A Method and System for Detecting Ship Targets on the Sea Surface by a Polarimetric Radar

Abstract

The present invention relates to a method and system for detecting sea surface ship targets by a polarimetric radar. The method is characterized in that it includes: acquiring polarimetric radar image data to be detected and performing preprocessing; performing pre-detection on the preprocessed polarimetric radar data, and marking ship targets in the polarimetric radar data to generate initial ship samples; extracting tensor data slices for each pixel in the initial ship samples to generate a ship sample set; using the ship sample set to train a semi-supervised support tensor machine based on Tucker decomposition to obtain the detection result of ship targets in the polarimetric radar image data to be detected. The present invention can be widely applied in the field of target detection.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to a method and system for detecting sea surface ship targets using a polarization radar. Background Art

[0002] Radar is an active detection system that is not affected by environmental factors such as weather, fog, and light, and can work all day and all weather. Compared with single-channel radar, polarimetric radar obtains multi-channel observation data by transmitting and receiving electromagnetic waves in two different polarization modes, and obtains the radar cross-section (RCS) of electromagnetic waves of the target in different polarization modes, thereby obtaining richer scattering characteristics, which helps to extract target information in complex scenes. Therefore, polarimetric radar has a wide range of applications in military, aviation, remote sensing and other fields. In recent years, the use of polarimetric radar for sea surface ship target detection has become an important application in the field of radar remote sensing. In civilian use, sea surface ship target detection helps manage shipping traffic and marine fishery activities, and can effectively combat illegal activities such as smuggling. In military use, sea surface ship target detection can monitor the location of key targets that need to be attacked in wartime and other information to ensure military initiative. Therefore, for a large amount of polarimetric radar sea surface data, it is necessary to timely and effectively discover and extract ship targets, and it is necessary to carry out research on polarimetric radar target detection algorithms for sea surface ships.

[0003] At present, most polarization radar sea surface ship target detection systems can be divided into the following two types: 1) Target detection based on sea clutter model. In the target detection framework based on the sea clutter model, in medium and low sea conditions, the sea surface clutter signal strength is weak and the ship target echo signal strength is strong. The ship target detection system statistically models the sea surface clutter signal and uses the constant false alarm detection method to obtain the threshold, thereby completing the ship target detection. Then, the existing priors such as ship size and ship scattering characteristics are used to remove false alarm results from the detection results, and the final ship detection results are given. 2) Target detection based on the polarization characteristics of ship targets. In the target detection framework based on the polarization characteristics of ship targets, the detection system mainly focuses on how to extract the polarization characteristics of the target for target detection. Since the ship target is a man-made target, its electromagnetic scattering characteristics are quite different from those of the sea surface. Therefore, there are polarization characteristics differences between the ship target and the sea surface, so the ship can be distinguished from the sea clutter. By comprehensively considering the target backscattering power, scattering characteristics (such as Bragg scattering, secondary scattering, volume scattering), polarization decomposition coefficients, polarization entropy and other characteristics, the ship detection result can be obtained. In addition, setting a certain optimization objective function to integrate the information of different polarization channels is also a way to extract target features. For example, based on the clutter fluctuation criterion, a polarization whitening filter (PWF) can be derived.

[0004] However, the existing ship detection methods have certain limitations. The detection method based on the sea clutter model has the advantage that few hyperparameters are required to be set manually. The detection system can adaptively learn the statistical model of sea clutter and detect ship targets based on it. Therefore, the detection system can have good generalization performance under different observation sensors and different scene conditions. However, its limitations are that the existing sea clutter model can usually only represent the sea surface in medium and low sea conditions. In high sea conditions, the distribution of sea clutter is complex and random, and it is difficult to establish a good sea clutter model. The monitoring framework based on the polarization characteristics of ship targets has the advantage that artificial ship targets generally have some significant polarization scattering characteristics. Usually, this feature can be distinguished from the naturally formed sea surface regardless of medium and low sea conditions or high sea conditions. These features can be used to effectively detect ship targets in some complex scenes. However, the polarization characteristics of the target will change with various factors such as ship type, sensor type, scene conditions, etc., and it is difficult for artificially designed polarization features to achieve good performance in all situations. This is because these polarization features incorporate some polarization information in the process of artificial design, but also ignore a lot of polarization information. In this case, in order to effectively detect targets under various ship types, sensor types, and sea scene conditions, it is necessary to propose new methods to retain polarization information as much as possible, or to perform targeted feature extraction based on the corresponding sensors, ship types, and sea scene conditions. This usually requires supervised polarization feature learning, but supervised learning faces limitations such as data acquisition. Summary of the invention

[0005] In view of the above problems, the object of the present invention is to provide a method and system for detecting sea surface ship targets using polarization radar, which can completely retain all spatial information and polarization information in the neighborhood.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions: In a first aspect, a method for detecting sea surface ship targets using a polarization radar is provided, comprising:

[0007] Acquire polarization radar image data to be detected and perform preprocessing;

[0008] Pre-detect the pre-processed polarimetric radar data, and mark the ship targets in the polarimetric radar data to generate initial ship samples;

[0009] A tensor data slice is extracted for each pixel in the initial ship sample to generate a ship sample set;

[0010] The ship sample set is used to train the semi-supervised support tensor machine based on Tucker decomposition, and the ship target detection results in the polarimetric radar image data to be detected are obtained.

