Non-rigid image registration method and system based on rotation invariant feature flow

Abstract

The invention discloses a non-rigid image registration method based on rotation invariant feature flow, which mainly solves the problem that non-rigid distorted images with larger rotation angles cannot be registered in the prior art, and adopts the scheme as follows: the pixel gradient direction, the amplitude, the image gradient direction and the image gradient direction of any input image pair are calculated; a plurality of candidate directions are obtained through direction neighborhood amplitude weighting, and a plurality of main directions are screened; performing rotation normalization in a pixel neighborhood based on the plurality of main directions to obtain feature vectors in a plurality of unified directions, and generating descriptor vectors through splicing; calculating various parameters such as data matching, displacement constraint, flow field smoothness, direction consistency constraint and the like according to the descriptor vector to obtain a characteristic flow objective function; and iteratively optimizing the target function to calculate a displacement field meeting a multi-constraint condition, and moving the pixel points of the source image to corresponding positions to obtain a registration image aligned with the target image. According to the method, high-precision and robust registration of non-rigid distortion images with large rotation angles can be achieved, and the method can be used for complex deformation scenes such as remote sensing monitoring and medical image analysis.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and remote sensing image analysis, and specifically relates to a non-rigid image registration method and system, which can be used for pixel-level alignment in complex deformation scenarios such as remote sensing monitoring and medical image analysis. Background Art

[0002] Image registration is one of the core tasks in computer vision and remote sensing image analysis, and is widely used in many fields such as remote sensing monitoring, medical image analysis, target tracking, and autonomous driving. Its basic goal is to establish a spatial correspondence relationship between images acquired under different times, different perspectives, or different sensor conditions, so that they are aligned in the same coordinate system, thereby providing support for subsequent tasks such as image analysis, target recognition, and scene reconstruction.

[0003] Existing image registration methods can generally be divided into two categories: traditional handcrafted feature-based methods and deep learning-based methods. Among them:

[0004] Traditional image registration methods mainly rely on manually designed feature extraction algorithms to achieve image matching by detecting key points and constructing feature descriptors. For example, the patent document with the application number CN202310776833.6 discloses an "image registration method based on multi-source homogeneous structure point features". This method first obtains features by extracting image phase information, calculates the gradient magnitude and angle of the reference image and the image to be registered respectively using an improved Sobel operator, and constructs a direction histogram within the circular neighborhood of the feature points; then selects the direction corresponding to the main peak of the gradient histogram as the main direction of the feature points, and combines the gradient magnitude, angle, and main direction to generate a highly consistent feature descriptor in the log-polar coordinate system; finally, efficiently removes mismatched points through the random sample consensus algorithm. Since this method needs to construct a registration model by sparse feature point sampling, its accuracy is limited by the randomness of the local key point distribution and the accuracy of feature matching, and is easily affected by noise interference or feature loss, resulting in deviation in model estimation and insufficient stability.

[0005] Deep learning-based image registration technology shows significant potential in performance optimization due to its powerful feature extraction ability and end-to-end learning advantages. For example, the patent document with the application number CN201611034020.6 discloses a "non-rigid image registration method based on image features and TV-L1". It uses the SIFT method or the SURF method to extract feature points from the floating image and the reference image, and performs rough registration using the extracted feature points; then, on the basis of rough registration, uses the TV-L1 method based on the optical flow field for fine registration. Although this method has relatively high accuracy, it has the following two deficiencies:

[0006] First, since it adopts a hierarchical and progressive multi-stage optimization strategy for coarse registration and then fine registration, it relies on sparse feature point matching in the coarse registration stage. Moreover, in the case of uneven feature distribution or the existence of mismatches, initial pose deviations are likely to occur. This kind of error will form an accumulation effect in the fine registration stage, leading to the deformation field optimization falling into a local sub-optimal solution.

[0007] Second, because it uses the traditional TV-L1 model, there are modeling limitations when dealing with large-angle rotational transformations, and multiple resamplings and iterative adjustments are required for compensation. This not only significantly increases the computational complexity but also affects the stability of the registration process. Summary of the Invention

[0008] The purpose of the present invention is to propose a non-rigid image registration method based on rotation-invariant feature flow in view of the above-mentioned deficiencies of the existing technologies, so as to achieve image registration in the case of complex distortions and improve the accuracy, stability, and robustness of the registration.

