A multi-source remote sensing image registration method based on improved PIIFD feature description

Abstract

The present invention proposes a multi-source remote sensing image registration method based on improved PIIFD feature description, which belongs to the field of visual image processing. Specifically, the following steps are performed: first, feature points of the multi-source remote sensing image to be registered and the reference image are extracted respectively; each feature point is described using the improved PIIFD descriptor to obtain a feature descriptor vector corresponding to each feature point; then, a BBF method is used to perform bidirectional initial matching using the feature descriptor vector, and incorrect matches are eliminated using the feature main direction and RANSAC method to obtain precise matching point pairs; a corresponding number of precise matching point pairs are selected according to different transformation modes to calculate the transformation matrix, and the least squares method is used to optimize the parameters of the transformation matrix; finally, the transformation matrix with optimized parameters is multiplied by the image to be registered to obtain the final registration result. The present invention solves the problem of difficult multi-scale image registration, has certain advantages in both the number of matches and registration accuracy, and is highly practical.

Description

Technical Field

[0001] The invention belongs to the field of image scene generation, and in particular relates to a multi-source remote sensing image registration method based on improved PIIFD feature description. Background Art

[0002] Image registration is the process of geometrically aligning two or more images of the same scene taken at different times, from different viewpoints, and by different sensors.

[0003] In recent years, due to the widespread application of multi-sensor vision systems, multi-source remote sensing image acquisition technology has continued to develop and plays an important role in pattern recognition, medical imaging, remote sensing and modern military fields. However, the problem of multimodal image registration has gradually emerged.

[0004] The registration accuracy of existing multi-source remote sensing image registration methods is still low. They can usually only handle the registration of remote sensing images of two modalities and cannot take into account multiple modalities; and they cannot handle the image scale differences caused by imaging from different sensors.

[0005] Existing registration methods can be roughly divided into two categories: region-based registration methods and feature-based registration methods.

[0006] The region-based registration method mainly uses image grayscale information to establish a similarity measure for image registration. However, the existing region-based registration method has different degrees of problems with image modality, intensity transformation, complex spatial transformation and computational complexity, so its application is greatly limited.

[0007] Feature-based registration methods are more robust to image intensity variations and noise. Commonly used features include point features, line features, and region features. Since line and region features have regional characteristics and cannot accurately determine matching locations, most algorithms primarily extract point features.

[0008] Existing multi-source remote sensing image methods such as PSO-SIFT, RIFT, SURF-RPM, etc. are unable to process multi-scale and multi-source remote sensing image data, and thus cannot well register images. Summary of the Invention

[0009] In order to overcome the shortcomings of the above-mentioned multi-source remote sensing image registration method, the present invention proposes a multi-source remote sensing image registration method based on an improved PIIFD feature description, which can adapt to more modalities and has multi-scale image processing capabilities.

[0010] The multi-source remote sensing image registration method based on the improved PIIFD feature description comprises the following steps:

[0011] Step 1: For the multi-source remote sensing image and the reference image to be registered, the KAZE method is selected to extract the feature points of each of the two images;

[0012] The specific steps are as follows:

[0013] First, based on nonlinear filtering, a nonlinear scale space is constructed for the image to be registered or the reference image respectively;

[0014] The nonlinear scale space includes O scale space groups and S sub-layers;

[0015] Then, for the i-th group i-th sublayer (o i ,s i ) A random pixel in the image is compared with its eight neighboring pixels in the same layer and the nine pixels at the same position in the upper and lower layers. If the pixel is larger or smaller than all neighboring pixels, the pixel is considered an extreme point.

[0016] Then, the response value is calculated by scaling the Hessian determinant, and the extreme value corresponding to the response value is the KAZE feature point;

[0017] The Hessian formula is:

[0018]

[0019] where l xx and L yy is the second-order partial derivative of the brightness L in the x or y direction, L xy is the mixed second-order partial derivative of the brightness L in the x and y directions, σ 2 is the number of Octave scale space groups;

[0020] Step 2: Describe each feature point using the improved PIIFD descriptor to obtain the feature descriptor vector corresponding to each feature point;

[0021] The specific process is:

[0022] First, for the feature point σ i , calculate its scale factor μ i , as follows:

[0023]

[0024] offset is a constant, usually 1.6; λ i To adjust the scale range of variables, the sub-pixel approximate coordinates of the feature points are calculated as follows:

[0025]

[0026] x=(x,y,λ)T ,

[0027] Where L(x) is the approximation of the Laplace operator, is the approximate value of the spatial coordinate. The approximate coordinate is calculated by the above formula, so as to calculate the scale factor; λ is a random variable in [-1,1].

