A multi-frame radar echo land-sea segmentation method based on spectral clustering algorithm

Abstract

The present invention relates to a multi-frame radar echo sea-land segmentation method based on a spectral clustering algorithm, comprising: acquiring the current frame echo data information of a sea surface scanning area, and performing phase unwrapping on the current frame echo data information to obtain unwrapped phase information; constructing an undirected graph of each resolution cell by using the unwrapped phase information of the same resolution cell in multi-frame echo data, and constructing a Laplacian matrix for each undirected graph; combining the Ratio Cut criterion in the spectral clustering algorithm, calculating the second smallest eigenvalue of the Laplacian matrix as the phase similarity measure of the resolution cell, and obtaining a measure matrix of all resolution cells; using the K-means clustering algorithm, and combining the measure matrix to segment the sea surface scanning area into a sea clutter area and a land clutter area to obtain an initial sea-land segmentation result; performing morphological filtering on the initial sea-land segmentation result to obtain a final sea-land segmentation result. This method solves the problems existing in the single-frame radar echo sea-land segmentation algorithm, such as poor performance, low quality, and low result stability.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and particularly relates to a multi-frame radar echo land-sea segmentation method based on a spectral clustering algorithm. Background Art

[0002] Under the background of sea clutter, the received echo is composed of land clutter from land areas, sea clutter from ocean areas, target clutter, etc., and the composition of the echo is very complex. The purpose of land-sea segmentation is to separate the land clutter and sea clutter in the received echo. This is because for subsequent target detection, the echo power of land clutter and island reef clutter is relatively large, which will cause great interference to target detection and an ideal detection result cannot be obtained. After separating interference factors such as land clutter, only sea surface target detection is performed on the ocean area, which can avoid the influence of interference factors on subsequent processing. After land-sea segmentation, the process of target detection can effectively reduce the processing time during detection, improve efficiency, and avoid wasting time in unnecessary processes. Land-sea scene segmentation plays an important role in detection, and the quality of its segmentation determines whether sea clutter can be accurately distinguished from other clutter. If the segmentation result is inaccurate, the range of subsequent radar target detection will be incorrect, directly affecting the accuracy of target detection.

[0003] In recent years, for the land-sea segmentation technology under the background of sea clutter, many scholars and researchers have applied some image segmentation theories to solve the land-sea segmentation problem. These methods are mainly based on Synthetic Aperture Radar (SAR) images or satellite images. However, these algorithms perform land-sea segmentation through grayscale images rather than radar echo characteristics. In the scanning mode, due to the limitation of the actual aperture of the radar antenna, the azimuth resolution is much lower than that of SAR. In addition, the coherent integration time in the scanning mode is dozens of milliseconds shorter. The land-sea segmentation of a maritime reconnaissance radar in the scanning mode essentially does not belong to the category of image processing. When the detection scene becomes complex, relying solely on extracting image features for segmentation, the image features often cannot clearly distinguish land and sea clutter, and the performance of land-sea segmentation will decline. Therefore, in the context of a relatively complex detection scene, it is an ideal way to extract measures from radar echo data to distinguish land and sea clutter through characteristic analysis. The commonly used land-sea method based on the characteristics of a single-frame radar echo is to analyze the measured data and, based on the characteristic that the Doppler spectral width of sea clutter is greater than that of land clutter, propose a phase linearity scheme to describe the difference in spectral width between the two. By calculating the phase linearity of the radar echo as a measure and using binary segmentation to distinguish the two, an initial land-sea segmentation result is obtained.

[0004] However, the land-sea segmentation method based on the characteristics of single-frame radar echoes only relies on the radar echo data of a single scanning period, resulting in a large difference in the similarity of land-sea clutter scenes among multiple groups of radar echo data, thus showing problems such as poor land-sea segmentation performance and low stability of segmentation results. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] An embodiment of the present invention provides a multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm, including the steps of:

[0007] S1. Obtain the current frame echo data information of the sea surface scanning area, and perform phase unwrapping on the current frame echo data information to obtain unwrapped phase information;

[0008] S2. Use the unwrapped phase information of the same resolution cell in the multi-frame echo data information to construct an undirected graph of the resolution cell, and construct a Laplacian matrix for each undirected graph, where the number of frames of the multi-frame echo data information is greater than or equal to 5 frames;

[0009] S3. Combine the Ratio Cut criterion in the spectral clustering algorithm, calculate the second smallest eigenvalue of the Laplacian matrix as the phase similarity measure of the resolution cell, and obtain the measure matrix of all resolution cells;

[0010] S4. Use the K-means clustering algorithm, and combine the measure matrix to divide the sea surface scanning area into a sea clutter area and a land clutter area to obtain an initial land-sea segmentation result;

[0011] S5. Perform morphological filtering on the initial land-sea segmentation result to obtain a land-sea segmentation result.

