Target multi-angle intrinsic feature mining method based on laplacian eigenmap

Abstract

The application provides a target multi-angle intrinsic feature mining method and system based on Laplace feature mapping, comprising: extracting target multi-angle SAR image features; constructing a target original high-dimensional observation feature space according to the multi-angle SAR image features; performing Laplace feature mapping on the original high-dimensional observation feature space to obtain a mapped low-dimensional intrinsic feature space; and obtaining a distribution relationship between a target angle and a manifold surface through the low-dimensional intrinsic feature space. The application faces multi-angle target SAR image samples, adopts a Laplace feature mapping method to find a low-dimensional manifold surface in a high-dimensional observation feature space, and proves the consistency between the target angle change rule and the manifold surface rule, thereby laying a foundation for subsequent multi-angle target feature representation and feature matching applications.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intrinsic feature mining, in particular to a target multi-angle intrinsic feature mining method and system based on Laplace feature mapping. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) has been widely used in various fields due to its all-weather and all-day high-resolution observation capabilities, such as flood monitoring, ocean monitoring, agricultural surveying, and topographic mapping. Multi-angle SAR can obtain more rich angle information of the target, so the interpretation advantage of multi-angle SAR image is more obvious.

[0003] For multi-angle SAR images, due to the large difference in the backscattering characteristics of electromagnetic waves by targets at different angles, the image representation of the same target at different angles also differs greatly, causing great difficulty in interpreting and reading SAR images. Based on this, the present application seeks the consistency of angle and feature rules by analyzing multi-angle SAR image samples of typical targets at different observation angles, finds a feature description method that can effectively represent the angle change of the target, and lays a foundation for subsequent feature representation and feature matching applications of multi-angle targets.

[0004] A multi-angle SAR image target recognition method based on EfficientNet and BiGRU in Radar, Vol. 10, No. 6, December 2021, proposes a multi-angle SAR target recognition model based on EfficientNet and BiGRU, and uses island loss to train the model, and uses the same MSTAR dataset as the present application to perform multi-angle SAR target recognition experiments. This article mainly consists of an EfficientNet network for extracting single image features and a BiGRU network for further extracting multi-angle SAR image sequence timing features to form target features for recognition, which is different from the gray feature, shape feature and texture feature used in the present application. In this article, island loss and cross-entropy loss are used as the training method for target recognition tasks, which is different from the manifold learning method of seeking low-dimensional intrinsic features in high-dimensional feature space in the present application. Compared with the present application, the present application can find the angle distribution change rule of different angle targets in the intrinsic feature space, which is more conducive to the subsequent classification and recognition of targets based on this rule.

[0005] The document Radar Science and Technology, Vol. 9, No. 2, April 2020, Multi-angle observation spaceborne SAR technology research summarizes the new technology of multi-angle observation spaceborne SAR, and discusses the application directions of multi-angle observation spaceborne SAR imaging processing algorithm and image radiation quality improvement, sidelobe suppression, etc. The main research content is from the perspective of multi-angle SAR imaging mechanism, and analyzes the quality characteristics of multi-angle SAR imaging. The present application is feature mining and angle change rule exploration for image targets after obtaining multi-angle SAR images, and the article belongs to different but complementary research directions.

[0006] The document Radar Science and Technology, Vol. 9, No. 2, April 2020, Multi-angle SAR moving target detection technology and its experimental verification research on Gaofen 3, based on the advantages of long-time observation of multi-angle SAR on the scene and large synthetic aperture angle, the moving target detection ability is researched, and the logarithmic background difference detection algorithm is proposed. The application verification of multi-angle moving target detection on Gaofen 3 staring spotlight mode is carried out. But the research object of the literature is a moving target, and the mining of static targets is lacking.

[0007] The document Radar Science and Technology, Vol. 7, No. 5, October 2018, Unmanned aerial vehicle classification and recognition based on double radar micro-motion feature fusion, first uses multiple radars to observe the target from different angles; then the collected radar data is respectively subjected to short-time Fourier transform to obtain time-frequency spectrum; then the principal component analysis is used to extract features from the time-frequency spectrum, and the features obtained by two different angle radar sensors are fused together; finally, support vector machine is used for training and classification recognition. However, the feature extraction object of the article is not SAR image, but time-frequency spectrum in the collected radar data, in addition, the principal component analysis method used in the article needs the linearization hypothesis of the intrinsic manifold surface, which is not suitable for natural images.