[0011] Further, the pre-detection of the pre-processed polarization radar data and marking of the ship targets in the polarization radar data to generate an initial ship sample includes:

[0012] The pre-processed polarimetric radar data are pre-detected using a modified polarimetric whitening filter to obtain the corrected pixel intensities in single-view and multi-view conditions respectively.

[0013] According to the corrected pixel intensities in single-view and multi-view, the ship targets in the polarimetric radar data are marked to generate initial ship samples.

[0014] Furthermore, the corrected pixel intensity includes the corrected pixel intensity in single view and multi-view, wherein:

[0015] The corrected pixel intensity z in single vision is:

[0016]

[0017] The corrected pixel intensity z in multi-view is:

[0018]

[0019] Among them, S hh , S hv and S vv are the scattering components corresponding to the HH channel, HV channel, and VV channel of the polarization radar scattering matrix; y ij , i, j = 1, 2, 3 are the components of the positive definite Hermite matrix, representing the category label; r 1 is the first-order similarity parameter; r 2 is the second-order similarity parameter; α 0 is the weight before the first-order similarity parameter; the superscript * indicates complex conjugation; the parameters ρ, γ, ε and σ hh for:

[0020]

[0021] Among them, E means to find the expectation.

[0022] Furthermore, extracting a tensor data slice for each pixel in the initial ship sample to generate a ship sample set includes:

[0023] For each pixel in the initial ship sample, a tensor data slice is extracted;

[0024] Extract all tensor data slices with category labels as ships to form a ship sample set.

[0025] Furthermore, the ship sample set is used to train the semi-supervised support tensor machine based on Tucker decomposition to obtain the ship target detection result in the polarization radar image data to be detected, including:

[0026] According to the ship sample set, the classification interface decision function of the Tucker decomposition-based support tensor machine is trained, and the Tucker decomposition-based support tensor machine is expanded into a semi-supervised form;

[0027] Calculate the classification interface decision function for each remaining tensor data slice and output the category label;

[0028] If the maximum number of iterations has not been reached and there is a new ship sample whose category label is ship, then select a new tensor data slice from the new ship sample that maximizes the inner product operation between the weight parameter and the tensor data slice under the meaning of Tucker decomposition, set its category label to ship, add the tensor data slice to the ship sample set, and re-perform iterative training; otherwise, end the iteration;

[0029] According to the classification interface decision function obtained from the last round of iterative training, category labels are generated for all tensor data slices, and the ship target detection results are returned.

[0030] Furthermore, the classification interface decision function of the support tensor machine based on Tucker decomposition is trained according to the ship sample set, and the support tensor machine based on Tucker decomposition is expanded to a semi-supervised form, including:

[0031] Set the core tensor of the tensor data slice and perform Tucker decomposition on all tensor data slices;

[0032] Initialize the ship sample set, set the maximum number of algorithm iterations and support tensor machine hyperparameters;

[0033] Based on the Tucker decomposition of all tensor data slices, as well as the set maximum number of iterations and tensor support hyperparameters, the classification interface decision function for the tensor support machine is trained;

[0034] The weight parameters and bias of the classification interface decision function are calculated to obtain the classification interface decision function.

[0035] Furthermore, the classification interface decision function f(x) is:

[0036]

[0037] in, is the weight parameter; b is the bias; A k (k=1, 2, ..., N) is the tensor data slice in The projection of each mode of k is the n-modular product of the tensor data slices, The core tensor of Tucker decomposition for slicing tensor data; J N is the dimension of projection on each channel.

[0038] In a second aspect, a polarization radar sea surface ship target detection system is provided, comprising:

[0039] A data preprocessing module, used to obtain polarization radar image data to be detected and perform preprocessing;

[0040] A pre-detection module is used to perform pre-detection on the pre-processed polarization radar data, mark the ship targets in the polarization radar data, and generate initial ship samples;

[0041] A tensor data slice extraction module is used to extract a tensor data slice for each pixel in the initial ship sample to generate a ship sample set;

[0042] The model training module is used to train the semi-supervised support tensor machine based on Tucker decomposition using a ship sample set to obtain the ship target detection results in the polarization radar image data to be detected.