[0012] In an embodiment of the present invention, step S2 includes:

[0013] S21. Take the unwrapped phase information of the resolution cell as the data of the vertex in the undirected graph, and take the resolution cells under the same wave position and the same range gate in one frame of echo as a vertex of the undirected graph. Obtain several vertices from the resolution cells under the same wave position and the same range gate in several frames of echo, and connect the several vertices to obtain the undirected graph;

[0014] S22. Use the weight matrix of the undirected graph and the degree matrix of the vertex to construct the Laplacian matrix.

[0015] In an embodiment of the present invention, the vertex expression of the undirected graph is:

[0016] {φ 1 (m,n), φ 2 (m,n), …, φ i (m,n)}

[0017] Among them, φ i (m,n) represents the unwrapped phase sequence of the resolution cell at the m-th range bin and the n-th Doppler bin in the i-th frame of echo.

[0018] In an embodiment of the present invention, the Laplacian matrix is:

[0019] L = D - W

[0020] Among them, L is the Laplacian matrix, D is the degree matrix, w ij is the similarity degree between vertex v i and vertex v j , D ii is the degree of each vertex, and the degree of a vertex represents the number of edges associated with the vertex. W is the weight matrix. e ij is the edge connecting vertices, and E is the set of edges of the undirected graph, E = {e ij = <v i , v j > | v i , v j ∈ V}, and V is the set of vertices.

[0021] In an embodiment of the present invention, step S3 includes:

[0022] S31. Calculate the minimized cost function when optimally cutting the undirected graph into several subgraphs according to the weight matrix of the undirected graph;

[0023] S32. Improve the minimized cost function according to the constraint of the objective function of the Ratio Cut criterion in the spectral clustering algorithm to obtain the objective function;

[0024] S33. Combine the Laplacian matrix, the eigenvectors of the Laplacian matrix, and the properties of the non-normalized Laplacian matrix to optimize the objective function into a constrained optimization problem model;

[0025] S34. Refer to the Rayleigh - Ritz theorem, solve the second - smallest eigenvalue of the Laplacian matrix as the solution of the constrained optimization problem model, and use the second - smallest eigenvalue of the Laplacian matrix as the phase similarity measure of the resolution cell, thereby obtaining the measure matrix of all resolution cells.

[0026] In one embodiment of the present invention, the minimized cost function is:

[0027]

[0028] where B1, …, B k are the divided subgraphs, is the complement of B i , represents the sum of the weights of the edges connecting B i and , and k represents the number of divided subgraphs.

[0029] In one embodiment of the present invention, the objective function is:

[0030]

[0031] where |B i | represents the number of vertices in the subgraph B i . When k = 2, the objective function is denoted as

[0032] In one embodiment of the present invention, step S33 includes:

[0033] Define the feature vector as:

[0034] f = (f1, f2, …, f n ) ∈ R n

[0035]

[0036] where B is a subgraph, V is the set of vertices, |B| represents the number of vertices in the subgraph B, and R n is an n-dimensional vector space;

[0037] Substitute the feature vector into the properties of the unnormalized Laplacian matrix to obtain the equation:

[0038]

[0039] where |B| represents the number of vertices in the subgraph B, |V| represents the number of vertices in the original graph, and both |V| and |B| are constants, n is the number of vertices, and L is the Laplacian matrix; from the above equation, it can be obtained that minimizing the objective function is equivalent to minimizing f T Lf;

[0040] Combined with minimizing f T Lf, the objective function is optimized into a constrained optimization problem model:

[0041] Satisfy

[0042] f T Lf = λf T f = λn

[0043] Where n is the number of vertices and λ is the eigenvalue of the Laplacian matrix.