[0008] Patent document CN114519778A discloses a target three-dimensional reconstruction method of multi-angle SAR data, comprising: acquiring a sub-aperture image sequence of multi-angle SAR; segmenting each sub-aperture image, segmenting each sub-aperture image into a first target region and a first background region, and marking the first target region and the first background region to obtain a first mask image; determining a shadow protection region in the first background region according to the first mask image, marking the shadow protection region to obtain a second mask image; creating an initial 3D voxel grid corresponding to the sub-aperture image sequence; removing non-target voxels in the initial 3D voxel grid according to the second mask image to generate a 3D point cloud model corresponding to the sub-aperture image sequence; and three-dimensionally reconstructing the target according to the 3D point cloud model. However, the research content of patent document CN114519778A is based on multi-angle SAR sub-aperture image sequence to extract scene elevation information, and the data collected by this extraction method is limited in the application of multi-angle target classification and identification.

[0009] Patent document CN106897985B discloses a multi-angle SAR image fusion method based on visibility classification, calculates the multi-view visibility LI of the non-overlapping region pixels Inon_overlap of each azimuth view image according to the binary segmentation results of the overlapping region and the non-overlapping region of the image pixels, and automatically realizes fine classification of image pixels through the visibility index. However, the application purpose of patent document CN106897985B is image fusion, and the processing object adopted is image pixels, not the image features used for constructing the original high-dimensional observation space in the present application. SUMMARY

[0010] In view of the defects in the prior art, the purpose of the present application is to provide a target multi-angle intrinsic feature mining method and system based on Laplace feature mapping.

[0011] According to the target multi-angle intrinsic feature mining method based on Laplace feature mapping provided by the present application, the following steps are included:

[0012] Step S1: Extracting target multi-angle SAR image features;

[0013] Step S2: Constructing a target original high-dimensional observation feature space according to the multi-angle SAR image features;

[0014] Step S3: Performing Laplace feature mapping on the original high-dimensional observation feature space to obtain a mapped low-dimensional intrinsic feature space;

[0015] Step S4: Obtaining the distribution relationship between the target angle and the manifold surface through the low-dimensional intrinsic feature space.

[0016] Preferably, the SAR image features include: basic pixel features, grayscale features, shape features, and texture features.

[0017] The SAR image features are obtained in a filtered image at a set scale.

[0018] Preferably, step S2 includes:

[0019] Step S2.1: maximum-minimum value normalization processing is performed on the multi-angle SAR image features respectively, and the calculation formula is as follows:

[0020]

[0021] wherein f(i) is an i-th feature value, f min is the minimum value among all feature values, f max is the maximum value among all feature values, and F(i) is a normalized feature value normalized to the range of [0, 1];

[0022] Step S2.2: all processed features are concatenated to obtain an original high-dimensional observation space feature matrix, denoted as HighDimF(N, M), wherein N is the number of samples, and M is the feature dimension after concatenation.

[0023] Preferably, step S3 includes:

[0024] Step S3.1: the original high-dimensional observation space feature matrix is taken as a data set X in an original high-dimensional observation space, the neighborhood of each data point is calculated, and an adjacency graph is constructed.

[0025] Step S3.2: the adjacency graph is weighted to obtain a weight matrix W, as follows:

[0026]

[0027] wherein W ij represents a weight value matrix given to an edge connecting a near neighbor point of each point in the adjacency graph constructed in the original high-dimensional observation space; e represents a natural base; t represents a heat kernel function coefficient; x i represents an i-th data point in the data set in the original high-dimensional observation space; x j represents a j-th data point in the data set in the original high-dimensional observation space; J i represents a sample point x i in the original high-dimensional observation space; and

[0028] Step S3.3: a degree matrix D is calculated according to the weight matrix, and the calculation formula is as follows:

[0029] D ii = Σ j W ij

[0030] wherein D ii represents a metric matrix of the weighted adjacency graph, the metric matrix is in the form of a diagonal matrix, and the value on the diagonal line of the matrix is the sum of the weight values of the edges between each sample point and the near neighbor points;

[0031] Step S3.4: obtaining a Laplacian matrix L through the difference between the metric matrix and the weight matrix, and the calculation formula is as follows:

[0032] L=D-W

[0033] Step S3.5: performing generalized eigenvalue decomposition on the Laplacian matrix L, and the calculation formula is as follows:

[0034] Ly=λDy

[0035] taking the second smallest eigenvalue to the first d+1 smallest eigenvalues λ2,…,λ d+1 , corresponding to the eigenvectors U=[u2,…,u d+1 , to obtain a low-dimensional embedding representation , wherein λ represents an eigenvalue, y represents an eigenvector, d represents the first d dimensions obtained by sorting the eigenvalues in ascending order after eigenvalue decomposition, k represents a low-dimensional intrinsic feature dimension, y i represents the eigenvector corresponding to the first 2-d+1 eigenvalues in the eigenvector of the i-th sample point;

[0036] Further, the data set Y in the low-dimensional intrinsic space, that is, the low-dimensional intrinsic space feature matrix LowDimF(N,d), wherein Y={y i ,i=1,…,N}.