[0043] In a third aspect, a processing device is provided, comprising computer program instructions, wherein the computer program instructions, when executed by the processing device, are used to implement the steps corresponding to the above-mentioned polarization radar sea surface ship target detection method.

[0044] In a fourth aspect, a computer-readable storage medium is provided, on which computer program instructions are stored, wherein the computer program instructions, when executed by a processor, are used to implement the steps corresponding to the above-mentioned polarization radar sea surface ship target detection method.

[0045] The present invention adopts the above technical solution, which has the following advantages:

[0046] 1. The present invention adopts a modified polarization whitening filter for pre-detection and a semi-supervised support tensor machine based on Tucker decomposition for ship target detection. The support tensor machine allows input in the form of tensors, so that all polarization information and spatial information of the neighborhood data can be completely retained, so that the sea surface polarization data can retain more high-dimensional structural information, which is helpful to improve the polarization radar sea surface ship target detection capability, and is of great significance to the detection of sea surface ships and the development and protection of marine resources.

[0047] 2. The present invention performs detection in a semi-supervised form, using initial ship samples in polarization radar image data and performing classification interface training. The initial ship samples are generated by a modified polarization whitening filter, and then the classification interface is used to select several previously unlabeled ship targets with the highest confidence and add them to the ship sample set to retrain the classification interface. The above process is repeated, and when the iterative process stops, the ship target detection is completed according to the classification interface. This semi-supervised method overcomes the limitations of data acquisition that supports tensor machines as supervised learning, and can be applied to various sensor types and sea surface scenarios.

[0048] 3. The input data of the present invention is the polarization radar image to be detected, and the output is the ship target detection result. The result is in the form of a binary image, which indicates that each pixel is a ship target or a sea clutter signal. Compared with the existing technology, the polarization characteristics can be fully retained, so that it is suitable for different ships, sensors, and scene conditions, and overcomes the disadvantage that the tensor machine as a supervised learning algorithm cannot be directly trained using unlabeled data.

[0049] In summary, the present invention can be widely used in the field of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Throughout the accompanying drawings, the same reference numerals are used to represent the same components. In the accompanying drawings:

[0051] Figure 1 It is a schematic diagram of a method flow provided by an embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of polarization radar data provided by an embodiment of the present invention;

[0053] Figure 3 is a schematic diagram of an initial ship target selected from polarization radar data provided by an embodiment of the present invention;

[0054] Figure 4 is a schematic diagram of a tensor data slicing process provided by an embodiment of the present invention;

[0055] Figure 5 It is a schematic diagram of ship target detection results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0057] It should be understood that the terms used herein are only for the purpose of describing specific example embodiments and are not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms "one", "an" and "said" as used herein may also be meant to include plural forms. The terms "include", "comprise", "contain", and "have" are inclusive, and therefore specify the existence of stated features, steps, operations, elements and / or parts, but do not exclude the existence or addition of one or more other features, steps, operations, elements, parts, and / or combinations thereof. The method steps, processes, and operations described herein are not interpreted as necessarily requiring them to be performed in the specific order described or illustrated, unless the execution order is clearly indicated. It should also be understood that additional or alternative steps may be used.

[0058] Although the terms first, second, third, etc. can be used in the text to describe multiple elements, components, regions, layers and / or sections, these elements, components, regions, layers and / or sections should not be limited by these terms. These terms can only be used to distinguish an element, component, region, layer or section from another region, layer or section. Unless the context clearly indicates, terms such as "first", "second" and other numerical terms do not imply order or sequence when used in the text. Therefore, the first element, component, region, layer or section discussed below can be referred to as the second element, component, region, layer or section without departing from the teaching of the example embodiments.

[0059] In the existing ship target detection framework based on polarization features, which polarization information is retained by polarization feature fusion depends on manual design, and some polarization information is lost in the process of extracting features. Therefore, it is difficult to apply to various types of ships, sensor types, and sea scenes, thus affecting the detection performance. In addition, the existing technology usually performs feature learning for given ships, sensors, and scenes in a supervised learning manner, but there are limitations in data acquisition. The polarization radar sea surface ship target detection method and system provided in an embodiment of the present invention generates initial ship samples through a polarization whitening filter, and then iteratively trains the classification interface in a semi-supervised learning form, which can overcome the disadvantage that the tensor machine as a supervised learning algorithm cannot be directly trained using unlabeled data. In addition, due to the use of a support tensor machine based on Tucker decomposition, it allows direct input in tensor form and target detection. The multi-polarization and multi-phase polarization radar data naturally has a tensor form. Direct input in the form of a tensor can completely retain the structural information in multiple dimensions such as space, time, and polarization channels, as well as all polarization information of targets in the neighborhood, avoiding the loss of polarization information due to feature extraction, which helps to improve the performance of ship target detection.