[0028] Then, by the scaling factor μ i Multiplying with the standard quantitative k, the detection neighborhood size of the improved adaptive multi-scale PIIFD descriptor is (kμ)*(kμ);

[0029] Each feature point corresponds to a different scale factor, so the detection neighborhood of the multi-scale PIIFD descriptor of each feature is also different.

[0030] Finally, within the detection neighborhood of size (kμ)*(kμ), the improved PIIFD descriptor is calculated to extract the descriptor;

[0031] Specifically: from the square neighborhood of (kμ)*(kμ), extract an area consisting of 16 small squares, and the area of ​​each small square is Corresponding to a direction histogram.

[0032] By calculating the sum of opposite directions, the 16 direction histograms with an average coverage of 0~2π(0°, 22.5°, ..., 337.5°) are converted into 8 degenerate direction histograms with an average coverage of 0~π(0°, 22.5°, ..., 157.5°). Then, the linear descriptor constructed by the degenerate direction histogram and the row vector of its 180° rotation matrix is ​​normalized to obtain the feature point σ i The corresponding descriptor vector is 128 in length.

[0033] Step 3: Use the BBF method to perform bidirectional initial matching using the feature descriptor vector, and use the feature main direction to eliminate false matches, and then use the RANSAC method to eliminate false matches to obtain precise matching point pairs;

[0034] Specifically:

[0035] Step 301: For the current feature point a of the image to be registered, calculate the Euclidean distance between the feature point a and all feature points in the reference image in sequence, select the closest feature point that meets the conditions, match it with the feature point a, and add it to the initial matching result;

[0036] The conditions that are met are as follows:

[0037] The closest feature point and the second closest feature point in the reference image satisfy the following formula:

[0038]

[0039] Where d1 is the Euclidean distance between the feature point a and the nearest feature point, d2 is the Euclidean distance between the feature point a and the next nearest feature point, and η is a preset threshold.

[0040] Step 302: Count the initial matching results and eliminate rotation errors using the main feature direction;

[0041] Specifically:

[0042] Randomly select two sets of feature points corresponding to the initial matching, and their main directions are and N is the number of matching pairs, is the main direction angle of the two sets of matching feature points, so the main direction angle difference of the feature for:

[0043]

[0044] After twisting the image to the same direction, remove the images that do not meet the The remaining matching pairs are the matching results after eliminating the rotation error.

[0045] Step 303: Use RANSAC to remove incorrect matches;

[0046] Specifically, for the current iteration, four non-collinear matching point pairs are randomly selected from the matching pairs after eliminating the rotation error as sample data to calculate the transformation matrix H, which is recorded as the matrix model M;

[0047] Then, the cost function error between each matching pair and the matrix model M is calculated in turn. If the error is less than the set threshold, the current matching pair is added to the internal point set I;

[0048] The number of iterations increases by 1, and 4 non-collinear matching point pairs are randomly selected as samples again to calculate the matrix model M1. Similarly, the cost function error between each matching point pair and the matrix model M1 is calculated, and the matching points with an error less than the set threshold are added to the internal point set I1;

[0049] Until the set number of iterations is reached, the inner point set with the largest number of matching point pairs is selected as the precise matching point pair;

[0050] Step 4: Select a corresponding number of precise matching point pairs from the precise matching point pairs according to the different transformation modes to calculate the transformation matrix, and optimize the parameters of the transformation matrix using the least squares method;

[0051] Transformation methods include: similarity transformation, affine transformation or projective transformation;

[0052] Similarity transformation corresponds to two sets of precise matching point pairs, affine transformation requires three sets of precise matching point pairs, and projection transformation requires four sets of precise matching point pairs;

[0053] Step 5: Multiply the transformation matrix with optimized parameters by the image to be registered to obtain the final registration result.