[0037] Preferably, the distribution relationship between the target angle and the manifold surface is formed into a low-dimensional manifold surface through the three intrinsic dimensions in the obtained low-dimensional intrinsic feature space, a multi-angle target sample index is established, and the relationship between the target angle and the manifold surface law is explored.

[0038] When the target angle and the manifold surface law are explored, the exploration is performed in the form of synchronous dynamic playing of the image and the low-dimensional intrinsic feature space scatter plot.

[0039] According to the target multi-angle intrinsic feature mining system based on Laplacian eigenmap provided by the application, the system comprises:

[0040] Module M1: extracting target multi-angle SAR image features;

[0041] Module M2: constructing a target original high-dimensional observation feature space according to the multi-angle SAR image features;

[0042] Module M3: Laplace Eigen Mapping is performed on the original high-dimensional observation feature space to obtain a mapped low-dimensional Eigen feature space;

[0043] Module M4: a distribution relationship between a target angle and a manifold surface is obtained through the low-dimensional Eigen feature space.

[0044] Preferably, the SAR image features include basic pixel features, grayscale features, shape features, and texture features.

[0045] The SAR image features are obtained in a filtered image at a set scale.

[0046] Preferably, module M2 includes:

[0047] Module M2.1: maximum and minimum value normalization processing is performed on the multi-angle SAR image features respectively, and the calculation formula is as follows:

[0048]

[0049] wherein f(i) is an i-th feature value, f min is a minimum value among all feature values, f max is a maximum value among all feature values, and F(i) is a normalized feature value normalized to the range of [0, 1];

[0050] Module M2.2: all processed features are concatenated to obtain an original high-dimensional observation space feature matrix, denoted as HighDimF(N, M), wherein N is a sample number, and M is a feature dimension after concatenation.

[0051] Preferably, module M3 includes:

[0052] Module M3.1: the original high-dimensional observation space feature matrix is taken as a data set X in the original high-dimensional observation space, a neighborhood of each data point is calculated, and an adjacency graph is constructed.

[0053] Module M3.2: a weight matrix W is obtained by weighting the adjacency graph, as follows:

[0054]

[0055] wherein W ij represents a weight value matrix given to an edge connecting a near neighbor of each point in the adjacency graph constructed in the original high-dimensional observation space; e represents a natural base; t represents a heat kernel function coefficient; x i represents an i-th data point in the data set in the original high-dimensional observation space; x j represents a j-th data point in the data set in the original high-dimensional observation space; J i represents a sample point x iA set of neighbor points in the original high-dimensional observation space;

[0056] Module M3.3: calculating a degree matrix D according to the weight matrix, and the calculation formula is as follows:

[0057] D ii =∑ j W ij

[0058] Wherein, D ii represents a degree matrix of the weighted adjacent graph, and the degree matrix is in the form of a diagonal matrix, and the value on the diagonal line of the matrix is the sum of the weight values of the edges between each sample point and the neighbor points;

[0059] Module M3.4: obtaining a Laplacian matrix L by the difference between the degree matrix and the weight matrix, and the calculation formula is as follows:

[0060] L=D-W

[0061] Module M3.5: performing generalized eigenvalue decomposition on the Laplacian matrix L, and the calculation formula is as follows:

[0062] Ly=λDy

[0063] Taking the second smallest eigenvalue to the first d+1 smallest eigenvalues λ2,…,λ d+1 , corresponding to the eigenvectors U=[u2,…,u d+1 ], a low-dimensional embedding is obtained, and the low-dimensional embedding is represented as Wherein, λ represents an eigenvalue, y represents an eigenvector, d represents the first d dimensions obtained by sorting the eigenvalues in ascending order after eigenvalue decomposition, k represents a low-dimensional intrinsic feature dimension, and y i represents the eigenvector corresponding to the first 2-d+1 eigenvalues in the eigenvector of the i-th sample point.

[0064] Further, the data set Y in the low-dimensional intrinsic space is obtained, and the data set Y is a low-dimensional intrinsic space feature matrix LowDimF(N,d), wherein Y={y i ,i=1,…,N}.

[0065] Preferably, the target angle and the distribution relationship of the manifold surface are formed into a low-dimensional manifold surface by three intrinsic dimensions in the obtained low-dimensional intrinsic feature space, a multi-angle target sample index is established, and the relationship between the target angle and the manifold surface law is explored.

[0066] When the target angle and the manifold surface law are explored, the exploration is performed in the form of synchronous dynamic playing of the image and the low-dimensional intrinsic feature space scatter plot.