[0009] The technical idea to achieve the above purpose is as follows: By adopting rotation-invariant feature flow to handle large-scale transformations, rotations, and non-rigid effects brought about by complex distortions; and through feature flow optimization technology, it is ensured that the registration process is pixel-level registration, guaranteeing the registration accuracy, stability, and robustness.

[0010] According to the above idea, the technical solution of the present invention includes:

[0011] 1. A non-rigid image registration method based on rotation-invariant feature flow, characterized by comprising:

[0012] (1) Calculate the pixel gradient direction θ i，j and amplitude G for any pair of input images, obtain multiple candidate directions through neighborhood amplitude weighting of the direction, and sort and screen these candidate directions to select multiple main directions;

[0013] (2) Based on multiple main directions, perform rotation normalization within the neighborhood of key points to obtain multiple feature vectors with unified directions, and splice multiple feature vectors to generate a descriptor vector;

[0014] (3) Calculate the data matching term, displacement constraint term, flow field smoothing term, and direction consistency constraint term according to the descriptor vector, and add the terms to obtain a feature flow objective function;

[0015] (4) Based on the constructed objective function, calculate the displacement field that satisfies multiple constraint conditions through an iterative optimization algorithm;

[0016] (5) Use the displacement field to determine the position of each pixel, and move the pixel points in the source image to the corresponding positions to obtain a registered image aligned with the target image.

[0017] Further, based on multiple principal directions, rotation normalization is performed within the neighborhood of key points, and its implementation includes the following:

[0018] With each pixel as the center, a neighborhood window of size 3t×3t is determined, and the neighborhood window is divided into k directions with the center point as the axis, where t is a hyperparameter for controlling the window size set by the user;

[0019] Statistically, within the neighborhood window, the proportion of the gradient of the pixel in each direction is obtained, and a one-dimensional vector n of size k×t×t is obtained in each direction to i describe the local features of the pixel;

[0020] Rotate the neighborhood window according to each direction of each pixel so that all directions of each pixel are aligned to the same reference direction.

[0021] 2. A non-rigid image registration system based on a rotation-invariant feature flow, characterized by comprising:

[0022] A principal direction screening module for calculating the gradient direction and amplitude of the input image, obtaining multiple candidate directions by weighted neighborhood amplitudes in the same direction, and sorting and screening out multiple principal directions;

[0023] A descriptor vector generation module for rotating and normalizing and splicing one-dimensional feature vectors of multiple principal directions to obtain a descriptor vector for each pixel point;

[0024] A feature flow objective function calculation module for calculating a data matching term, a displacement constraint term, a flow field smoothing term, and a direction consistency constraint term respectively, and then adding the terms to obtain a feature flow objective function;

[0025] A feature flow optimization module for calculating the feature flow objective function for the descriptor and iteratively solving for the displacement field that minimizes the objective function;

[0026] A displacement field mapping module for using the displacement field to determine the destination position coordinates of each pixel in the source image to move the pixel points in the source image to the corresponding positions in the target image to complete the alignment of the registered images.

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

[0028] Firstly, since the present invention uses a principal direction screening module, it can extract multiple principal directions that fully express image features, so that the image features can be described from multiple angles and accurately. At the same time, through several optimal principal directions selected, the data dimension can be effectively reduced and computing resources can be saved.

[0029] Second, in the present invention, since the descriptor vector generation module is used to rotate, normalize and splice the one-dimensional feature vectors of multiple principal directions to obtain the descriptor vector of each pixel point, it can not only effectively eliminate the influence of image rotation on feature extraction, making the descriptor more robust to angle changes, but also simplify the data structure while ensuring the description accuracy, which helps to reduce the computational resource consumption of subsequent processing;

[0030] At the same time, since the feature information of multiple principal directions can achieve multi-angle and omni-directional feature fusion after splicing, the discrimination ability of the descriptor is significantly improved.

[0031] Third, in the present invention, since the direction consistency constraint term is added to the feature flow objective function calculation module, it can not only ensure the stable consistency of feature description in each direction, but also effectively reduce the direction deviation caused by rotation or perspective change, ensuring that the feature description is more stable and reliable, and significantly improving the anti-interference ability of the system in multi-angle and complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is the implementation flowchart of the non-rigid image registration method provided in Embodiment 1 of the present invention;

[0033] Figure 2 is the structural block diagram of the non-rigid image registration system provided in Embodiment 2 of the present invention;

[0034] Figure 3 is the registration simulation result diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The embodiments and effects of the present invention will be further described in detail below with reference to the drawings.