[0060] Example 1

[0061] like Figure 1 As shown, this embodiment provides a polarization radar sea surface ship target detection method, comprising the following steps:

[0062] 1) Obtain the polarimetric radar image data to be detected and perform preprocessing, specifically:

[0063] 1.1) Obtain polarization radar image data to be detected.

[0064] Specifically, each pixel in the polarimetric radar image data is a 3*3 coherent matrix, such as Figure 2 Shown is an example of a pseudo-color image of polarimetric radar image data.

[0065] 1.2) Filter the coherent spots of the acquired polarization radar image data.

[0066] Specifically, the filter used is a Boxcar filter, which is a filtering method that takes the average pixel value in a sliding window. For single-channel image data, the relationship between the pixel value after filtering and the pixel value in the sliding window is:

[0067]

[0068] in, is the polarization radar image data after filtering; Y i,jis the original polarimetric radar image data; M and N are the sizes of the filter window; i and j are the row and column numbers corresponding to the image filtering area, respectively.

[0069] The polarization radar image data in the present invention is multi-channel image data, and the Boxcar filter filters each element of the polarization coherence matrix separately. For the diagonal elements, since they are real numbers, the image data generated by the diagonal elements can be directly filtered; for the non-diagonal elements, since they are complex numbers, the real part and the imaginary part are window filtered separately; since the matrix is ​​conjugate symmetric, a total of 3+3×2=9 channels of image data need to be filtered, and the filtering method of each channel is consistent with the filtering method of single-channel radar image data.

[0070] Specifically, since the window parameters M and N are key parameters affecting Boxcar filtering, the larger the M and N are, the greater the smoothness is, but more details may be lost. Therefore, the window size is generally between 5 and 11, and the default value in the present invention is M=N=5.

[0071] 2) Using the modified polarization whitening filter, pre-detect the pre-processed polarization radar data, and mark the ship targets in the polarization radar data to generate the initial ship samples, specifically:

[0072] 2.1) The pre-processed polarimetric radar data are pre-detected using a modified polarimetric whitening filter to obtain the corrected pixel intensities in single-view and multi-view, respectively.

[0073] Specifically, the polarization scattering matrix of the polarization radar has four components. Generally, for reciprocal media, the two cross-polarization components are equal, which can be simplified into a vector form:

[0074] v=(S hh S hv S vv ) T (2)

[0075] Among them, S hh , S hv and S vv They are the scattering components corresponding to the polarization radar scattering matrix HH (horizontal transmission and horizontal reception) channel, HV (horizontal transmission and vertical reception) channel, and VV (vertical transmission and vertical reception) channel; the superscript T represents the transpose; v is the polarization feature vector, which is often considered to obey the multivariate complex Gaussian distribution when there is no texture modulation of the backscattering coefficient of the scene surface, and its probability density function P v (v) is:

[0076]

[0077] Wherein, the superscript H represents the conjugate transpose; q is the dimension of the polarization eigenvector, and q=3; C is the covariance matrix, and C=E(vv H ) is the covariance matrix of the polarization eigenvector v, E represents the expectation; |C| represents the determinant of the covariance matrix C. The above formula shows that the observations of each polarization channel are considered to be zero mean, that is, E(v) = 0. Assume that the cross-polarization component S hv and the in-phase polarization component S hh , S vv When there is no coupling between the polarization components, but there is coupling between the in-phase polarization components, the covariance matrix C is:

[0078]

[0079] The superscript * indicates complex conjugation, and the parameters ρ, γ, ε, and σ hh for:

[0080]

[0081] For multi-look polarimetric radar image data, its multi-look covariance matrix V is:

[0082]

[0083] Where N is the number of views, v i is the polarization eigenvector under viewing number i; the superscript H represents the conjugate transpose.

[0084] For a uniform scene, the multi-view covariance matrix V is considered to obey the multivariate complex Wishart distribution, and its PDF (probability density function) is:

[0085]

[0086] Where N is the number of views; q is the dimension of the polarization eigenvector; Tr(·) represents the trace of the matrix; G(N,q) is the normalization coefficient of the PDF, and Γ(·) represents the Gamma function.