[0067] Compared with the prior art, the present application has the following beneficial effects:

[0068] 1. The present application faces multi-angle target SAR image samples, adopts Laplace feature mapping method to find low-dimensional manifold surface in high-dimensional observation feature space, and proves the consistency of the target angle change rule and the manifold surface rule, thereby laying a foundation for subsequent multi-angle target feature representation and feature matching application.

[0069] 2. The present application is aimed at the acquired SAR multi-angle target image data set, and explores the rule of angle and feature space intrinsic manifold surface from the perspective of image features. The exploring means proposed in the present application can be widely applied to further applications such as multi-angle target classification and identification.

[0070] 3. The present application adopts Laplace feature mapping as a nonlinear manifold learning method, which does not need to satisfy the linearization assumption condition of feature space, and is more suitable for natural images.

[0071] 4. When exploring the target angle and the manifold surface rule, the present application adopts the form of synchronous dynamic playing of image and low-dimensional intrinsic feature space scatter diagram, so that the rule of different angle targets corresponding to the low-dimensional manifold surface along the surface can be clearly observed. BRIEF DESCRIPTION OF DRAWINGS

[0072] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments with reference to the following drawings:

[0073] Figure 1 The present application is a flowchart.

[0074] Figure 2 The present application is a schematic diagram of exploring the low-dimensional manifold surface and the angle relationship of multi-angle targets. DETAILED DESCRIPTION

[0075] The present application will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present application. These all belong to the protection scope of the present application.

[0076] The present application faces multi-angle target SAR image samples, extracts suitable target features to construct original high-dimensional observation feature space, adopts Laplace feature mapping method to find low-dimensional manifold surface in high-dimensional observation feature space, and proves the consistency of the target angle change rule and the manifold surface rule, thereby laying a foundation for subsequent multi-angle target feature representation and feature matching application.

[0077] Example 1

[0078] According to the present invention, a method for mining multi-angle intrinsic features of a target based on Laplacian eigenmaps is provided, such as... Figure 1 As shown, it includes:

[0079] Step S1: Extract multi-angle SAR image features of the target. Considering that traditional statistical features based on SAR image distribution are no longer applicable to high-resolution SAR image targets, more structural, texture, and shape features suitable for high-resolution SAR image targets are introduced. Furthermore, to construct the most complete possible original high-dimensional observation feature space for the target, different SAR image features need to be considered separately. These features include: basic pixel features, grayscale features, shape features, and texture features. Specifically, these include basic pixel features reflecting the image, grayscale features reflecting the target's radiation intensity information (e.g., grayscale histograms), shape features reflecting the target's shape characteristics (e.g., edge orientation histograms, EOH), and texture features reflecting the characteristics between the target and the background (e.g., grayscale co-occurrence matrix (GLCM) and local binary mode (LBP) operators).

[0080] Considering the impact of speckle noise, which is unique to SAR images, on feature extraction, the image features used to construct the original high-dimensional observation feature space are obtained from filtered images at a certain scale.

[0081] Specifically, the extraction methods for different features are further described:

[0082] Basic pixel features

[0083] The multi-angle target image samples used are image blocks of the same pixel size, such as 128*128. For samples of different sizes, downsampling or upsampling can be used to normalize the image pixel size. For the normalized image samples, the same rules (such as row-wise or column-wise) can be used to flatten the two-dimensional matrix into a one-dimensional row vector, such as: I2D(128,128) becomes I1D(1,128*128). The basic pixel feature matrix composed of N multi-angle target image samples can be denoted as F_OrigPixel(N,128*128).

[0084] Gray-scale histogram

[0085] The grayscale histogram is calculated from multi-angle target image samples using the following formula:

[0086]

[0087] Where A is the grayscale coefficient, such as 255 for an 8-bit image; n is the total number of quantization bits in the histogram; p is the number of quantization bits for each level in n; and x is the grayscale value of a pixel in the image. The grayscale histogram feature matrix of N multi-angle target image samples can be denoted as F_Hist(N,n).

[0088] Edge Orientation Histogram (EOH)

[0089] Edge Orientation Histogram (EOH) is based on the statistical characteristics of image edges, and can better reflect the shape and edge information of the target. Each pixel on the edge corresponds to an edge gradient direction, and the edge can be regarded as being composed of edge pixel points of a specific direction. The statistics of the gradient direction of the edge pixel points can represent the shape of the target. First, the Canny operator is used to perform edge detection on the image sample to obtain an edge image, and then the normal vector direction angle of each edge pixel point is calculated according to the following formula after Gaussian filtering:

[0090]

[0091] where g(i,j) is the gray scale of the image after Gaussian filtering. The angle space of 360° is quantized into 8 directions with 45° as a unit, the statistics of the normal vector direction angle θ of each edge point falling in the 8 directions is calculated, and the edge orientation histogram is obtained. The edge orientation histogram feature matrix of N multi-angle target image samples can be denoted as F_Eoh(N,8).