[0044] In an embodiment of the present invention, step S4 includes:

[0045] S41. Using the K-means clustering algorithm, preliminarily divide the measure matrix into a first subclass and a second subclass, and calculate the first measure mean of the first subclass and the second measure mean of the second subclass;

[0046] S42. Compare the first measure mean and the second measure mean, and determine the sea clutter region corresponding to the smaller value of the first measure mean and the second measure mean, and determine the land clutter region corresponding to the larger value of the first measure mean and the second measure mean.

[0047] In an embodiment of the present invention, step S5 includes:

[0048] Use the operator of binary morphology to perform morphological filtering on the sea-land preliminary segmentation result to eliminate the isolated pixel clusters in the sea clutter region and fill the holes in the land clutter region to obtain the sea-land segmentation result of the sea surface scanning region.

[0049] Compared with the prior art, the beneficial effects of the present invention:

[0050] The multi-frame radar echo sea-land segmentation method of the present invention combines the radar echo data of multiple scanning cycles, constructs an undirected graph for each resolution unit, and calculates the second smallest eigenvalue of the Laplacian matrix corresponding to the undirected graph as the phase similarity measure of the resolution unit, realizing the sea-land segmentation of multi-frame radar echoes, solving the problems of poor segmentation performance, low quality, and low result stability caused by limited data samples in the traditional sea-land segmentation algorithm based on single-frame radar echoes, having better robustness to outliers, and improving the consistency of the sea-land segmentation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow schematic diagram of a multi-frame radar echo sea-land segmentation method based on the spectral clustering algorithm provided by an embodiment of the present invention;

[0052] Figure 2 It is a schematic diagram of an undirected graph provided by an embodiment of the present invention;

[0053] Figure 3A comparison chart of the phase linearity method provided by the embodiments of the present invention and the initial segmentation result of the multi-frame radar echo sea-land segmentation method based on the spectral clustering algorithm;

[0054] Figure 4 A morphological filtering result chart of the phase linearity method for multiple scanning periods provided by the embodiments of the present invention and the multi-frame radar echo sea-land segmentation method based on the spectral clustering algorithm;

[0055] Figure 5 A Hausdorff distance chart of the phase linearity method provided by the embodiments of the present invention and the multi-frame radar echo sea-land segmentation method based on the spectral clustering algorithm. Detailed implementation manners

[0056] The following further describes the present invention in detail with reference to specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0057] Embodiment 1

[0058] Please refer to Figure 1 , Figure 1 A flowchart of a multi-frame radar echo sea-land segmentation method based on the spectral clustering algorithm provided by the embodiments of the present invention. The multi-frame radar echo sea-land segmentation method based on the spectral clustering algorithm includes the following steps:

[0059] S1. Obtain the current frame echo data information of the sea surface scanning area, and perform phase unwrapping on the current frame echo data information to obtain the unwrapped phase information.

[0060] Specifically, use the radar transmitter to send signals to the sea surface, and use the radar receiver to receive the echo data reflected by the sea surface scanning area to obtain the radar echo data; the radar echo data is divided into pure clutter data and target echo data, and the pure clutter data is further divided into sea clutter data and ground clutter data. Assume that in one frame of echo, there are M range cells, N wave positions, and the number of accumulated pulses is K. Each element value in the radar echo data is a complex number. Due to the periodicity of the phase of the complex number, the obtained phase value is limited to the range of [-π, π), and the directly obtained complex phase is called the wrapped phase, and it is necessary to perform unwrapping on it to obtain the true phase.

[0061] Assume that for a certain wave position, in one scan, if the number of echo sequence pulses is K, since the radar echo data is a complex number, Q represents the imaginary part and I represents the real part. The wrapped phase sequence of each resolution cell can be expressed as:

[0062]

[0063] where I represents the real part data of the radar echo data, and Q represents the imaginary part data of the radar echo data.

[0064] Within a resolution cell, there is a wrapped phase sequence containing K wrapped phase values

[0065]

[0066] Judge the wrapped phase sequence The difference from the true phase sequence φ Due to the periodicity of complex phases, the difference between the wrapped phase and the true phase value is an integer multiple of 2π. If φ(k + 1) = φ(k) + δ(k), then φ(k + 1) = φ(k) + δ(k). If φ(k + 1) = φ(k) + δ(k), then φ(k + 1) = φ(k) + δ(k). This is executed in a loop until after taking the absolute value and then rounding φ(k + 1) = φ(k) + δ(k); in the case where there is noise in the radar echo, when the phase difference is close to π, phase ambiguity may occur. Therefore, we select a fixed value λ to further judge φ(k + 1) = φ(k) + δ(k) after addition and subtraction: If φ(k + 1) = φ(k) + δ(k), then φ(k + 1) = φ(k) + δ(k). If φ(k + 1) = φ(k) + δ(k), then φ(k + 1) = φ(k) + δ(k); then through calculation, the true phase value is obtained as: φ(k + 1) = φ(k) + δ(k).