[0054] The advantages of the present invention are:

[0055] The present invention provides a multi-source remote sensing image registration method based on improved PIIFD feature description. Compared with existing methods, it can register image data of multiple modalities. The improved PIIFD descriptor solves the problem of difficult multi-scale image registration. It has certain advantages in both the number of matches and registration accuracy, and is highly practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a multi-source remote sensing image registration method based on improved PIIFD feature description of the present invention;

[0057] Figure 2 This is a reverse addition calculation effect diagram of the present invention;

[0058] Figure 3 This is a schematic diagram of the present invention using the main direction histogram to select correctly matched feature point pairs;

[0059] Figure 4 This is the effect diagram of the present invention using the main feature direction and RANSAC method to eliminate incorrect matches;

[0060] Figure 5 Schematic diagram of selecting the main direction consistency and RANSAC algorithms for mismatch removal in the present invention;

[0061] Figure 6 The figure is a schematic diagram of pictures obtained by transforming a picture according to different transformation methods according to the present invention. DETAILED DESCRIPTION

[0062] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of the present invention.

[0063] The present invention proposes a multi-source remote sensing image registration method based on an improved PIIFD feature description, which has a wider range of adaptive modalities and multi-scale image processing capabilities. Specifically, the method comprises the following steps: first, the multi-source remote sensing images to be registered are preprocessed to obtain a registration image, and the KAZE method is selected to extract feature points; then, the extracted feature points are described using an improved PIIFD descriptor; and the feature descriptions are matched; finally, the feature point matching results are subjected to parameter estimation and model transformation using a similarity transformation, an affine transformation, or a projective transformation, and the model parameters are calculated using the least squares method. The present invention can register image data from multiple modalities and solves the problem of difficult multi-scale image registration through the improved PIIFD descriptor. It has certain advantages in both the number of matches and registration accuracy, and is highly practical.

[0064] The multi-source remote sensing image registration method based on the improved PIIFD feature description is as follows: Figure 1 As shown, the following steps are included:

[0065] Step 1: For the multi-source remote sensing image and the reference image to be registered, the KAZE method is selected to extract the feature points of each of the two images;

[0066] For multi-source images to be registered, a group of images is usually divided into images to be registered and reference images, preprocessed to obtain a reference image, and the KAZE method is selected to extract feature points of the reference image and the image to be registered respectively;

[0067] The specific steps are as follows:

[0068] First, based on nonlinear filtering, a nonlinear scale space is constructed for the image to be registered or the reference image respectively;

[0069] The nonlinear scale space includes O scale space groups and S sub-layers; the relationship between each layer is expressed as:

[0070]

[0071] Where σ0 is the benchmark scale level, o i and s i are the number of Octave scale space groups and sublevel sub-layers,

[0072] o i ∈[0,1,...,O-1],s i ∈[0, 1, ..., S-1], i∈[0, 1, ..., N], where N is the total number of images after nonlinear filtering.

[0073] Next, the pixel unit σ i Converted to time units, the expression of the scale factor of each sublayer is:

[0074]

[0075] t i is the evolution time. According to a set of evolution times, the nonlinear scale space is obtained by using the AOS algorithm.

[0076] Then, for the i-th group i-th sublayer (o i ,s i ) A random pixel in the image is compared with its eight neighboring pixels in the same layer and the nine pixels at the same position in the upper and lower layers. If the pixel is larger or smaller than all neighboring pixels, the pixel is considered an extreme point.

[0077] Then, in the scale space, the response value of each pixel in each layer of the image is calculated by the scale-normalized Hessian determinant. The extreme value corresponding to the response value is the KAZE feature point.

[0078] For example, if the scale space has three layers, the value of the pixel point, that is, the feature point, is the maximum of the 26 values ​​of the surrounding eight and the upper and lower layers, and the extreme value is taken. The response value is the value of the Hessian determinant normalized by the extreme value. The Hessian formula is:

[0079]

[0080] Among them L xx and L yy is the second-order partial derivative of the brightness L in the x or y direction, L xy is the mixed second-order partial derivative of the brightness L in the x and y directions, σ 2 is the number of Octave scale space groups;

[0081] Step 2: Describe each feature point using the improved PIIFD descriptor to obtain the feature descriptor vector corresponding to each feature point;

[0082] Feature points are described using the improved PIIFD descriptor, primarily by adaptively extracting the region of feature description. In multimodal images, different modal images typically have different resolutions and viewing areas, which results in scale variations. PIIFD uses a fixed neighborhood size (typically 40*40) and cannot represent changes in feature scale. The SIFT algorithm detects feature points in scale space, providing scale information for each feature point. This scale information is then used to determine the neighborhood size for extracting descriptors, thereby achieving scale invariance. Therefore, drawing on the SIFT algorithm's principles, this paper requires an adaptive neighborhood area to achieve scale invariance and accurately describe features.