[0092] Gray-Level Co-occurrence Matrix (GLCM)

[0093] First, the gray-level co-occurrence matrix of the multi-angle image sample is calculated, and then the contrast, correlation, energy and homogeneity feature values in the matrix are calculated according to the following formulas:

[0094] Contrast: Energy:

[0095] Correlation: Homogeneity:

[0096] where p(i,j) is the gray-level co-occurrence matrix. The gray-level co-occurrence matrix feature matrix of N multi-angle target image samples is denoted as F_Glcm(N,4).

[0097] Local Binary Pattern (LBP)

[0098] The local binary pattern (LBP) is in a window, taking the center pixel of the window as a threshold, comparing the gray values of adjacent pixels with it. If the value of the surrounding pixel is greater than the center pixel value, the position of the pixel point is marked as 1, otherwise 0. The histogram statistics of the encoding of all pixels in an image block (LBP type as the horizontal axis, the number of occurrences as the vertical axis) are obtained to obtain the LBP feature, which reflects the texture information of the image. The local binary pattern feature matrix of N multi-angle target image samples is denoted as F_Lbp(N,m), wherein m is the number of local binary pattern features.

[0099] Step S2: constructing a target original high-dimensional observation feature space according to the multi-angle SAR image features; step S2 includes: step S2.1: performing maximum and minimum value normalization processing on the multi-angle SAR image features respectively, and the calculation formula is as follows:

[0100]

[0101] Wherein f(i) is the i-th feature value, f min is the minimum value among all feature values, f max is the maximum value among all feature values, and F(i) is the normalized feature value normalized to the range of [0, 1]. The original high-dimensional observation space feature matrix composed of all features of N multi-angle target image samples in series is denoted as HighDimF(N,M), wherein N is the number of samples, and M is the feature dimension after series connection.

[0102] Step S2.2: concatenating all processed features to obtain the original high-dimensional observation space feature matrix denoted as HighDimF(N,M), wherein N is the number of samples, and M is the feature dimension after series connection.

[0103] Step S3: under the optimization criterion of maintaining the neighborhood characteristics of the feature space, performing Laplacian Eigenmaps on the original high-dimensional observation feature space to obtain a mapped low-dimensional intrinsic feature space; the Laplacian Eigenmaps (LE) maintains the local neighborhood relationship, and the criterion can be described as: the data points with close distance in the original high-dimensional observation space should still maintain close distance after mapping.

[0104] Specifically, first, an adjacency graph G is constructed: the original high-dimensional observation space feature matrix is taken as a data set X in the original high-dimensional observation space, the neighborhood of each data point is calculated, and thus the adjacency graph is constructed. That is, the data set X in the original high-dimensional observation space is the original high-dimensional observation space feature matrix HighDimF(N,M) composed of all features of N multi-angle target image samples in series.

[0105] Then, a weighted adjacency graph W is constructed: the weighting of the adjacency graph yields the weight matrix W, as shown in the following equation:

[0106]

[0107] Among them, W ij This represents the weight matrix assigned to the edges connecting the nearest neighbors of each point in the adjacency graph constructed in the original high-dimensional observation space; e represents the natural base; t represents the heat kernel function coefficient, empirically set to 1; x i x represents the i-th data point in the dataset within the original high-dimensional observation space; j J represents the j-th data point in the dataset within the original high-dimensional observation space; i Represents sample point x i The set of nearest neighbors in the original high-dimensional observation space;

[0108] Next, the degree matrix D is calculated based on the weight matrix, using the following formula:

[0109] D ii =∑ j W ij

[0110] Where D ii The metric matrix represents the weighted adjacency graph. The metric matrix is ​​in the form of a diagonal matrix, and the values ​​on the diagonal are the sum of the weights of the edges between each sample point and its nearest neighbor.

[0111] Then, the Laplacian matrix L is obtained by subtracting the degree matrix from the weight matrix. The calculation formula is as follows:

[0112] L=DW

[0113] Finally, solve for the d-dimensional embedding Y: solve the optimization problem using the LE criterion. Depend on This optimization problem can be transformed into a generalized eigenvalue decomposition of the Laplacian matrix L, calculated as follows:

[0114] Ly=λDy

[0115] Take its second smallest eigenvalue to the first d+1 smallest eigenvalues ​​λ2,…,λ d+1 The corresponding eigenvectors U = [u2, ..., u] d+1 The low-dimensional embedding representation is obtained as follows: Where λ represents the eigenvalue, y represents the eigenvector, d represents the first d dimensions obtained by sorting the eigenvalues ​​in ascending order after eigenvalue decomposition, and k represents the dimension of the low-dimensional intrinsic features. iY represents the feature vector corresponding to the first 2~d+1 feature values in the i-th sample point feature vector; and then the data set Y in the low-dimensional intrinsic space is obtained, that is, the low-dimensional intrinsic space feature matrix LowDimF(N, d), wherein Y={y i ,i=1,…,N}.