[0067] Furthermore, by performing phase unwrapping on the current frame of echo data information, an unwrapped phase sequence is obtained, and the part of the unwrapped phase sequence used to construct an undirected graph is used as the unwrapped phase information of the knot.

[0068] S2. Use the unwrapped phase information of the same resolution cell in the multi-frame echo data information to construct an undirected graph of the resolution cell, and construct a Laplacian matrix for each undirected graph, where the number of frames of the multi-frame echo data information is greater than or equal to 5 frames. Specifically, it includes the steps:

[0069] S21. Use the unwrapped phase information of the resolution cell as the data of the vertices in the undirected graph, and use the resolution cells of the same wave position and the same range gate in one frame of echo as a vertex of the undirected graph. Several vertices are obtained from the resolution cells of the same wave position and the same range gate in several frames of echo, and several vertices are connected to obtain an undirected graph.

[0070] Please refer to Figure 2 , Figure 2 which is a schematic diagram of an undirected graph provided by an embodiment of the present invention. As Figure 2 shown, the edges connecting vertex to vertex have no direction, and the weight value of the edge between two points is fixed, that is, the weight w i of the edge connecting v j and v ij = w ji , and this spectral graph is an undirected graph.

[0071] In the undirected graph constructed in this embodiment, vertices correspond to the phase sequences of a certain resolution cell in different frame echoes. It can be understood that for the same resolution cell, an undirected graph is formed, and for multiple resolution cells, multiple undirected graphs are formed.

[0072] Furthermore, to enable subsequent measures to accurately describe the similarity between data points, the number of vertices in the undirected graph cannot be too small. Therefore, after receiving the fifth frame echo, the data is processed. That is to say, the number of frames of radar echo data is greater than or equal to 5 frames, and the number of vertices in each undirected graph is greater than or equal to 5.

[0073] Specifically, an undirected graph is constructed using the unwrapped phase sequence of vertices, and the vertex expression is:

[0074] {φ 1 (m,n), φ 2 (m,n), …, φ i (m,n)}(3)

[0075] where φ i (m,n) represents the unwrapped phase sequence of the resolution cell of the m-th range cell and the n-th wave position in the i-th frame echo.

[0076] S22. Construct a Laplacian matrix using the weight matrix and degree matrix of the vertices of the undirected graph.

[0077] Specifically, let G = (V, E) be an undirected graph, where V is the set of vertices and E = {e ij = <v i , v j > | v i , v j ∈ V} is the set of edges. The greater the weight of the edge e ij connecting two points, the more similar the two points v i and v j . The weight matrix is denoted as W, and the weight matrix is a square matrix whose dimension is the same as the number of vertices in the undirected graph. For the undirected graph G, the adjacency matrix and the weight matrix can be used to represent the connection relationship and similarity degree between points in it.

[0078] Let the matrix W be the weight matrix of the undirected graph G. The undirected graph similarity matrix W is obtained through the similarity between two data, and its definition is given by the following formula:

[0079]

[0080] where w ij is an element in the weight matrix, which represents the similarity between vertex v i and vertex vj The similarity degree between, e ij is the edge connecting vertices, and E is the set of edges of the undirected graph, E = {e ij = <v i , v j > | v i , v j ∈ V}, where V is the set of vertices.

[0081] In the spectral clustering algorithm, the weight matrix W is usually defined by the similarity matrix. The elements in the weight matrix W can be represented by the similarity between vertex v i and v j as follows:

[0082]

[0083] where σ is the scale factor, and the scale factor σ = 3.

[0084] The degree of a vertex in an undirected graph refers to the sum of the weights of the edges connected to the vertex in the undirected graph. The definition of the degree matrix D is as follows:

[0085]

[0086] where w ij is the similarity degree between vertex v i and vertex v j , and D ii is the degree of each vertex. The degree of a vertex represents the number of edges associated with the vertex.