[0036] Embodiment 1, a non-rigid image registration method based on rotation-invariant feature flow

[0037] The input of the present invention is a pair of images to be registered, including a source image and a target image. The registration process is to register the source image and output a registered source image to make it as close as possible to the target image.

[0038] Referring to Figure 1 , the implementation steps of this example are as follows:

[0039] Step 1, calculate the direction and magnitude of the gradient.

[0040] Calculate the pixel gradient direction θ i，j and magnitude G for any pair of input source image and target image. The formulas are as follows:

[0041]

[0042] G = (Gi , G j )

[0043] where G i is the horizontal direction of the pixel value, and G j is the vertical component of the pixel value.

[0044] Step 2: Weight the directions in the neighborhood to determine all the principal directions of each pixel.

[0045] 2.1) Determine a neighborhood R centered on each pixel (x, y) of the image. Divide the neighborhood window into several candidate directions with the center point as the axis. For a certain candidate direction θ k calculate the magnitude D k of each direction θ x，y (θ k ) of each pixel (x, y) in the image:

[0046]

[0047] where (i, j) represents other pixel points appearing in the neighborhood, i represents the abscissa of the pixel, and j represents the ordinate of the pixel.

[0048] 2.2) Find a direction θ k in all the candidate directions θ * such that its magnitude D x，y (θ * ) satisfies the following condition:

[0049] D x，y (θ * ) ≥ D x，y (θ k ).

[0050] 2.3) Find all the local extreme directions θ k in all the candidate directions θ x such that the magnitude D x (θ x，y ) of θ x satisfies the following condition:

[0051]

[0052] 2.4) The user sets a threshold coefficient β, and takes the candidate directions with magnitudes greater than βD x，y (θ * ) and being local extremes among all the candidate directions as the principal directions.

[0053] Step 3: Select several principal directions from all the principal directions, normalize them, and generate a descriptor vector.

[0054] 3.1) Sort all the principal directions θ k in descending order according to D x，y (θ k ) to obtain the first principal direction θ1, the second principal direction θ2, …… the nth principal direction θ n of each pixel, where n is a hyperparameter set by the user;

[0055] 3.2) Select the first e principal directions according to the sorting result, where e is a hyperparameter set by the user;

[0056] 3.3) Taking each pixel as the center, determine a neighborhood window of size 3t×3t. Rotate the neighborhood window according to each direction of each pixel so that all directions of each pixel are aligned to the same reference direction to complete the rotation normalization;

[0057] 3.4) Divide the neighborhood window into k directions with the center point of the neighborhood window as the axis, and count the proportion of the pixel gradient in each direction within the neighborhood window. Obtain a one-dimensional vector of size k×t×t in each direction to describe the local features of the pixel, where t is a hyperparameter for controlling the window size set by the user;

[0058] 3.5) Concatenate the one-dimensional vectors of the first e principal directions of each pixel point in sequence from head to tail to obtain a descriptor vector of size e×k×t×t for each pixel point, where k is the number of directions into which the neighborhood window is divided, and t and e are hyperparameters for controlling the window size set by the user.

[0059] Step 4, set the feature flow objective function.

[0060] The present invention needs to obtain the displacement field and perform registration according to the displacement field. Only by optimizing the feature flow objective function can the displacement field be obtained. In this step, the calculation formula of the feature flow objective function is set, and its implementation includes the following:

[0061] 4.1) Calculate the data matching term L1(w):

[0062]

[0063] where p is the pixel position, w is the displacement vector, S1(p) represents the rotation-invariant feature descriptor extracted at the pixel p in the first image, and S2(p+w) represents the rotation-invariant feature descriptor extracted at the pixel p+w in the second image;

[0064] 4.2) Calculate the displacement constraint term L2(w):

[0065]

[0066] where \(u(p)\) is the horizontal displacement of the displacement vector \(w\), and \(v(p)\) is the vertical displacement of the displacement vector \(w\);

[0067] 4.3) Calculate the flow field smoothing term \(L3(w)\):

[0068]

[0069] where \(\alpha\) and \(d\) are two hyperparameters with different values, and \(v\) is the spatial neighborhood of the pixel;

[0070] 4.4) Calculate the direction consistency constraint term \(L4(w)\):

[0071]

[0072] 4.5) Add the above four terms to obtain the feature flow objective function \(E(w)\):

[0073] \(E(w)=L1(w)+L2(w)+L3(w)+L4(w)\).