[0087] The above discussion is based on the assumption of a uniform scene, but most actual scenes are non-uniform. There are two factors in radar measurement: one is the coherent speckle noise caused by multipath interference of coherent waves scattered by randomly distributed scenes; the other is the texture characteristic, that is, the spatial fluctuation of the non-uniform scene. Regardless of single-view or multi-view, in coherent radar image data, coherent speckles are usually considered to be a multiplicative noise. A commonly used single-view polarization coherence model assumes that the single-view observation vector y of the polarization feature vector of the polarization radar is the product of a Gamma-distributed texture scalar factor t and a complex vector v representing that the coherent speckle obeys a multivariate complex Gaussian distribution, and the two are considered to be independent of each other. The mathematical expression is:

[0088]

[0089] In this model, it is assumed that the texture scalar factor t affects each polarization channel equally. According to this multiplicative speckle model, the single-view observation vector y of the polarization radar obeys a conditional complex Gaussian distribution, and its PDF is:

[0090]

[0091] Wherein, q is the dimension of the polarization eigenvector; C is the covariance matrix given in the above formula (4).

[0092] For the multi-view case, the observation quantity Y of its covariance matrix is:

[0093]

[0094] Among them, t i 、v i are the texture variables and polarization coherent speckle noise feature vectors of the i-th single-view sample data. In general, the texture variables have a higher spatial correlation than the coherent speckle noise feature vector, or the spatial variation of the texture variables is much slower than the spatial variation of the polarization coherent speckle noise. Therefore, it can be considered that the texture variables used for multi-view averaging are approximately equal, so:

[0095] Y=t·V (11)

[0096] Among them, Y is the observation of the polarization covariance matrix; V is the observation of the multi-view polarization coherent speckle noise covariance matrix. Therefore, it can be deduced that the observation of the covariance matrix Y obeys the conditional complex Wishart distribution, and its PDF is:

[0097]

[0098] Usually, the coherent speckle is measured by the ratio of the standard deviation s of the image to its mean m, s / m. The guiding principle of the polarization whitening filter is to minimize s / m. Let y be the single-view observation vector of the polarization radar, that is, the observation value of the polarization characteristic vector in the single-view case, A 1 is a q×q positive definite Hermite matrix, q=3 is the dimension of the polarization feature vector, and the pixel intensity z is constructed as:

[0099] z=y H Ay (13)

[0100] By derivation, we can obtain the positive definite Hermite matrix A 1 :

[0101] A 1 =C -1(14)

[0102] Therefore, the final solution to minimize the speckle is:

[0103]

[0104] The above formula (15) is the solution of the polarization whitening filter in single-view.

[0105] For the multi-view case, let Y be the observation of the polarization covariance matrix, A 2 is a q×q positive definite Hermite matrix, q=3 is the dimension of the polarization feature vector, and the pixel intensity z is constructed as:

[0106] z=Tr(A 2 Y) (16)

[0107] It can be deduced that the solution of the polarization whitening filter in multi-view is:

[0108]

[0109] Among them, y ij , i, j = 1, 2, 3 is a positive definite Hermite matrix A 2 of each component.

[0110] Specifically, the sea clutter is corrected based on the polarization whitening filter. For a pixel in the given polarization radar image data to be detected, its first-order similarity parameter r can be calculated. 1 and the second-order similarity parameter r 2 :

[0111]

[0112]

[0113] Select a certain sea surface area (40×40 is selected by default in this embodiment) and calculate the first-order similarity parameter r pixel by pixel. 1 and the second-order similarity parameter r 2 , and calculate g(α):

[0114]

[0115] Among them, g(α) is the auxiliary function for calculating the correction coefficient; k is the kth region, and α is the weight before the first-order similarity parameter. The function of the above formula (20) with respect to α has a minimum value, and it can be found that the minimum value of g is α=α 0 At this time, let the weight α that makes the auxiliary function g take the minimum value be 0 is the first-order correction coefficient. The solution of the modified polarization whitening filter needs to be multiplied by the correction coefficient (α 0r 1 +r 2 ) 2 , that is, the corrected pixel intensity z in single vision is:

[0116]

[0117] The corrected pixel intensity z in multi-view is:

[0118]

[0119] 2.2) According to the corrected pixel intensities in single view and multi-view, the ship targets x(i, j) in the polarimetric radar data are marked to generate initial ship samples.

[0120] Specifically, after obtaining the solution of the modified polarization whitening filter, that is, the modified pixel intensity z, the constant false alarm detection (CFAR) method can be used to determine the detection threshold, and n ship targets that exceed the threshold the most are selected as initial ship samples. In this embodiment, n=1 is selected by default. The initial ship targets selected in this embodiment are as follows: Figure 3 shown.

[0121] 3) Extract tensor data slices for each pixel in the initial ship sample to generate a ship sample set.