[0087] The Laplacian matrix L of the undirected graph G is defined as:

[0088] L = D - W (7)

[0089] where L is the Laplacian matrix, D is the degree matrix, and W is the weight matrix.

[0090] S3. Combining with the Ratio Cut criterion in the spectral clustering algorithm, calculate the second smallest eigenvalue of the Laplacian matrix as the phase similarity measure of the resolution unit, and obtain the measure matrix of all resolution units.

[0091] Specifically, due to land echoes, the phase correlation degree between consecutive multi-frame echo data is greater than that of sea clutter. When multiple undirected graphs formed by multi-frame echo phase data points are divided into two subgraphs, if the resolution unit is a land echo, the sum of the weights of the edges connected in each subgraph of the two subgraphs will be greater than the weight sum when the resolution unit is a sea echo. Therefore, the cut value of the graph can be used as a measure to describe the phase similarity between multi-frame radar echoes.

[0092] Step S3 specifically includes:

[0093] S31. Calculate the minimized cost function when optimally cutting the undirected graph into several subgraphs according to the weight matrix of the undirected graph.

[0094] Specifically, when optimally cutting the undirected graph, the minimized cost function is:

[0095]

[0096] where B1, …, B k are the divided subgraphs, is the complement of B i , represents the sum of the weights of the edges connecting B i and , and k represents the number of divided subgraphs.

[0097] S32. Improve the minimized cost function according to the constraint of the objective function of the Ratio Cut criterion in the spectral clustering algorithm to obtain the objective function.

[0098] Specifically, the meaning of the cost function is that when dividing the graph G into k subgraphs, the total weight of the edges connecting each subgraph that need to be deleted during the division. In order to achieve the purpose of the spectral clustering algorithm, the weights of the edges between different subgraphs should be as small as possible, and the weights of the edges within the subgraph should be as large as possible. The purpose can be achieved by minimizing cut(B1, …, B k ). However, if only considering minimizing cut(B1, …, B k ), an imbalance problem will occur.

[0099] To avoid the problem of unbalanced minimum cut graph caused by only considering minimizing the cost function, in the objective function Ratio Cut of the ratio cut criterion under the spectral clustering algorithm, the scale of each subgraph is constrained, that is, in order to make each subgraph have a sufficient size and considering the number of vertices in each subgraph, the objective function should make B1, …, B k as large as possible. Therefore, the objective function is improved to:

[0100]

[0101] where |B i | represents the number of vertices in the subgraph B i . When k = 2, the objective function is denoted as

[0102] It can be seen from the above Ratio Cut function that it fully considers the scale of the subgraph, that is, the number of vertices in the subgraph. If |B i|Too small, the value of the cost function will become larger, and the obtained point set is not the optimal undirected graph segmentation result at this time.

[0103] S33. Combine the Laplacian matrix, the eigenvectors of the Laplacian matrix, and the properties of the unnormalized Laplacian matrix to optimize the objective function into an optimization problem model with constraints.

[0104] Specifically, given a point set Define the eigenvector f = (f1, f2,..., f n ) ∈ R n , and satisfy the following conditions:

[0105]

[0106] Among them, B is a subgraph, V is the set of vertices, |B| represents the number of vertices in the subgraph B, and R n is an n-dimensional vector space.

[0107] Substitute the eigenvector into the properties of the unnormalized Laplacian matrix to obtain the equation:

[0108]

[0109] Among them, |B| represents the number of vertices in the subgraph B, |V| represents the number of vertices in the original graph, and both |V| and |B| are constants, n is the number of vertices, and L is the Laplacian matrix.

[0110] It can be seen from equation (11) that since |V| is a constant, minimizing the objective function is equivalent to minimizing f T Lf.

[0111] Furthermore, combine minimizing f T Lf to optimize the objective function into an optimization problem model with constraints:

[0112]

[0113] Among them, n is the number of vertices, λ is the eigenvalue of the Laplacian matrix L, and f is the eigenvector of the Laplacian matrix L.

[0114] S34. Refer to the Rayleigh - Ritz theorem, solve the second - smallest eigenvalue of the Laplacian matrix as the solution of the optimization problem model with constraints, and use the second - smallest eigenvalue of the Laplacian matrix as the phase similarity measure of the discrimination unit, so as to obtain the measure matrix of all discrimination units.