[0083] The specific process is:

[0084] First, for the i-th feature point σ i , (the adaptive neighborhood is based on the scale information of the feature point to describe different feature neighborhoods, so different scale factors need to be taken for different feature points), calculate its scale factor μ i , as follows:

[0085]

[0086] offset is a constant, usually 1.6; λ i To adjust the scale range of variables, the sub-pixel approximate coordinates of the feature points are calculated as follows:

[0087]

[0088] x=(x,y,λ) T ,

[0089] Where L(x) is the approximation of the Laplace operator, is the approximate value of the spatial coordinate. The approximate coordinate is calculated by the above formula, so as to calculate the scale factor; λ is a random variable in [-1,1].

[0090] Then, by the scaling factor μ i Multiplying with the custom standard quantitative k, the improved adaptive multi-scale PIIFD descriptor detection neighborhood size is (kμ)*(kμ),

[0091] A custom standard quantity is then multiplied by the scale factor to determine the neighborhood area, which allows for varying detection ranges for each feature. Each feature point corresponds to a different scale factor, so the detection neighborhood of the multi-scale PIIFD descriptor for each feature is also different.

[0092] Finally, within the detection neighborhood of size (kμ)*(kμ), the improved PIIFD descriptor is calculated to extract the descriptor;

[0093] Specifically: first calculate the magnitude and direction of the image gradient, and use the continuous average square gradient to calculate the main direction of the feature point; from the square neighborhood of (kμ)*(kμ), extract an area consisting of 16 small squares, and the area of ​​each small square is Corresponding to a direction histogram.

[0094] like Figure 2As shown in Figure 1, by calculating the sum of opposite directions, the 16 direction histograms with an average coverage of 0 to 2π (0°, 22.5°, ..., 337.5°) are converted into 8 degenerate direction histograms with an average coverage of 0 to π (0°, 22.5°, ..., 157.5°). Then, the linear descriptor constructed by the degenerate direction histogram and the row vector of its 180° rotation matrix is ​​normalized to obtain the feature point σ i The corresponding descriptor vector is 128 in length.

[0095] Step 3: Use the BBF method to perform bidirectional initial matching using the feature descriptor vector, and use the feature main direction to eliminate false matches, and then use the RANSAC method to eliminate false matches to obtain precise matching point pairs;

[0096] Specifically:

[0097] Step 301: For the current feature point a of the image to be registered, calculate the Euclidean distance between the feature point a and all feature points in the reference image in sequence, select the closest feature point that meets the conditions, match it with the feature point a, and add it to the initial matching result;

[0098] The conditions that are met are as follows:

[0099] The closest feature point and the second closest feature point in the reference image satisfy the following formula:

[0100]

[0101] Where d1 is the Euclidean distance between the feature point a and the nearest feature point, d2 is the Euclidean distance between the feature point a and the next nearest feature point, and η is a preset threshold.

[0102] Step 302: Count the initial matching results and eliminate rotation errors using the main feature direction;

[0103] Specifically:

[0104] Randomly select two sets of feature points corresponding to the initial matching, and their main directions are and N is the number of matching pairs, is the main direction angle of the two sets of matching feature points, so the main direction angle difference of the feature for:

[0105]

[0106] After twisting the image to the same direction, remove the images that do not meet the The remaining matching pairs are the matching results after eliminating the rotation error.

[0107] For example, if a set of images is matched, the main direction of the feature points ensures the rotation invariance of the feature points. When the images are rotated to the same position, theoretically, the main directions of the features between the correct matching pairs are the same, while the directions of the incorrect matching pairs are different. Therefore, when the main direction difference of a pair of matching pairs is significantly different from that of other matching pairs, it is considered to be a pair of incorrect matching pairs. In this paper, when the images are rotated to the same position, the main direction angle difference of the two matching feature points is less than 5°, which is considered to be a pair of correct matching pairs. Figure 3 As shown, this embodiment uses a histogram to count, with 5° as the interval, the range of the histogram x-axis is [0°, 360°), and the y-axis statistics are included in the corresponding interval. Finally, the feature pairs in the interval with the most feature pairs in the histogram are taken as the correctly matched feature point pairs in this paper.