[0122] Specifically, the pre-detection in step 2) is pixel-level, and it is necessary to extract tensor data slices for each pixel to support the training of the tensor machine. The purpose of extracting tensor data slices is to retain all polarization information and structural information in the pixel neighborhood in the form of tensors. Assume that the size of the selected image window is w, and in this embodiment, w=3 is the default. Consider a certain pixel x(i, j), and denote the neighborhood of pixel x(i, j) with a size of w as X {i，j} The extraction of tensor data slices needs to reflect the pixel neighborhood X in addition to polarization information. {i，j} The neighborhood of other adjacent pixels (for example, the neighborhood X of pixel x(i+1, j+1) {i+1，j+1} ). The tensor data slice extracted from pixel x(i, j) is recorded as By neighborhood X {i，j} , X {i+1，j+1} , X {i +1，j-1} , X {i-1，j+1} , X {i-1，j-1} Combination composition, i.e. tensor data slicing for:

[0123]

[0124] Among them, each neighborhood X {i，j} The superscript {i, j} indicates the neighborhood of pixel (i, j). The first two dimensions are the length and width of the neighborhood window, and the third dimension is the vector form v of the corresponding pixel polarization covariance matrix C, which is:

[0125]

[0126] Therefore, the tensor data slice extracted for pixel x(i, j) is When supporting tensor machine detection in the future, The category of +1 indicates that the pixel x(i, j) is a ship, and the category of -1 indicates that the pixel x(i, j) is sea clutter. The schematic diagram of the above tensor data slicing process is as follows Figure 4 shown.

[0127] The specific process of this step is:

[0128] 3.1) Extract tensor data slices for each pixel in the initial ship sample

[0129] Specifically, at the beginning of the iteration, according to the ship target x(i, j) marked in step 2), the corresponding tensor data slice is extracted: At the end of a round of iteration, a tensor data slice with category +1 is returned In any case, the corresponding tensor data slice can be found And its category label y i，j Set to +1.

[0130] 3.2) Extract all tensor data slices with the category label of ship to form a ship sample set, which is used to train the classification interface supporting the tensor machine in the next round of iteration.

[0131] 4) Using the ship sample set, the semi-supervised support tensor machine based on Tucker Decomposition is trained to obtain the ship target detection results in the polarization radar image data to be detected, specifically:

[0132] 4.1) According to the ship sample set, the classification interface decision function of the Tucker decomposition-based support tensor machine is trained, and the Tucker decomposition-based support tensor machine is expanded to a semi-supervised form.

[0133] Specifically, for a tensor machine that supports Tucker decomposition, set the tensor data slice The Tucker decomposition of is expressed as:

[0134]

[0135]

[0136] Among them, I N for The dimension of each channel; for the various elements of for Each element of k is the n-modular product of the tensor data slices; Slice tensor data The core tensor of Tucker decomposition, A 1k , A 2k (k=1, 2, ..., N) is the tensor data slice exist The projection of each mode, J N is the dimension of projection on each channel.

[0137] The decision function of the classification interface supporting tensor machines is Input variables are tensor data slices

[0138]

[0139] in, is the weight parameter, b is the bias, and <·, ·> represents the inner product operation between tensor data slices under the meaning of Tucker decomposition. The specific form is:

[0140]

[0141] Among them, g w (i 1 ,i 2 , ..., i N )for Core tensors after Tucker decomposition Each element of g x (j 1 , j 2 , ..., j N )for Core tensors after Tucker decomposition the various elements of for and The direct inner product of , k is the subscript of the channel; Slice tensor data exist Projections of each mode.

[0142] The goal of supporting tensor machines is to find the decision function, that is, the weight parameters in the decision function and bias b, where the weight parameter for:

[0143]

[0144] in, for Projection onto the k-th channel after Tucker decomposition.

[0145] Assume that the tensor data slice The corresponding class label is y p ∈{-1, 1}, all The number is denoted as P, and the subscript p indicates that from all The subscripts generated by the selection satisfy 1≤p≤P, and D is the selected hyperparameter. The support tensor machine based on Tucker decomposition is expressed as:

[0146]

[0147] Among them, ξ i It is an auxiliary parameter.

[0148] The dual problem is:

[0149]

[0150] Among them, α p With y p is the dual auxiliary variable used in solving the problem.

[0151] The above dual problem can be solved by using the Sequential Minimal Optimization (SMO) algorithm. The weight parameters of the original problem can be obtained from solving the dual problem. The solution to obtain the classification interface decision function that supports the tensor machine

[0152]

[0153] Therefore, the specific process of this step is:

[0154] 4.1.1) Setting the core tensor of tensor data slice And perform Tucker decomposition on all tensor data slices to facilitate the subsequent calculation of tensor inner products.

[0155] 4.1.2) Initialize the ship sample set (i.e., category +1), and set the maximum number of algorithm iterations Mmax (default Mmax = 50) and the supporting tensor machine hyperparameter D.