[0115] Specifically, from f T Lf = λf TFrom \(f = \lambda n\), it can be seen that to find satisfy is equivalent to finding the minimum value of \(\lambda\). Therefore, as long as the minimum eigenvalue \(\lambda\) of \(L\) is calculated. However, the minimum eigenvalue of the Laplacian matrix \(L\) is 0, and the elements of the corresponding eigenvector are all 1, so the condition \(f\perp1\) is not satisfied.

[0116] At this time, the Rayleigh - Ritz theorem needs to be referred to.

[0117] Let \(A\in C\) n×n be a Hermitian matrix. For the real function:

[0118]

[0119] Then it can be obtained that:

[0120]

[0121] According to the Rayleigh - Ritz theorem, for equation (12), the second - smallest eigenvalue and eigenvector of the Laplacian matrix \(L\) can be taken.

[0122] Based on the above derivation, it can be known that the solution of equation (12) is equivalent to finding the second - smallest eigenvalue of the Laplacian matrix corresponding to the undirected graph \(G\). Therefore, this value can be equivalent to the cut value of the undirected graph as a measure to describe the similarity of the echo phases of the same range bin and the same Doppler bin in each frame of echo.

[0123] Therefore, in this embodiment, all resolution cells are traversed, and when the undirected graph formed by each resolution cell is divided into two sub - graphs in turn, the second - smallest eigenvalue of its Laplacian matrix forms a measure matrix \(\lambda\) i ; and the second - smallest eigenvalue is used as the phase similarity measure of the resolution cell of the \(m\) - th range cell and the \(n\) - th Doppler bin.

[0124] S4. Use the K - means clustering algorithm to combine the measure matrix to divide the sea - surface scanning area into a sea clutter area and a land clutter area to obtain the initial land - sea segmentation result. Specifically, it includes:

[0125] S41. Use the K - means clustering algorithm to initially divide the measure matrix into a first subclass and a second subclass, and calculate the first measure mean of the first subclass and the second measure mean of the second subclass.

[0126] Specifically, use the K - means clustering algorithm to divide the measure matrix \(\lambda\) of the detection scene iPerform preliminary classification, dividing it into a first subclass and a second subclass, and marking 0 or 1. Then, calculate the first measure mean of the first subclass and the second measure mean of the second subclass respectively.

[0127] S42. Compare the first measure mean and the second measure mean, and determine the sea clutter area as the sea surface scanning area corresponding to the smaller value of the first measure mean and the second measure mean, and determine the land clutter area as the sea surface scanning area corresponding to the larger value, to obtain the initial land-sea segmentation result.

[0128] Specifically, the similarity between the phases of ocean echoes in different frames is small, so the corresponding measure is also small. On the contrary, the similarity between the phases of land echoes in different frames is large, so the corresponding measure is larger than that of ocean echoes. Therefore, the type of radar echo is judged by the measure means of the two categories. The radar echo with a smaller measure mean is sea clutter, and the radar echo with a larger measure mean is land clutter, so as to obtain the initial land-sea segmentation result.

[0129] In this embodiment, the K-means algorithm is used for classification. This algorithm does not require determining a classification threshold compared with the adaptive threshold segmentation method, and has a simple process and good segmentation performance.

[0130] S5. Perform morphological filtering on the initial land-sea segmentation result to obtain the land-sea segmentation result.

[0131] Specifically, perform morphological filtering on the initial land-sea segmentation result through the operators of binary morphology, filter out the burrs or isolated pixel points smaller than the structural element in the sea clutter area, and fill the holes smaller than the structural element in the land clutter area to obtain the land-sea segmentation result of the sea surface scanning area, and realize the land-sea segmentation of the detection scene.

[0132] Further, when the radar works in the next scanning cycle and the receiver receives the next frame of echo, repeat the above steps to update the land-sea segmentation result.

[0133] The multi-frame radar echo land-sea segmentation method in this embodiment combines the radar echo data of multiple scanning cycles, constructs an undirected graph for each resolution cell, and calculates the second smallest eigenvalue of the Laplacian matrix corresponding to the undirected graph as the phase similarity measure of the resolution cell, realizing the land-sea segmentation of multi-frame radar echoes, solving the problems of poor segmentation performance, low quality, and low result stability caused by limited data samples in the traditional land-sea segmentation algorithm based on single-frame radar echoes, having better robustness to outliers, and improving the consistency of the land-sea segmentation result. Embodiment Two

[0134] On the basis of Embodiment One, this embodiment further illustrates the effect of the multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm through simulation experiments.