[0108] Step 303: Use RANSAC to remove incorrect matches;

[0109] Specifically, for the current iteration, four non-collinear matching point pairs are randomly selected from the matching pairs after eliminating the rotation error as sample data to calculate the transformation matrix H, which is recorded as the matrix model M;

[0110] Then, the cost function error between each matching pair and the matrix model M is calculated in turn. If the error is less than the set threshold, the current matching pair is added to the internal point set I;

[0111] The number of iterations increases by 1, and 4 non-collinear matching point pairs are randomly selected as samples again to calculate the matrix model M1. Similarly, the cost function error between each matching point pair and the matrix model M1 is calculated, and the matching points with an error less than the set threshold are added to the internal point set I1;

[0112] Until the set number of iterations is reached, the inner point set with the largest number of matching point pairs is selected as the precise matching point pair; the matrix model at this time is the optimal model, and the corresponding cost function is the smallest. The purpose of RANSAC is to find the optimal parameter matrix model so that the number of data points that satisfy the matrix model is the largest.

[0113] like Figure 4 The figure shows the schematic diagram of removing mismatches based on the consistency of the main direction and removing mismatches based on the RANSAC algorithm. Figure 4 (a) Bilateral matching graph of AM-PIIFD feature descriptors extracted from a set of infrared and visible images; Figure 4 (b) Schematic diagram of removing mismatches based on the consistency of the main direction of the test image; Figure 4 (c) is a schematic diagram of removing mismatches of the test image based on the RANSAC algorithm.

[0114] Step 4: Select a corresponding number of precise matching point pairs from the precise matching point pairs according to the different transformation modes to calculate the transformation matrix, and optimize the parameters of the transformation matrix using the least squares method;

[0115] Transformation methods include: similarity transformation, affine transformation or projective transformation;

[0116] Similarity transformations correspond to two sets of precise matching point pairs, affine transformations require three sets, and projection transformations require four sets. Therefore, when the number of matching pairs is less than or equal to one, it is indicated that registration cannot be completed and no transformation model can be selected for registration. When there are two matching pairs, similarity transformation is selected for registration. When there are three matching pairs, this article selects affine transformation for registration. When there are four or more matching pairs, projection transformation is used for registration. The more matching pairs that can be matched, the more complex the transformation model can be, resulting in better registration results.

[0117] Step 5: Multiply the transformation matrix with optimized parameters by the image to be registered to obtain the final registration result with the reference image.

[0118] Example:

[0119] A multi-source remote sensing image registration method based on improved PIIFD feature description, the method comprising the following steps:

[0120] Step S1, pre-processing the image to be registered to obtain a registration image, and selecting the KAZE method to extract feature points;

[0121] The specific steps are as follows: First, a nonlinear scale space is constructed for each image. In the scale space, the response value of each layer image at each pixel is calculated by the scale-normalized Hessian determinant, where the extreme value of the response value is the KAZE feature point, as follows:

[0122] The nonlinear scale space described in S1.1 is constructed using nonlinear filtering, where the nonlinear filtering method can be described by a nonlinear partial differential equation as shown below:

[0123]

[0124] Where L is the brightness of the image, time t is the scale parameter, div and They represent gradient and divergence respectively, and c is the conductivity function:

[0125]

[0126] in:

[0127] is the gradient image after Gaussian smoothing of the original image, and k is the factor that controls the diffusion level, so that the nonlinear diffusion can adapt to the local features of the image.

[0128] Construct a nonlinear scale space, which includes O scale space groups and S sub-layers. The relationship between each layer is expressed as:

[0129] Then the pixel unit σ i Converted to time units, the scale factor of each sub-layer is expressed as t i =σ i 2 / 2,t i is the evolution time. According to a set of evolution times, a nonlinear scale space can be obtained by using the AOS algorithm.

[0130] S1.2 Feature point detection calculates the response value of each pixel in each layer of the image through the scale-normalized Hessian determinant, where the Hessian formula is:

[0131]

[0132] Among them L xx and L yy is the second-order partial derivative of the brightness L in the x or y direction, L xy It is the mixed second-order partial derivative of the brightness L in the x and y directions. Find the extreme value of the response in all filtered images and get the feature points.

