Sea surface weak target detection method based on weighted difference visibility graph characteristics

Abstract

The invention discloses a sea surface weak target detection method based on weighted difference visibility map features, which comprises the following steps: collecting a sea surface echo amplitude sequence, performing subsequence division on the sea surface echo amplitude sequence, then converting the sea surface echo amplitude sequence into a phase difference sequence, and performing normalization processing on the phase difference sequence; constructing a weighted difference visibility graph network by using the normalized phase difference sequence; feature extraction is carried out on the weighted difference visibility map network corresponding to each subsequence, and sea clutter map feature vectors of the subsequences are constructed by using the extracted features; constructing a training sample set by taking each sea clutter map feature vector as a training sample point; training a target detector by using the training sample set; and for a to-be-detected sea surface echo amplitude sequence, taking the sea clutter map feature vector corresponding to the sequence as a to-be-detected sample point, and determining whether a target exists in the to-be-detected sea surface echo amplitude sequence through a relative position relationship between the to-be-detected sample point and the target detector.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target detection, and specifically relates to a method for detecting weak targets on the sea surface based on weighted difference visibility map features, which is suitable for target detection under low signal-to-clutter ratio conditions and improves the detection performance of radar for weak targets on the sea surface. Background Art

[0002] Detection of weak targets in complex sea clutter background is a major challenge for radar remote sensing. 2 ) is small, the speed is low, and the radar echo of the sea peak is similar to that of the target at a low grazing angle, which makes the sea clutter overlap with the weak target or even cover the target, seriously restricting the target detection performance of the maritime radar. The existing target detection method based on feature theory distinguishes sea clutter and target echo by studying their differences in the transform domain. However, the single feature extracted in the transform domain often has the disadvantages of requiring a long observation time and limited description of sea surface echo information when dealing with different sea conditions, which greatly affects the separability of the target and sea clutter in the feature space. Summary of the invention

[0003] In view of the complexity of the impact of sea clutter on radar target detection, the present invention provides a method for detecting weak targets on the sea surface based on weighted difference visibility map features, which solves the problem of weak target detection under the background of sea clutter from the perspective of combining multi-dimensional features in the image domain, and extracts weighted difference visibility map features of sea clutter and target echoes to achieve effective target detection.

[0004] In order to achieve the above tasks, the present invention adopts the following technical solutions:

[0005] A method for detecting faint targets on the sea surface based on weighted difference visibility map features, comprising:

[0006] Collecting a sea surface echo amplitude sequence, dividing the sea surface echo amplitude sequence into subsequences and converting it into a phase difference sequence, and normalizing the phase difference sequence;

[0007] The normalized phase difference sequence is used to construct a weighted difference visibility graph network;

[0008] Feature extraction is performed on the weighted difference visibility graph network corresponding to each subsequence, and the sea clutter graph feature vector of the subsequence is constructed using the extracted features;

[0009] Each sea clutter image feature vector is used as a training sample point to construct a training sample set; the target detector is trained using the training sample set;

[0010] For the sea surface echo amplitude sequence to be detected, the sea clutter map feature vector corresponding to the sequence is used as the sample point to be tested. Through the relative position relationship between the sample point to be tested and the target detector, it is determined whether there is a target in the sea surface echo amplitude sequence to be detected.

[0011] Furthermore, the converting the sea surface echo amplitude sequence into a phase difference sequence after dividing the sea surface echo amplitude sequence into subsequences includes:

[0012] The sea surface echo amplitude sequence received by the radar is divided into subsequences of length N. After division, the m-th subsequence x at the nth time is obtained. m (n); the phase difference sequence is calculated as follows:

[0013]

[0014] Where τ represents the time interval, the superscript * represents the complex conjugate, and arg(·) represents the phase of the complex variable.