[0135] 1. Experimental data

[0136] The measured sea clutter data used in this example are the data published on the official website of Radar Journal by Naval Aviation University in 2020 (numbered 20200722150529_738_scanning~20200722150529_841_scanning). The radar is installed at the test point of the First Yantai Seawater Bathing Beach, with an installation height of 80 m. The radar operates in the scanning mode, with a scanning speed of 24 r / min, a PRF of 1.6 kHz, and operates at a range of 6 nm. 44 consecutive frames of echo data are selected from it, and each set of data has 8442 range cells. The measured data are processed. First, each long pulse data segment is divided into multiple pulse segments with a length of K as simulated scanning data, and the pulse length K = 16.

[0137] 2. Simulation experiment

[0138] Using the measured sea clutter data, the existing phase linearity method and the method of this embodiment are respectively used for land-sea clutter scene segmentation, and the results are as Figure 3 shown, Figure 3 This is a comparison chart of the initial segmentation results of the phase linearity method provided by the embodiment of the present invention and the multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm. Among them, Figure 3 (a) represents the initial segmentation result obtained by using the phase linearity method; Figure 3 (b) represents the initial segmentation result of the land-sea clutter scene obtained by using the method of this embodiment. Figure 3 (a) and Figure 3 (b) have the wave position dimension on the horizontal axis and the range dimension on the vertical axis. White represents land and black represents the sea. From Figure 3 it can be seen that because multiple frames of echo data are used in the calculation, the number of isolated pixel clusters in the initial segmentation result obtained by the method of this embodiment is significantly less than the number of isolated pixel clusters in the result obtained by the phase linearity method.

[0139] 44 consecutive frames of echo data are selected from the measured sea clutter data. The method of this embodiment starts processing from the 5th scanning cycle and is compared with the segmentation method of the phase linearity method based on single-frame echo. The results are as Figure 4 shown, Figure 4 This is a morphological filtering result chart of the phase linearity method between multiple scanning cycles and the multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm provided by the embodiment of the present invention. From Figure 4 it can be seen that combining the land-sea segmentation results of multiple scanning cycles, the segmentation performance of the method of this embodiment is better, and it has better robustness to outliers.

[0140] Meanwhile, based on the above data, the true value is determined, the contour of the central island of the land-sea segmentation result is extracted, and the Hausdorff distance of the selected boundary curve between two frames of echoes is calculated to evaluate the quality of the land-sea segmentation result. The calculation results are as Figure 5 shown Figure 5 is the Hausdorff distance map of the phase linearity method and the multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm provided by the embodiment of the present invention. From Figure 5 it can be seen that in terms of the similarity of the land contours between different echoes, the multi-scan cycle echo data processed by the land-sea segmentation algorithm in this embodiment has less difference than the single-frame echo data processed by the phase linearity method, reducing the similarity difference of the land-sea clutter scenes between multi-frame radar echo data, and the result consistency is better, indicating that the land-sea segmentation performance of the method in this embodiment is better.

[0141] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should all be regarded as belonging to the protection scope of the present invention.

Claims

1. A multi-frame radar echo land-sea segmentation method based on spectral clustering algorithm, characterized in that Including the steps: S1. Obtain the current frame echo data information of the sea surface scanning area, and perform phase unwrapping on the current frame echo data information to obtain unwrapped phase information; S2. Use the unwrapped phase information of the same resolution cell in multiple frames of echo data information to construct an undirected graph of the resolution cell, and construct a Laplacian matrix for each undirected graph, where the number of frames of the multiple frames of echo data information is greater than or equal to 5 frames; S3. Combine the Ratio Cut criterion in the spectral clustering algorithm, calculate the second smallest eigenvalue of the Laplacian matrix as the phase similarity measure of the resolution cell, and obtain the measure matrix of all resolution cells; S4. Use the K-means clustering algorithm, and combine the measure matrix to divide the sea surface scanning area into a sea clutter area and a ground clutter area to obtain a preliminary sea-land segmentation result; S5. Perform morphological filtering on the preliminary sea-land segmentation result to obtain a sea-land segmentation result.

2. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 1, wherein Step S2 includes: S21. Use the unwrapped phase information of the resolution cell as the data of the vertices in the undirected graph, and use the resolution cells of the same wave position and the same range gate in one frame of echo as a vertex of the undirected graph. Obtain several vertices from the resolution cells of the same wave position and the same range gate in several frames of echo, and connect the several vertices to obtain the undirected graph; S22. Use the weight matrix of the undirected graph and the degree matrix of the vertices to construct the Laplacian matrix.

3. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 1, characterized in that The vertex expression of the undirected graph is: {φ 1 (m,n), φ 2 (m,n), …, φ i (m,n)} Among them, φ i (m,n) represents the unwrapped phase sequence of the resolution cell of the m-th range bin and the n-th Doppler bin in the i-th frame echo.

4. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 1, characterized in that The Laplacian matrix is: L = D - W Among them, L is the Laplacian matrix, and D is the degree matrix. w ij is the similarity degree between vertex v i and vertex v j . D ii is the degree of each vertex. The degree of a vertex represents the number of edges associated with that vertex. W is the weight matrix. e ij is the edge connecting vertices. E is the set of edges of the undirected graph. E = {e ij = <v i , v j > | v i , v j ∈ V}, and V is the set of vertices.

5. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 1, wherein Step S3 includes: S31. Calculate the minimization cost function when the undirected graph is optimally cut into several subgraphs according to the weight matrix of the undirected graph; S32. Improve the minimization cost function according to the constraint of the objective function of the Ratio Cut criterion in the spectral clustering algorithm for the scale of each subgraph to obtain the objective function; S33. Combine the Laplacian matrix, the eigenvector of the Laplacian matrix, and the properties of the unnormalized Laplacian matrix to optimize the objective function into a constrained optimization problem model; S34. Refer to the Rayleigh-Ritz theorem, solve the second smallest eigenvalue of the Laplacian matrix as the solution of the constrained optimization problem model, and use the second smallest eigenvalue of the Laplacian matrix as the phase similarity measure of the resolution cell, so as to obtain the measure matrix of all resolution cells.

6. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 5, wherein The minimization cost function is: Among them, B1, …, B k are the divided subgraphs, is the complement of B i , represents the sum of the weights of the edges connecting B i and , and k represents the number of divided subgraphs.

7. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 5, characterized in that The objective function is: Among them, |B i | represents the number of vertices in subgraph B i When k = 2, the objective function is denoted as 8. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 5, wherein Step S33 includes: Define the eigenvector as: f = (f1, f2, …, f n ) ∈ R n Among them, B is a subgraph, V is a set of vertices, |B| represents the number of vertices in subgraph B, and R n is an n-dimensional vector space; Substitute the eigenvector into the properties of the unnormalized Laplacian matrix to obtain an equation: Among them, |B| represents the number of vertices in subgraph B, |V| represents the number of vertices in the original graph, and both |V| and |B| are constants. n is the number of vertices, and L is the Laplacian matrix. From the above equation, it can be obtained that minimizing the objective function is equivalent to minimizing f T Lf; Combined with the minimization of f T Lf, the objective function is optimized into an optimization problem model with constraints: f satisfies f⊥1, f T Lf = λf T f = λn where n is the number of vertices, and λ is the eigenvalue of the Laplacian matrix.

9. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 1, characterized in that Step S4 includes: S41. Use the K-means clustering algorithm to preliminarily divide the measure matrix into a first subclass and a second subclass, and calculate the first measure mean of the first subclass and the second measure mean of the second subclass; S42. Compare the first measurement mean and the second measurement mean, and determine the sea clutter region corresponding to the smaller value of the first measurement mean and the second measurement mean, and determine the land clutter region corresponding to the larger value of the first measurement mean and the second measurement mean, so as to obtain the initial land-sea segmentation result.

10. The multi-frame radar echo land-sea segmentation method based on the spectral clustering algorithm according to claim 1, characterized in that Step S5 includes: Use the operator of binary morphology to perform morphological filtering on the initial land-sea segmentation result, so as to eliminate the isolated pixel clusters existing in the sea clutter region and fill the holes existing in the land clutter region, and obtain the land-sea segmentation result of the sea surface scanning region.

Images