[0133] Step S2, performing feature description on the feature points extracted in step S1 using the improved PIIFD descriptor;

[0134] The specific steps are: first, adaptively obtain the feature description area; then, calculate the improved PIIFD descriptor;

[0135] S2.1 The present invention sets the scale factor of the characteristic point of the response to μ, as follows:

[0136]

[0137] The offset is 1.6, o i and s i The number of Octave scale space groups and sublevel sub-layers to which the current feature point belongs, λ i is a variable, and is obtained by calculating the sub-pixel approximate coordinates of the feature points as follows:

[0138]

[0139] x=(x,y,λ) T ,

[0140] Where L(x) is the approximation of the Laplace operator, is the approximate value of the spatial coordinate. i , it can be determined that the detection neighborhood size of the improved adaptive multi-scale PIIFD descriptor is (kμ)*(kμ), where the default value of k is 6, so that a variable detection range is adopted for each feature.

[0141] S2.2 extracts the descriptor. First, the magnitude and direction of the image gradient are calculated. The continuous average square gradient is used to calculate the main direction of the feature point. In order to obtain better accuracy and computational efficiency, the changing square neighborhood determined in S2.1 is divided into regions. The extraction region consists of 16 small squares. The area of ​​each small square is Corresponding to a direction histogram. By calculating the sum of opposite directions, the 16 direction histograms with an average coverage of 0 to 2π (0°, 22.5°, ..., 337.5°) are converted into 8 degenerate direction histograms with an average coverage of 0 to π (0°, 22.5°, ..., 157.5°). This is to achieve invariance when the gradient is reversed. Then, the linear descriptor constructed using the obtained direction histogram and the row vector of its 180° rotation matrix is ​​normalized to obtain a descriptor vector of length 128.

[0142] Step S3, performing feature matching on the feature description obtained in step S2;

[0143] The specific steps are to perform a bidirectional matching on the feature description obtained in step S2 using the BBF method, first use the main direction of the feature to eliminate false matches, and then use the RANSAC method to eliminate false matches;

[0144] S3.1 performs a bilateral matching using the BBF method. The matching strategy is the Euclidean distance between the feature points corresponding to the registered image and all features in the other registered image. When the closest feature point and the next closest feature point satisfy the following formula: Then the match is considered correct;

[0145] Where d1 is the Euclidean distance between the feature point and the nearest feature point, d2 is the Euclidean distance between the feature point and the next nearest feature point, and η is a preset threshold.

[0146] S3.2 The specific steps of consistency of feature main direction are: statistics are performed on the initial matching obtained in S3.1, and the main directions of the two sets of feature points corresponding to the initial matching are and N is the number of matching pairs, is the main direction angle of the two sets of multi-source image feature points, so the main direction angle difference of the feature for:

[0147]

[0148] After eliminating the rotation error, remove the unsatisfactory Matching pairs, the remaining matching matching pairs, and then RANSAC is used to eliminate the wrong matches. Specifically, 4 matching point pairs are randomly selected from the above matching matching pairs as sample data to estimate parameters, and the distance of all matching point pairs under the estimated transformation parameters is calculated. If the distance is less than the preset threshold, the matching point is considered to be an internal point, otherwise it is an external point. Repeated iterations are performed to take the internal point as the precise matching point pair.

[0149] In step S4, the feature point matching result obtained in step S3 is subjected to parameter estimation and model transformation using one of similarity transformation, affine transformation and projection transformation, and the model parameters are calculated using the least squares method.