[0116] Step S4: obtaining the distribution relationship between the target angle and the manifold surface through the low-dimensional intrinsic feature space. The distribution relationship between the target angle and the manifold surface is formed by the three intrinsic dimensions in the obtained low-dimensional intrinsic feature space to form a low-dimensional manifold surface, to establish a multi-angle target sample index, and to explore the relationship between the target angle and the manifold surface law; when exploring the relationship between the target angle and the manifold surface law, the multi-angle target feature analysis platform prepared according to the application is adopted in the form of synchronous dynamic playing of the image and the low-dimensional intrinsic feature space scatter plot, so that the law of the different angle targets along the surface on the low-dimensional manifold surface can be clearly observed.

[0117] Further, the application is further described by taking the multi-angle sample library in the MSTAR data set as a research example:

[0118] Step one, target multi-angle SAR image feature extraction, respectively obtaining the basic pixel feature matrix F_OrigPixel(N, 128*128) composed of N multi-angle target image samples, the gray histogram feature matrix F_Hist(N, n), the edge direction histogram feature matrix F_Eoh(N, 8), the gray co-occurrence matrix feature matrix F_Glcm(N, 4), and the local binary pattern feature matrix F_Lbp(N, m).

[0119] Step two, original high-dimensional observation feature space establishment, respectively performing maximum and minimum value normalization processing on each type of feature in step one, and concatenating the features to a high-dimensional feature vector to obtain the original high-dimensional observation space feature matrix HighDimF(N, M) composed of all normalized features of the N multi-angle target image samples.

[0120] Step three, Laplace eigenmap dimension reduction, performing Laplace eigenmap dimension reduction on the original high-dimensional observation space feature matrix HighDimF(N, M) in step two to obtain the low-dimensional intrinsic space feature matrix LowDimF(N, d).

[0121] Step four, as shown in Figure 2 shown, the relationship between the low-dimensional manifold surface and the multi-angle target is explored, the first three-dimensional scatter plot of the low-dimensional intrinsic space feature matrix LowDimF(N, d) obtained in step three is taken to obtain the manifold surface of the multi-angle target sample in the low-dimensional intrinsic space, and the change law of the target along with the change of the angle on the manifold surface is observed through the point index on the manifold surface and the image sample association.

[0122] Embodiment two

[0123] The application further provides a target multi-angle intrinsic feature mining system based on Laplace feature mapping. A person skilled in the art can realize the target multi-angle intrinsic feature mining system based on Laplace feature mapping by executing the steps of the target multi-angle intrinsic feature mining method based on Laplace feature mapping, that is, the target multi-angle intrinsic feature mining method based on Laplace feature mapping can be understood as a preferred embodiment of the target multi-angle intrinsic feature mining system based on Laplace feature mapping.

[0124] According to the application, a target multi-angle intrinsic feature mining system based on Laplace feature mapping is provided, which comprises:

[0125] Module M1: extracting target multi-angle SAR image features; the SAR image features comprise basic pixel features, gray features, shape features and texture features; the SAR image features are obtained in a filtered image at a set scale.

[0126] Module M2: constructing a target original high-dimensional observation feature space according to the multi-angle SAR image features; module M2 comprises:

[0127]

[0128] wherein f(i) is an ith feature value, f min is the minimum value in all feature values, f max is the maximum value in all feature values, and F(i) is a normalized feature value normalized to the range of [0, 1].

[0129] Module M2.2: concatenating all processed features to obtain an original high-dimensional observation feature matrix, denoted as HighDimF(N, M), wherein N is the number of samples, and M is the feature dimension after concatenation.