[0015] Furthermore, the step of constructing a weighted difference visibility graph network using the normalized phase difference sequence includes:

[0016] Step 2.1, build visibility graph network VG

[0017] Based on the visibility graph theory, the visibility criterion for the phase difference sequence x(n), n=1,2,...,N is defined as follows:

[0018] x(k)<x(j)+(x(i)-x(j))(ik) / (ji),i<k<j

[0019] Where i, k, j represent the data points in the phase difference sequence, and x(i), x(j), and x(k) represent the amplitudes at the corresponding data points in the sequence;

[0020] For any two data points i and j in the sequence, if any point k between them and its amplitude satisfy the above formula, then the data points i and j are regarded as two mutually visible nodes in the visibility graph network, and there is an edge between the two nodes;

[0021] The normalized phase difference sequence of each subsequence is converted into a visibility graph network, where each data point in the phase difference sequence is a node in the network;

[0022] Step 2.2: Construct the horizontal visibility graph network HVG

[0023] Two data points i and j are considered as two connected nodes in the horizontal visibility graph network if the following geometric criteria are met in the sequence x(a):

[0024] x(i),x(j)>x(k) (for all k,i<k<j)

[0025] According to the above definition, the visibility map network established for the phase difference sequence is converted into a horizontal visibility map network network;

[0026] Step 2.3, construct the difference visibility graph network DVG

[0027] The difference visibility graph network is defined as follows:

[0028] E(DVG)=E(VG) / E(HVG), V(DVG)=V(VG)=V(HVG)

[0029] Wherein, E(DVG), E(VG) and E(HVG) represent the sets of edges of the DVG network, VG network and HVG network respectively, and V(DVG), V(VG) and V(HVG) represent the sets of nodes of the DVG network, VG network and HVG network respectively;

[0030] Through the above formula, on the basis of VG network and HVG network, the phase difference sequence is represented by DVG network;

[0031] Step 2.4, construct the weighted difference visibility graph network WDVG

[0032] On the basis of the DVG network, the arctan function is used to assign weights to the edges between two nodes with a connection relationship in the DVG network to characterize the changes between the quantity data points with a link relationship in the phase difference sequence, thereby obtaining a weighted difference visibility graph network; the definition of the weight is as follows:

[0033] W ij =arctan(x(j)-x(i)) / (ji)

[0034] Among them, W ij Represents the weight of the edge between node i and node j in the DVG network; the weighted adjacency matrix WA is defined as follows:

[0035]

[0036] Among them, W ij represents the element in the i-th row and j-th column of the adjacency matrix; if W ij =W ji ≠0, indicating that node i is connected to node j, and the edge weight is W ij ; If W ij =W ji =0, indicating that there is no edge between the two nodes.

[0037] Furthermore, the feature extraction of the weighted difference visibility graph network corresponding to each subsequence includes:

[0038] (1) Weighted graph weight peak height WPH

[0039] WPH=max(W ij )

[0040] Among them, W ij represents the element in the i-th row and j-th column of the weighted adjacency matrix WA;

[0041] (2) Weighted Graph Complexity (WGC)

[0042] WGC=4κ(1-κ)

[0043] κ=λ max -2cos(π / (N+1)) / N-1-2cos(π / (N+1))

[0044] Among them, λ max represents the maximum eigenvalue of the weighted adjacency matrix WA;

[0045] (3) Weighted Graph Entropy (WGE)

[0046]

[0047] Among them, k i represents the degree of node i in WDVG, and p(k) represents the distribution of the degree.

[0048] Furthermore, the extracted features are used to construct the sea clutter map feature vector of the subsequence, specifically:

[0049] ρ m =[ρ1(x m ),ρ2(x m ),ρ3(x m )] T

[0050] Among them, x m represents the mth sea surface echo subsequence, ρ1(x m ),ρ2(x m ),ρ3(x m ) means for the subsequence x m The vector composed of the WPH features extracted in the corresponding WDVG, the vector composed of the WGC features extracted, and the vector composed of the WGE features extracted. The superscript T represents the transposition operation.

[0051] Furthermore, each sea clutter image feature vector is used as a training sample point to construct a training sample set; and the target detector is trained using the training sample set, including:

[0052] Step 5.1: construct a training sample set ρ. Each training sample point in the training sample set is a sea clutter map feature vector ρ constructed for the subsequence. m ; and calculate the original convex hull S containing all sea clutter map feature vectors based on the training sample points n ;

[0053] Step 5.2: Use the training samples to convex hull S n Specifically, the redundant local space and its corresponding triangulation surface are found in the convex hull, and the training samples are selected as the inline points for internal segmentation.