[0074] Step 5, use the objective function to obtain the optimal displacement field.

[0075] To achieve the precise registration of the source image and the target image, it is necessary to optimize the objective function to find the optimal displacement field \(w\) * , so an iterative method needs to be used to optimize the objective function \(E(w)\) to obtain the displacement field \(w\) that minimizes the objective function \(E(w)\) * . Its implementation includes the following:

[0076] (5.1) Initialize the displacement field: Set the initial displacement field \(w\) as a zero vector, that is, all pixels are initially not displaced;

[0077] (5.2) Substitute the currently calculated displacement field \(w\) into the feature flow objective function \(E(w)\) to obtain the value of the current objective function;

[0078] (5.3) Subtract the value of the feature flow objective function in the previous iteration from the value of the current feature flow objective function and take the absolute value \(T\):

[0079] \(T = |E(w\) old ) - E(w)|

[0080] (5.4) Compare the absolute value with the set threshold;

[0081] If the absolute value \(T\) is less than the set threshold, stop the iteration and output the displacement field \(w\), and this displacement field is the required displacement field \(w\) * ;

[0082] If the absolute value \(T\) is greater than the set threshold, save the current objective function value and execute step (5.5);

[0083] (5.5) Calculate the gradient of the current displacement field: And correct the displacement field formula w

[0084] Return to step (5.2), where u is the horizontal component of w, v is the vertical component of w,

[0085] x and y are the abscissa and ordinate of the pixel point respectively.

[0086] Step 6, complete image registration according to the displacement field.

[0087] 6.1) Determine the displacement of a single pixel from the new coordinate according to the displacement field:

[0088] For any pixel (x, y) in the source image, a displacement vector w can be determined x，y , that is, the value of the displacement field w at the pixel (x, y);

[0089] According to the displacement vector w of each pixel x，y , accurately determine the displacement of each pixel in the source image through the following formula:

[0090] w x,y =(t x , t y )

[0091] where t x represents the horizontal displacement of the pixel (x, y), and t y represents the vertical displacement of the pixel (x, y);

[0092] 6.2) Calculate the new coordinates of the pixel:

[0093] For the pixel with the original position (x, y) in the source image, calculate the new position coordinates according to the obtained displacement

[0094]

[0095] 6.3) Perform pixel movement to complete registration:

[0096] Using the above calculation results, move each pixel in the source image to the new coordinates, so as to achieve pixel-level alignment with the target image. After completing the pixel displacement, that is, the source image and the target image are accurately aligned in space, and image registration is achieved.

[0097] Through this method, image distortion caused by factors such as shooting angle and sensor difference can be effectively corrected, and the accuracy of image analysis or subsequent processing can be improved.

[0098] Embodiment 2 A non-rigid image registration system based on rotation-invariant feature flow

[0099] Refer to Figure 2 , this example includes: a main direction screening module 1, a descriptor vector generation module 2, a feature flow objective function calculation module 3, a feature flow optimization module 4, and a displacement field mapping module 5. Among them,

[0100] The main direction screening module 1 includes: a pixel gradient solving sub-module 11, a candidate direction obtaining sub-module 12, and a direction sorting and screening sub-module 13; the descriptor vector generation module 2 includes: a rotation normalization sub-module 21 and a descriptor vector splicing sub-module 22. The working principle of the entire system is as follows:

[0101] The pixel gradient solving sub-module 11 is used to calculate the gradient direction and amplitude of any pixel in the input source image and target image, and input the gradient direction and amplitude to the candidate direction obtaining sub-module 12; the candidate direction obtaining sub-module 12 performs weighted calculation on the amplitudes of each direction neighborhood of each pixel based on the pixel gradient direction to obtain multiple candidate directions, and inputs the candidate directions to the direction sorting and screening sub-module 13; the direction sorting and screening sub-module 13 sorts the candidate directions, selects several candidate directions as the main direction, and then inputs the main direction to the rotation normalization sub-module 21.

[0102] The rotation normalization sub-module 21 performs rotation normalization processing on the neighborhood window of each pixel according to the input main direction of each pixel, so that the neighborhood window of the pixel is aligned to the same reference direction, and counts the proportion of the gradient of the pixel in the neighborhood window in each main direction to obtain a one-dimensional vector describing the local feature of the pixel in each main direction, and then inputs the one-dimensional vector of the main direction to the descriptor vector splicing sub-module 22; the descriptor vector splicing sub-module 22 sequentially splices the one-dimensional vectors of all the main directions end to end to obtain a descriptor vector, and then inputs the descriptor vector to the feature flow objective function calculation module 3.