[0156] 4.1.3) According to the Tucker decomposition of all tensor data slices, as well as the set maximum number of iterations Mmax and the tensor machine hyperparameter D, the classification interface decision function supporting the tensor machine is trained.

[0157] Specifically, a sequential minimum optimization algorithm is used to solve the dual problem of the above formula (31).

[0158] 4.1.4) Based on the solution of the dual problem, the weight parameters of the classification interface decision function are calculated and bias b, we get the classification interface decision function (Here the tensor data is sliced Expressed in Tucker decomposition form):

[0159]

[0160] 4.2) Calculate the classification interface decision function for each remaining tensor data slice and output the category label y i，j , where category y i，j = +1 means the target pixel is a ship, y i，j =-1 indicates that the target pixel is sea clutter.

[0161] 4.3) If the maximum number of iterations Mmax is not reached and there is a new ship sample in step 4.2) with a class label of +1, then select from the new ship sample the inner product operation between the weight parameter and the tensor data slice in the sense of Tucker decomposition The largest new tensor data slice The bigger the The corresponding pixel i, j is more likely to be a ship target), and its category label y is set i，j = +1, and slice the tensor data Add it to the ship sample set and go to step 4.1.3) to enter the next round of iteration; otherwise, end the iteration.

[0162] 4.4) According to the classification interface decision function obtained from the last round of iterative training, slice all tensor data Generate category label y i，j , returns the ship target detection result.

[0163] Example 2

[0164] This embodiment provides a polarization radar sea surface ship target detection system, including:

[0165] The data preprocessing module is used to obtain the polarization radar image data to be detected and perform preprocessing.

[0166] The pre-detection module is used to use a modified polarization whitening filter to pre-detect the pre-processed polarization radar data, mark the ship targets in the polarization radar data, and generate initial ship samples.

[0167] The tensor data slice extraction module is used to extract tensor data slices for each pixel in the initial ship sample to generate a ship sample set.

[0168] The model training module is used to train the semi-supervised support tensor machine based on Tucker decomposition using a ship sample set to obtain the ship target detection results in the polarization radar image data to be detected.

[0169] Example 3

[0170] This embodiment provides a processing device corresponding to the polarization radar sea surface ship target detection method provided in this embodiment 1. The processing device can be a processing device for a client, such as a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the method of embodiment 1.

[0171] The processing device includes a processor, a memory, a communication interface and a bus, and the processor, the memory and the communication interface are connected through the bus to complete mutual communication. The memory stores a computer program that can be run on the processing device, and the processing device executes the polarization radar sea surface ship target detection method provided in this embodiment 1 when running the computer program.

[0172] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0173] In some other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors of various types, which are not limited herein.

[0174] Example 4

[0175] This embodiment provides a computer program product corresponding to the polarization radar sea surface ship target detection method provided in this embodiment 1. The computer program product may include a computer-readable storage medium on which computer-readable program instructions for executing the polarization radar sea surface ship target detection method described in this embodiment 1 are loaded.

[0176] Computer readable storage media can be tangible devices that hold and store instructions used by instruction execution devices. Computer readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any combination thereof.

[0177] The above embodiments are only used to illustrate the present invention, wherein the structure, connection mode and manufacturing process of each component may be changed. Any equivalent transformations and improvements based on the technical solution of the present invention should not be excluded from the protection scope of the present invention.

Claims

1. A method for detecting sea surface ship targets using polarization radar. It is characterized in that include: Acquire polarization radar image data to be detected and perform preprocessing; Pre-detect the pre-processed polarimetric radar data, and mark the ship targets in the polarimetric radar data to generate initial ship samples; A tensor data slice is extracted for each pixel in the initial ship sample to generate a ship sample set; The ship sample set is used to train the semi-supervised support tensor machine based on Tucker decomposition, and the ship target detection results in the polarimetric radar image data to be detected are obtained. The ship sample set is used to train the semi-supervised support tensor machine based on Tucker decomposition to obtain the ship target detection result in the polarization radar image data to be detected, including: According to the ship sample set, the classification interface decision function of the Tucker decomposition-based support tensor machine is trained, and the Tucker decomposition-based support tensor machine is expanded into a semi-supervised form; Calculate the classification interface decision function for each remaining tensor data slice and output the category label; If the maximum number of iterations has not been reached and there is a new ship sample whose category label is ship, then select a new tensor data slice from the new ship sample that maximizes the inner product operation between the weight parameter and the tensor data slice under the meaning of Tucker decomposition, set its category label to ship, add the tensor data slice to the ship sample set, and re-perform iterative training; otherwise, end the iteration; According to the classification interface decision function obtained from the last round of iterative training, category labels are generated for all tensor data slices, and the ship target detection results are returned.