[0054] Step 5.3, when performing the internal sectioning, it may cause the internal training sample points contained in the convex hull area to appear in S after the internal sectioning operation. n In the case of external points, such training sample points are called external points; after the internal sectioning operation is completed, S n To perform external operations:

[0055] For S n , find the closest subdivision surface to the external point, and construct a new external subdivision surface with the three edges of the external point and the subdivision surface to surround all training sample points; finally, delete the redundant subdivision surface to complete the update;

[0056] Step 5.4, set the number of iterations, repeat step 5.3, and obtain the final concave hull decision region Ω;

[0057] Step 5.5, deleting a specified number of training sample points on the generated concave hull decision area surface;

[0058] Step 5.6: When the specified false alarm rate requirement is met, stop the iteration and convert the concave hull decision area Ω fina as an object detector.

[0059] Furthermore, the selection criteria of the inner section point are as follows: find the sectioning surface with the largest perimeter △ max and the longest side, and then find the corresponding training sample point as the inline point P o , and then use P o Construct 4 new subdivision surfaces and add them to S n , and delete △ max To update S n , complete S n The internal section.

[0060] Furthermore, the specified number of training sample points on the generated concave hull decision area surface are deleted as follows:

[0061] Step 5.5.1, based on the number Q of training sample points in the training sample set ρ and the preset false alarm rate Pf , calculate the number of false alarm points: N f =Q×P f ;

[0062] Step 5.5.2, based on the principle of maximizing the volume loss of the concave hull decision area, delete one training sample point at a time from all training sample points that satisfy the minimum concave hull decision area; based on the training sample set after deleting the training sample points, iteratively repeat steps 5.1 to 5.5 to complete the update of the training sample set and the concave hull decision area.

[0063] Furthermore, for the sea surface echo amplitude sequence to be detected, the sea clutter map feature vector corresponding to the sequence is used as a sample point to be detected, and whether there is a target in the sea surface echo amplitude sequence to be detected is determined by the relative position relationship between the sample point to be detected and the target detector, including:

[0064] For the sea surface echo amplitude sequence to be detected, subsequence division, phase difference sequence conversion, normalization processing, weighted difference visibility graph network construction, feature extraction, sea clutter map feature vector construction are carried out in sequence and used as the sample points to be tested;

[0065] If the ratio of the sample points to be tested that fall outside the concave hull decision area to all sample points to be tested is greater than a preset value, it is considered that there is a target in the sea surface echo amplitude sequence to be detected.

[0066] A terminal device comprises a processor, a memory and a computer program stored in the memory; when the processor is executed by a computer, the method for detecting faint targets on the sea surface based on weighted difference visibility graph features is implemented.

[0067] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the method for detecting faint targets on the sea surface based on weighted difference visibility graph features is implemented.

[0068] Compared with the prior art, the present invention has the following technical features:

[0069] The present invention proposes a method for detecting weak targets on the sea surface based on weighted difference visibility graph features. The method analyzes the underlying dynamic characteristics of the echo data sequence by constructing a weighted difference visibility graph of the radar received echo signal; extracts three features, WPH, WGC, and WGE, in the graph domain, and uses a concave hull learning algorithm to construct a target detector with controllable false alarms to complete the detection of weak targets on the sea surface. The results of the test using the IPIX radar measured data set show that the difference graph features extracted by the present invention have good separability in the feature space; the algorithm proposed in the present invention still has higher detection performance than the existing feature joint detection algorithm under the detection condition of low signal-to-noise ratio, and can effectively complete the detection of weak targets on the sea surface. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of the overall process of the method of the present invention;

[0071] Figure 2 The clutter feature sample point distribution, concave hull decision area and sample point distribution of the unit to be tested are obtained by extracting the sea clutter data measured by IPIX;

[0072] Figure 3 10 sets of IPIX measured sea clutter data at a false alarm rate of 1e -3 Under the same conditions, the detection performance of the method of the present invention is compared with that of the existing feature detection methods. DETAILED DESCRIPTION