[0130] Module M3: performing Laplace feature mapping on the original high-dimensional observation feature space to obtain a mapped low-dimensional intrinsic feature space; preferably, module M3 comprises:

[0131] Module M3.1: taking the original high-dimensional observation feature matrix as a data set X in the original high-dimensional observation space, calculating the neighborhood of each data point to construct an adjacency graph. Module M3.2: weighting the adjacency graph to obtain a weight matrix W, as follows:

[0132]

[0133] wherein W ij represents the weight matrix of the edges connecting the near neighbor points of each point in the constructed adjacent graph in the original high-dimensional observation space; e represents the natural base; t represents the kernel function coefficient, and the experience value is 1; x i represents the i th data point in the data set in the original high-dimensional observation space; x j represents the j th data point in the data set in the original high-dimensional observation space; J i represents the sample point x i in the original high-dimensional observation space; x

[0134] Module M3.3: calculating the degree matrix D according to the weight matrix, and the calculation formula is as follows:

[0135] D ii =∑ j W ij

[0136] wherein D ii represents the degree matrix of the weighted adjacent graph, and the degree matrix is in the form of a diagonal matrix, and the value on the diagonal line of the matrix is the sum of the weights of the edges between each sample point and the near neighbor points;

[0137] Module M3.4: obtaining the Laplacian matrix L by the difference between the degree matrix and the weight matrix, and the calculation formula is as follows:

[0138] L=D-W

[0139] Module M3.5: performing the generalized eigenvalue decomposition of the Laplacian matrix L, and the calculation formula is as follows:

[0140] Ly=λDy

[0141] taking the second smallest eigenvalue to the first d+1 smallest eigenvalues λ2,…,λ d+1 , corresponding to the eigenvectors U=[u2,…,u d+1 ], to obtain the low-dimensional embedding representation wherein λ represents the eigenvalue, y represents the eigenvector, d represents the first d dimensions obtained by sorting the eigenvalues in ascending order after the eigenvalue decomposition, k represents the low-dimensional intrinsic feature dimension, y i represents the eigenvector corresponding to the first 2~d+1 eigenvalues in the eigenvector of the i th sample point; and further obtaining the data set Y in the low-dimensional intrinsic space, that is, the low-dimensional intrinsic space feature matrix LowDimF(N,d), wherein Y={y i ,i=1,…,N}.

[0142] Module M4: obtaining the distribution relationship between the target angle and the manifold surface through the low-dimensional intrinsic feature space. The distribution relationship between the target angle and the manifold surface is formed into a low-dimensional manifold surface through three intrinsic dimensions in the obtained low-dimensional intrinsic feature space, a multi-angle target sample index is established, and the relationship between the target angle and the manifold surface is explored. When exploring the relationship between the target angle and the manifold surface, the image and the low-dimensional intrinsic feature space scatter plot are synchronously and dynamically played.

[0143] Those skilled in the art know that, in addition to implementing the system, device and each module thereof provided by the present application in the form of pure computer readable program code, the same program can also be realized in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers, etc. by logically programming the method steps. Therefore, the system, device and each module thereof provided by the present application can be considered as a hardware component, and the modules included therein for realizing various programs can also be considered as structures in the hardware component; the modules for realizing various functions can also be considered as both software programs for realizing methods and structures in the hardware component.

[0144] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be arbitrarily combined with each other without conflict.

Claims

1. A target multi-angle intrinsic feature mining method based on Laplacian Eigenmap, characterized in that, Comprising: Step S1: extracting target multi-angle SAR image features; the SAR image features include: basic pixel features, gray features, shape features, and texture features; the SAR image features are obtained in a filtered image under a set scale; Step S2: constructing a target original high-dimensional observation feature space according to the multi-angle SAR image features; step S2 comprises: Step S2.1: performing maximum and minimum value normalization processing on the multi-angle SAR image features respectively, and the calculation formula is as follows: where f(i) is the ith feature value, f min is the minimum value among all feature values, f max is the maximum value among all feature values, and F(i) is a normalized feature value normalized to the range [0, 1]; Step S2.2: concatenating all the processed features to obtain an original high-dimensional observation space feature matrix, denoted as HighDimF(N, M), wherein N is the number of samples, and M is the feature dimension after concatenation; Step S3: performing Laplace feature mapping on the original high-dimensional observation feature space to obtain a mapped low-dimensional intrinsic feature space feature matrix; step S3 comprises: Step S3.1: taking the original high-dimensional observation space feature matrix as a data set X in the original high-dimensional observation space, calculating the neighborhood of each data point, and thereby constructing an adjacency graph; Step S3.2: Weighting the adjacency graph results in a weight matrix as follows: wherein, represents a weight matrix assigned to edges connecting each point's near neighbors in the constructed adjacency graph in the original high-dimensional observation space; represents a natural base; t represents a thermal kernel coefficient; represents the i-th data point in the dataset in the original high-dimensional observation space; represents the j-th data point in the dataset in the original high-dimensional observation space; represents a sample point a set of near neighbors in the original high-dimensional observation space; Step S3.3: Compute the degree matrix from the weight matrix The formula is as follows: wherein denotes the metric matrix of the weighted adjacency graph, which is in the form of a diagonal matrix, the values on the diagonal of the matrix being the sum of the weights of the edges between each sample point and its nearest neighbors; Step S3.4: Obtain the Laplacian matrix by the difference between the degree matrix and the weight matrix The calculation formula is as follows: Step S3.5: Compute the generalized eigenvalues of the Laplacian matrix with the formula: Take the second smallest eigenvalue to the first smallest eigenvalue corresponding eigenvector , get low-dimensional embedding representation as ; wherein, indicates the eigenvalue, y indicates the eigenvector, d indicates the first d dimension obtained by sorting the eigenvalues in ascending order after eigenvalue decomposition, k indicates the low-dimensional intrinsic feature dimension, indicates the eigenvector corresponding to the first 2-d+1 eigenvalues in the eigenvector of the i-th sample point Further, the data set Y in the low-dimensional eigenspace is obtained, that is, the low-dimensional eigenspace feature matrix LowDimF(N, d), wherein, ; Step S4: obtaining a distribution relationship between a target angle and a manifold surface through the low-dimensional intrinsic feature space; the distribution relationship between the target angle and the manifold surface forms a low-dimensional manifold surface through three intrinsic dimensions in the obtained low-dimensional intrinsic feature space, establishes a multi-angle target sample index, and explores the relationship between the target angle and the manifold surface law; When the relationship between the target angle and the manifold surface law is explored, it is performed in the form of synchronous dynamic playing of an image and a low-dimensional intrinsic feature space scatter plot.