2. A method for detecting sea surface ship targets using polarization radar as claimed in claim 1, It is characterized in that The pre-detection of the pre-processed polarization radar data and marking of the ship targets in the polarization radar data to generate an initial ship sample includes: The pre-processed polarimetric radar data are pre-detected using a modified polarimetric whitening filter to obtain the corrected pixel intensities in single-view and multi-view conditions respectively. According to the corrected pixel intensities in single-view and multi-view, the ship targets in the polarimetric radar data are marked to generate initial ship samples.

3. A method for detecting sea surface ship targets using polarization radar as claimed in claim 2, It is characterized in that The corrected pixel intensity includes the corrected pixel intensity in single view and multi-view, wherein: The corrected pixel intensity z in single vision is: The corrected pixel intensity z in multi-view is: Among them, S hh , S hv and S vv are the scattering components corresponding to the HH channel, HV channel, and VV channel of the polarization radar scattering matrix; y ij , i,j=1,2,3 are the components of the positive definite Hermite matrix, representing the category label; r 1 is the first-order similarity parameter; r 2 is the second-order similarity parameter; α 0 is the weight before the first-order similarity parameter; the superscript * indicates complex conjugation; the parameters ρ, γ, ε and σ hh for: Among them, E means to find the expectation.

4. A method for detecting sea surface ship targets using polarization radar as claimed in claim 1, It is characterized in that The tensor data slice is extracted for each pixel in the initial ship sample to generate a ship sample set, including: For each pixel in the initial ship sample, a tensor data slice is extracted; Extract all tensor data slices with category labels as ships to form a ship sample set.

5. A method for detecting sea surface ship targets using polarization radar as claimed in claim 1, It is characterized in that According to the ship sample set, the classification interface decision function of the support tensor machine based on Tucker decomposition is trained, and the support tensor machine based on Tucker decomposition is expanded to a semi-supervised form, including: Set the core tensor of the tensor data slice and perform Tucker decomposition on all tensor data slices; Initialize the ship sample set, set the maximum number of algorithm iterations and support tensor machine hyperparameters; Based on the Tucker decomposition of all tensor data slices, as well as the set maximum number of iterations and tensor support hyperparameters, the classification interface decision function for the tensor support machine is trained; The weight parameters and bias of the classification interface decision function are calculated to obtain the classification interface decision function.

6. A method for detecting sea surface ship targets using polarization radar as claimed in claim 5, It is characterized in that The classification interface decision function for: in, is the weight parameter; b is the bias; A k (k=1,2,…,N) is the tensor data slice in The projection of each mode of k is the n-modular product of the tensor data slices, The core tensor of Tucker decomposition for slicing tensor data; J N is the dimension of projection on each channel.

7. A polarization radar sea surface ship target detection system, It is characterized in that include: A data preprocessing module, used to obtain polarization radar image data to be detected and perform preprocessing; A pre-detection module is used to perform pre-detection on the pre-processed polarization radar data, mark the ship targets in the polarization radar data, and generate initial ship samples; A tensor data slice extraction module is used to extract a tensor data slice for each pixel in the initial ship sample to generate a ship sample set; A model training module is used to train a semi-supervised support tensor machine based on Tucker decomposition using a ship sample set to obtain a ship target detection result in the polarization radar image data to be detected; The ship sample set is used to train the semi-supervised support tensor machine based on Tucker decomposition to obtain the ship target detection result in the polarization radar image data to be detected, including: According to the ship sample set, the classification interface decision function of the Tucker decomposition-based support tensor machine is trained, and the Tucker decomposition-based support tensor machine is expanded into a semi-supervised form; Calculate the classification interface decision function for each remaining tensor data slice and output the category label; If the maximum number of iterations has not been reached and there is a new ship sample whose category label is ship, then select a new tensor data slice from the new ship sample that maximizes the inner product operation between the weight parameter and the tensor data slice under the meaning of Tucker decomposition, set its category label to ship, add the tensor data slice to the ship sample set, and re-perform iterative training; otherwise, end the iteration; According to the classification interface decision function obtained from the last round of iterative training, category labels are generated for all tensor data slices, and the ship target detection results are returned.

8. A processing device, It is characterized in that It includes computer program instructions, wherein when the computer program instructions are executed by a processing device, they are used to implement the steps corresponding to the polarization radar sea surface ship target detection method described in any one of claims 1-6.

9. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores computer program instructions, wherein the computer program instructions, when executed by a processor, are used to implement the steps corresponding to the polarization radar sea surface ship target detection method according to any one of claims 1 to 6.

Images