[0073] In order to improve the detection performance of targets in complex sea surface environments, some scholars have proposed the comprehensive use of multi-dimensional information and the integration of high-dimensional refined information to provide a more refined description of complex sea surface echoes. Many studies have extracted joint features from different domains, including multi-fractal, phase, speckle and polarization features for target detection. Due to the complexity of the impact of sea clutter on radar detection, in order to fully explore the differences between sea clutter signals and target signals, detection methods are developing in the direction of absorbing emerging disciplines and technologies and mining information from different transformation fields. Network theory has shown significant advantages in the analysis of the dynamic characteristics of time series, and has been widely used in discrete signal processing, computer biology, the Internet industry and other fields in recent years.

[0074] Based on the above idea, the present invention analyzes the visibility graph (VG) and its variants in complex network theory, proposes a feature extraction method combined with weighted different visibility graph (WDVG), designs a concave hull detector with controllable false alarm to construct the sea clutter decision area, so as to improve the weak target detection performance under low signal-to-noise ratio conditions.

[0075] See also Figure 1 The present invention proposes a method for detecting faint targets on the sea surface based on weighted difference visibility graph features, comprising the following steps:

[0076] Step 1: collect the sea surface echo amplitude sequence for target detector training, and preprocess each sea surface echo amplitude sequence; the specific steps are as follows:

[0077] Step 1.1: Divide the sea surface echo amplitude sequence received by the radar into subsequences of length N. After division, the mth subsequence x at the nth time is obtained. m (n); the phase difference sequence is calculated as follows:

[0078]

[0079] Where τ represents the time interval, the superscript “*” represents the complex conjugate, arg(·) represents the phase of the complex variable, N is the subsequence length, that is, the time corresponding to the subsequence, and each moment n corresponds to a data point.

[0080] Step 1.2, normalize the phase difference sequence, map the data between [0,1] to reduce the difference between the data, and obtain the preprocessed phase difference sequence. The specific formula is as follows:

[0081]

[0082] Among them, x(n) represents the normalized phase difference sequence, φ mmax represents the maximum value in the phase difference sequence, φ mmin Represents the minimum value in the phase difference sequence.

[0083] Step 2: Use the normalized phase difference sequence to construct a weighted difference visibility graph network.

[0084] Step 2.1, build the visibility graph network (VG network).

[0085] Based on the visibility graph theory, the visibility criterion for the phase difference sequence x(n), n=1,2,...,N is defined as follows:

[0086] x(k)<x(j)+(x(i)-x(j))(ik) / (ji),i<k<j

[0087] Where i, k, j represent the data points in the phase difference sequence, and x(i), x(j), and x(k) represent the amplitudes at the corresponding data points in the sequence.

[0088] For any two data points i and j in the sequence, if any point k between them and its amplitude satisfy the above formula, then the data points i and j are regarded as two mutually visible nodes in the visibility graph network, and there is an edge between the two nodes; traverse all the data points in the sequence to obtain the adjacency matrix A:

[0089]

[0090] Among them, A ij represents the element in the i-th row and j-th column of the adjacency matrix; if A ij =A ji =1, indicating that node i is connected to node j; if A ij =A ji =0, indicating that there is no edge between the two nodes.

[0091] The normalized phase difference sequence of each subsequence is converted into a visibility graph network according to the above definition, where each data point in the phase difference sequence is a node in the network; the obtained VG network is G(V,E), where V represents the node set, corresponding to the amplitude of the original signal sequence; E represents the set of edges between nodes.

[0092] Step 2.2, construct the horizontal visibility graph network (HVG network). The horizontal visibility graph (HorizontalVisibility Graph, HVG) is an improved version of VG, which is computationally faster than VG, easy to process and allows analysis. HVG is defined as follows:

[0093] If the following geometric criteria are met in the sequence x(a), two data points i and j are regarded as two connected nodes in the HVG network:

[0094] x(i),x(j)>x(k) (for all k,i<k<j)

[0095] According to the above definition, the VG network established for the phase difference sequence is converted into an HVG network.

[0096] Step 2.3, construct the difference visibility graph network (DVG network).