2. A target multi-angle intrinsic feature mining system based on Laplacian Eigenmap, characterized in that, Comprising: Module M1: used for extracting target multi-angle SAR image features; the SAR image features include: basic pixel features, gray features, shape features, and texture features; the SAR image features are obtained in a filtered image under a set scale; Module M2: used for constructing a target original high-dimensional observation feature space according to the multi-angle SAR image features; module M2 comprises: Module M2.1: used for performing maximum and minimum value normalization processing on the multi-angle SAR image features respectively, and the calculation formula is as follows: where f(i) is the ith feature value, f min is the minimum value among all feature values, f max is the maximum value among all feature values, F(i) is the normalized feature value normalized to the range [0, 1]; Module M2.2: used for concatenating all the processed features to obtain an original high-dimensional observation space feature matrix, denoted as HighDimF(N, M), wherein N is the number of samples, and M is the feature dimension after concatenation; Module M3: used for performing Laplace feature mapping on the original high-dimensional observation feature space to obtain a mapped low-dimensional intrinsic feature space feature matrix; module M3 comprises: Module M3.1: used for taking the original high-dimensional observation space feature matrix as a data set X in the original high-dimensional observation space, calculating the neighborhood of each data point, and thereby constructing an adjacency graph; Module M3.2: for weighting said adjacency graph to obtain a weight matrix as follows: wherein, represents a weight matrix assigned to edges connecting each point's near neighbors in the constructed adjacency graph in the original high-dimensional observation space; represents a natural base; t represents a thermal kernel coefficient; represents the i-th data point in the dataset in the original high-dimensional observation space; represents the j-th data point in the dataset in the original high-dimensional observation space; represents a sample point a set of near neighbors in the original high-dimensional observation space; Module M3.3: for calculating a degree matrix from the weight matrix The calculation formula is as follows: wherein denotes the metric matrix of the weighted adjacency graph, which is in the form of a diagonal matrix, the values on the diagonal of the matrix being the sum of the weights of the edges between each sample point and its nearest neighbors; Module M3.4: for obtaining a Laplacian matrix by the difference between the degree matrix and the weight matrix The calculation formula is as follows: Module M3.5: for the generalized eigenvalue decomposition of the Laplacian matrix with the following formula: Take the second smallest eigenvalue to the first smallest eigenvalue corresponding eigenvector , get low-dimensional embedding representation as ; wherein, indicates the eigenvalue, y indicates the eigenvector, d indicates the first d dimension obtained by sorting the eigenvalues in ascending order after eigenvalue decomposition, k indicates the low-dimensional intrinsic feature dimension, indicates the eigenvector corresponding to the first 2~d+1 eigenvalues in the eigenvector of the i-th sample point. Further, the data set Y in the low-dimensional eigenspace is obtained, that is, the low-dimensional eigenspace feature matrix LowDimF(N, d), wherein, ; Module M4: used for obtaining a distribution relationship between a target angle and a manifold surface through the low-dimensional intrinsic feature space; the distribution relationship between the target angle and the manifold surface forms a low-dimensional manifold surface through three intrinsic dimensions in the obtained low-dimensional intrinsic feature space, establishes a multi-angle target sample index, and explores the relationship between the target angle and the manifold surface law; The target angle and the manifold surface law are searched through the forms of synchronous dynamic playing of the image and low-dimensional intrinsic feature space scatter plot.

Images