the pixel point (x, y); S4: Screening the extracted feature points according to the response values ​​of the feature points, retaining the feature points whose response values ​​are greater than a preset threshold, and removing unstable and weak feature points; Similarity analysis steps: S5: Calculate the Euclidean distance between the feature points of the two images. The formula is: Among them, p and q are the descriptors of two feature points. i ,q i is the i-th dimension component in the descriptor, and the feature point pairs whose distance is less than the preset similarity threshold are regarded as preliminary matching point pairs; S6: Use the random sampling consensus RANSAC algorithm to optimize the preliminary matching point pairs, with the number of iterations being 100-1000. By calculating the number of inliers in the model, the model with the largest number of inliers is selected as the optimal model. Steps to eliminate differences: S7: Analyze the position and grayscale differences of the successfully matched feature point pairs, and calculate the mean μ of the position differences x and μ y , where the mean of the position difference in the x and y directions and the mean of the grayscale difference μ g , the formula is in is the x-coordinate of the i-th pair of matching points in the two images, n is the number of matching point pairs, in is the y coordinate of the i-th pair of matching points in the two images, in is the gray value of the i-th pair of matching points in the two images; S8: Correct the image according to the mean of position and grayscale difference, using the affine transformation model, and the transformation matrix is: Among them, a, b, c, d, e, and f are the parameters to be determined. The transformation matrix is ​​solved by the least squares method. The formula is: in are the coordinates of the point in the source image, For the coordinates of the corresponding points in the target image, transform one of the images to eliminate the difference.

2. The image matching method based on image feature similarity analysis and difference elimination according to claim 1, characterized in that: In the image preprocessing step, the standard deviation of the Gaussian filter is adaptively adjusted according to the contrast of the image. If the contrast is low, the standard deviation is reduced, and if the contrast is high, the standard deviation is increased. The specific adjustment method is: first calculate the grayscale histogram of the image, count the distribution range of the grayscale value range, set a contrast threshold T, set it according to the empirical value, such as T = 50, if range < T, it is considered that the contrast is low, and the standard deviation is adjusted to If range>T, the contrast is considered high and the standard deviation is adjusted to 3. The image matching method based on image feature similarity analysis and difference elimination according to claim 1, characterized in that: In the feature extraction step, after using the SIFT algorithm to extract feature points, the principal component analysis PCA algorithm is further used to reduce the dimension of the feature point descriptors. The specific operation is: all the extracted feature point descriptors are combined into a matrix M, and the covariance matrix C = M is calculated. T M, solve the eigenvalues ​​and eigenvectors of the covariance C matrix, sort them from large to small according to the eigenvalue size, select the first k eigenvectors to form the projection matrix P, where k ranges from 32 to 64 dimensions, and project the feature point descriptor matrix onto the projection matrix P to obtain the reduced dimensionality feature point descriptor matrix M′=MP.

4. The image matching method based on image feature similarity analysis and difference elimination according to claim 1, characterized in that: In the similarity analysis step, before calculating the Euclidean distance, the feature point descriptor is normalized to make the descriptor have unit length. The specific normalization formula is: Where p′ is the normalized descriptor and p is the original descriptor. The normalization process is as follows: for each feature point descriptor p, calculate its length Then each component in the descriptor is divided by to obtain the normalized descriptor p′.

5. The image matching method based on image feature similarity analysis and difference elimination according to claim 1, characterized in that: In the difference elimination step, when calculating the mean of position and grayscale differences, the weighted average method is used to assign different weights according to the response value of the feature point. The higher the response value, the greater the weight. The specific calculation method is: let the response value of the i-th feature point be r i , first normalize the response value to get the normalized response value where r min and r max are the minimum and maximum values ​​of all feature point response values, respectively, and then calculate the weight w i =0.5+r′ i ×(1.5-0.5). Weighted mean of position differences and and the weighted mean of the grayscale differences The calculation formulas are Make the difference calculation more consistent with the importance of feature points.

Images