[0103] The feature flow objective function calculation module 3 calculates the data matching term, the displacement constraint term, the flow field smoothing term, and the direction consistency constraint term respectively according to the input descriptor vector, and then sums up each item to form a feature flow objective function and inputs it to the feature flow optimization module 4.

[0104] The feature flow optimization module 4, based on the objective function, continuously adjusts and optimizes the displacement field by means of iterative solution, so that the objective function reaches the minimum value, thereby obtaining the optimal displacement field, and then inputs the obtained displacement field to the displacement field mapping module 5.

[0105] The displacement field mapping module 5 calculates the new target position coordinates of each pixel in the source image by using the input displacement field, and moves each pixel to the new target position coordinates, and finally completes image registration.

[0106] The effects of the present invention can be further illustrated by the following simulation experiments:

[0107] I. Simulation Conditions

[0108] The simulation conditions of the present invention include:

[0109] 1) Hardware platform: A PC platform with stable performance is adopted, which has sufficient computing power and storage space to meet the requirements of image processing and calculation.

[0110] 2) Software platform: The MATLAB simulation environment is adopted, which has a rich built-in image processing toolbox and algorithm library, providing strong data processing and visualization support for the experiment.

[0111] 3) Test image data: The image data used in the experiment is a pair of RGB three-channel images to be registered. One of them is the source image, and the other is the target image. There is a rotation angle between the source image and the target image. The rotation angle of the target image is fixed at 0 degrees, while the source image originally has a counterclockwise rotation of more than 20 degrees.

[0112] II. Simulation Content and Results

[0113] Under the above simulation conditions, the present invention is used to Figure 3 register the source image described in Figure 3 a and the target image described in Figure 3 b, and the result is as shown in Figure 3 c. As can be seen from

[0114] c, after the image registration processing of the present invention, the rotation angle of the output image is also restored to 0 degrees, which fully proves that the rotation deviation between the two images has been effectively corrected, achieving precise registration, indicating that the present invention can still maintain the ability of high-precision alignment when processing images with large rotation errors, ensuring the perfect fusion of image details and color information.

[0115] It should be noted that the step numbers in the specification and claims of the present invention are only for a clear description of the implementation embodiments of the present invention for easy understanding, and their sequence numbers are not limited.

Claims

1. A non-rigid image registration method based on rotation-invariant feature flow, characterized in that Including: (1) Calculate the pixel gradient direction θ for any pair of input images i，j and amplitude G. Obtain multiple candidate directions by weighting the amplitude in the neighborhood of the direction, and sort these candidate directions to screen out multiple main directions; (2) Based on multiple principal directions, perform rotation normalization within the neighborhood of a pixel point to obtain multiple feature vectors in a unified direction, and splice the multiple feature vectors to generate a descriptor vector; (3) Calculate a data matching term, a displacement constraint term, a flow field smoothing term, and a direction consistency constraint term according to the descriptor vector, and add the terms to obtain a feature flow objective function; (4) Based on the constructed objective function, calculate a displacement field that satisfies multiple constraint conditions through an iterative optimization algorithm; (5) Use the displacement field to determine the position of each pixel, and move the pixel points in the source image to the corresponding positions to obtain a registered image aligned with the target image.

2. The method according to claim 1, wherein In (1) above, for any pair of input images, calculate their pixel gradient directions θ i，j , and magnitudes G. The formulas are as follows: G = (G i , G j ) where G i is the horizontal direction of the pixel value, and G j is the vertical direction component of the pixel value.

3. The method according to claim 1, wherein The multiple candidate directions θ are obtained by weighted neighborhood amplitude in the (1) passing direction, which is implemented as follows: k , and the implementation includes the following: First, calculate the intensity magnitude D of each direction θ for each pixel (x, y) in the image k x，y (θ k ):​ Among them, R is a neighborhood centered on the pixel (x, y), and (i, j) represents other pixel points that appear in the neighborhood. i represents the abscissa of the pixel, and j represents the ordinate of the pixel; θ i，j is the pixel gradient direction of the input image; Secondly, for all directions θ k , find the direction θ x，y (θ * ) that maximizes D * , set a threshold coefficient β, and retain the directions with an amplitude greater than βD x，y (θ * ) and that are local extrema as the principal directions of the pixels.