Claims

1. A multi-source remote sensing image registration method based on improved PIIFD feature description, characterized in that: The specific steps are as follows: Step 1: For the multi-source remote sensing image and the reference image to be registered, the KAZE method is selected to extract the feature points of each of the two images; Step 2: Describe each feature point using the improved PIIFD descriptor to obtain the feature descriptor vector corresponding to each feature point; The specific process is: First, for the feature points , calculate its scale factor , as follows: ， is a constant; is the variable for adjusting the scale range; S is the total number of sub-layers, and are the number of Octave scale space groups and sublevel sublayers respectively; Then, by the scale factor Multiplying with the standard quantitative k, the detection neighborhood size of the improved adaptive multi-scale PIIFD descriptor is ; Finally, in the size In the detection neighborhood, the improved PIIFD descriptor is calculated to extract the descriptor; Specifically: from In the square neighborhood, extract the area consisting of 16 small squares, and the area of ​​each small square is , corresponding to a direction histogram; By calculating the sum of opposite directions, the 16 direction histograms with an average coverage of 0~2π and directions of 0°, 22.5°,..., 337.5° are converted into 8 degenerate direction histograms with an average coverage of 0~π and directions of 0°, 22.5°,..., 157.5°. Then, the linear descriptor constructed by the degenerate direction histogram and the row vector of its 180° rotation matrix is ​​normalized to obtain the feature point The corresponding descriptor vector is 128 in length; Step 3: Use the BBF method to perform bidirectional initial matching using the feature descriptor vector, and use the feature main direction to eliminate false matches, and then use the RANSAC method to eliminate false matches to obtain precise matching point pairs; Specifically: Step 301: For the current feature point a of the image to be registered, calculate the Euclidean distance between the feature point a and all feature points in the reference image in sequence, select the closest feature point that meets the conditions, match it with the feature point a, and add it to the initial matching result; The conditions that are met are as follows: The closest feature point and the second closest feature point in the reference image satisfy the following formula: ， in is the Euclidean distance between the feature point a and the nearest feature point, is the Euclidean distance between the feature point a and the next closest feature point, is the preset threshold; Step 302: Count the initial matching results and eliminate rotation errors using the main feature direction; Specifically: Randomly select two sets of feature points corresponding to the initial matching, and their main directions are and , is the number of matching pairs, is the main direction angle of the two sets of matching feature points, so the main direction angle difference of the feature for: ， After twisting the image to the same direction, remove the images that do not meet the The remaining matching pairs are the matching results after eliminating the rotation error; Step 303: Use RANSAC to remove incorrect matches; Specifically, for the current iteration, four non-collinear matching point pairs are randomly selected from the matching pairs after eliminating the rotation error as sample data to calculate the transformation matrix H, which is recorded as the matrix model M; Then, the cost function error between each matching pair and the matrix model M is calculated in turn. If the error is less than the set threshold, the current matching pair is added to the inlier set I; The number of iterations increases by 1, and 4 non-collinear matching point pairs are randomly selected as samples again to calculate the matrix model M1. Similarly, the cost function error between each matching point pair and the matrix model M1 is calculated, and the matching points with an error less than the set threshold are added to the internal point set I1; Until the set number of iterations is reached, the inner point set with the largest number of matching point pairs is selected as the precise matching point pair; Step 4: Select a corresponding number of precise matching point pairs from the precise matching point pairs according to the different transformation modes to calculate the transformation matrix, and optimize the parameters of the transformation matrix using the least squares method; Step 5: Multiply the transformation matrix with optimized parameters by the image to be registered to obtain the final registration result.

2. The multi-source remote sensing image registration method based on improved PIIFD feature description according to claim 1, characterized in that: The step 1 is specifically as follows: First, based on nonlinear filtering, a nonlinear scale space is constructed for the image to be registered or the reference image respectively; The nonlinear scale space includes O scale space groups and S sub-layers; Then, for the Group sublayer A random pixel in the image is compared with its eight neighboring pixels in the same layer and the nine pixels in the upper and lower layers corresponding to the same position. If the pixel is larger or smaller than all neighboring pixels, the pixel is considered an extreme point. Finally, the response value is calculated by scaling the Hessian determinant, and the extreme value corresponding to the response value is the KAZE feature point; The Hessian formula is: ， in is the second-order partial derivative of the brightness L in the x or y direction, is the mixed second-order partial derivative of the brightness L in the x and y directions, is the number of Octave scale space groups.

3. The multi-source remote sensing image registration method based on improved PIIFD feature description according to claim 1, characterized in that: In the second step, the variables of the scale range are adjusted , is obtained by calculating the sub-pixel approximate coordinates of the feature points, as follows: ， ， in is the approximation of the Laplace operator, is the approximate value of the spatial coordinate. The approximate coordinate is calculated by the above formula, so as to calculate the scale factor. is a random variable in the range [-1, 1].

4. The multi-source remote sensing image registration method based on improved PIIFD feature description according to claim 1, characterized in that: In step 4, the transformation method includes: similarity transformation, affine transformation or projective transformation; Similarity transformation corresponds to two sets of precise matching point pairs, affine transformation requires three sets of precise matching point pairs, and projection transformation requires four sets of precise matching point pairs.

Images