4. The method according to claim 1, characterized in that, In the above (1), several main directions are sorted and selected from all the main directions. All the main directions θ k are sorted in descending order according to D x，y (θ k ), and the first main direction θ1, the second main direction θ2,... the nth main direction θ n of each pixel are obtained, where n is a hyperparameter set by the user.

5. The method according to claim 1, wherein In the above (2), based on multiple principal directions, rotation normalization is performed within the neighborhood of the key point, and its implementation includes the following: Taking each pixel as the center, determine a neighborhood window with a size of 3t×3t, and divide the neighborhood window into k directions with the center point as the axis, where t is a hyperparameter for controlling the window size set by the user; Statistically calculate the proportion of pixel gradients in each direction within the neighborhood window, and obtain a one-dimensional vector \(n\) of size \(k\times t\times t\) in each direction i used to describe the local features of pixels; Rotate the neighborhood window according to each direction of each pixel so that all directions of each pixel are aligned to the same reference direction.

6. The method according to claim 1, wherein In the above (2), splicing multiple feature vectors to generate a descriptor vector means selecting the first e from the multiple principal directions sorted and screened, and sequentially splicing their one-dimensional vectors end to end to obtain a descriptor vector with a size of e×k×t×t for each pixel point, where k is the number of directions divided by the neighborhood window, and t is a hyperparameter for controlling the window size set by the user.

7. The method according to claim 1, wherein In the above (3), constructing a feature flow objective function according to the descriptor vector, its implementation includes the following: (3a) Calculate data matches where p is the pixel position, w is the displacement vector, S1(p) represents the rotation-invariant feature descriptor extracted at pixel p in the first image, and S2(p + w) represents the rotation-invariant feature descriptor extracted at pixel p + w in the second image; (3b) Calculate the displacement constraint term where u(p) is the horizontal displacement of the displacement vector w, and v(p) is the vertical displacement of the displacement vector w; (3c) Calculate the flow field smoothing term: wherein α and d are hyperparameters with two different values, and ε is the spatial neighborhood of pixels; (3d) Calculate the direction consistency constraint term: (3f) Add the above four terms to obtain the feature flow objective function E(w): E(w) = L1(w) + L2(w) + L3(w) + L4(w).

8. The method according to claim 1, wherein Based on the constructed objective function in (4), solve for the displacement field that minimizes the objective function E(w). w *, and the formula is as follows: Among them, w* is the displacement field when the objective function E(w) takes the minimum value.

9. A non-rigid image registration system based on rotation-invariant feature flow, characterized in that, Including: A principal direction screening module for calculating the gradient direction and amplitude of the input image, obtaining multiple candidate directions by weighted neighborhood amplitudes in the same direction, and sorting and screening them to obtain multiple principal directions; A descriptor vector generation module for rotating and normalizing and splicing the one-dimensional feature vectors of multiple principal directions to obtain a descriptor vector for each pixel point; A feature flow objective function calculation module for calculating a data matching term, a displacement constraint term, a flow field smoothing term, and a direction consistency constraint term respectively, and then adding the terms to obtain a feature flow objective function; A feature flow optimization module for calculating the feature flow objective function for the descriptor and iteratively solving the displacement field that minimizes the objective function; A displacement field mapping module for using the displacement field to determine the destination position coordinates of each pixel moved in the source image, and moving the pixel points in the source image to the corresponding positions in the target image to complete the alignment of the registered image.

10. The system according to claim 9, wherein: The principal direction screening module includes: A pixel gradient solving sub-module, which is used to calculate the pixel gradient direction and the magnitude of the gradient for any pair of input images; A candidate direction obtaining sub-module, which is used to obtain all possible directions from the gradient direction, and weight the neighborhood magnitudes of all directions to obtain multiple candidate directions; A direction sorting and screening sub-module, which is used to sort the multiple candidate directions obtained by weighting, and then screen out multiple main directions; The descriptor vector generation module includes: A rotation normalization sub-module, which is used to rotate the neighborhood window of each pixel so that all the main direction directions of each pixel are aligned to the same reference direction, and count the proportion of the gradients of the pixels in the neighborhood window in each direction, and obtain a one-dimensional vector describing the local features of the pixels in each direction; A descriptor vector splicing sub-module, which sequentially splices the one-dimensional vectors of multiple main directions end to end to obtain a descriptor vector.

Images