[0097] In this scheme, the behavior difference sequence is finally represented by the Differential Visibility Graph (DVG) network; the DVG network is an extension based on the VG network and the HVG network, emphasizing the analysis of the differences between sequence data; the definition of DVG is as follows:

[0098] E(DVG)=E(VG) / E(HVG), V(DVG)=V(VG)=V(HVG)

[0099] Among them, E(DVG), E(VG) and E(HVG) represent the sets of edges of the DVG network, VG network and HVG network respectively, and V(DVG), V(VG) and V(HVG) represent the sets of nodes of the DVG network, VG network and HVG network respectively.

[0100] Through the above formula, the phase difference sequence can be represented by a DVG network based on the VG network and the HVG network.

[0101] Step 2.4, construct the weighted difference visibility graph network WDVG.

[0102] The DVG network constructed in step 2.3 is a binary network, which only gives the link information whether there is or not between two nodes; while the weighted graph that assigns edge weights according to the link strength of the edges in the DVG network retains the weight information in the network topology and can obtain more robust results.

[0103] Based on the DVG network, the present invention uses the arctan function to assign weights to the edges between two nodes with a connection relationship in the DVG network to characterize the changes between the quantity data points with a link relationship in the phase difference sequence, thereby obtaining a weighted difference visibility graph network; the definition of the weight is as follows:

[0104] W ij =arctan(x(j)-x(i)) / (ji)

[0105] Among them, W ij Represents the weight of the edge between node i and node j in the DVG network; the weighted adjacency matrix WA is defined as follows:

[0106]

[0107] Among them, W ij represents the element in the i-th row and j-th column of the adjacency matrix; if W ij =W ji ≠0, indicating that node i is connected to node j, and the edge weight is W ij ; If W ij =W ji =0, indicating that there is no edge between the two nodes.

[0108] Step 3: Perform feature extraction on the weighted difference visibility graph network.

[0109] Step 2 has obtained the weighted difference visibility graph network corresponding to each subsequence of the sea surface echo amplitude sequence; for the weighted difference visibility graph network corresponding to each subsequence, the following feature extraction is performed:

[0110] (1) Weighted graph weight peak height (WPH)

[0111] The weight information of the weighted graph edge plays an important role in distinguishing the strength of the WDVG edge link relationship, that is, the criticality. Therefore, WPH is defined as:

[0112] WPH=max(W ij )

[0113] Among them, W ij Represents the element in the i-th row and j-th column of the weighted adjacency matrix WA.

[0114] (2) Weighted Graph Complexity (WGC)

[0115] The Graph Complexity (GC) represents the global complexity of the graph structure and adjacency matrix. In WDVG, the weight of the edge is considered to be included in the calculation of WGC. The definition is as follows:

[0116] WGC=4κ(1-κ)

[0117] κ=λ max -2cos(π / (N+1)) / N-1-2cos(π / (N+1))

[0118] Among them, λ max Represents the maximum eigenvalue of the weighted adjacency matrix WA.

[0119] (3) Weighted Graph Complexity (WGE)

[0120] Graph entropy is a metric used to describe the complexity of a graph, reflecting the nonlinear dynamic characteristics of the system and is effective even in low SCR conditions. It is defined as follows:

[0121]

[0122] Among them, k i represents the degree of node i in WDVG, p(k) represents the distribution of degree; the degree k of node i in WDVG i Refers to the number of edges associated with node i. The larger the degree, the more important the role the node usually plays in the network.

[0123] Step 4: Based on the extracted three features of WPH, WGC and WGE, construct the sea clutter feature vector ρ of the subsequence m , expressed as:

[0124] ρ m =[ρ1(x m ),ρ2(x m ),ρ3(x m )] T

[0125] Among them, x m represents the mth sea surface echo subsequence, ρ1(x m ),ρ2(x m ),ρ3(x m ) means for the subsequence x mThe vector composed of the WPH features extracted in the corresponding WDVG, the vector composed of the WGC features extracted, and the vector composed of the WGE features extracted. The superscript T represents the transposition operation.

[0126] Step 5: Build the target detector.

[0127] The traditional fast convex hull learning detection method has the phenomenon of decision space redundancy, which is caused by the uneven distribution of features in space and non-convexity. In view of the above shortcomings, the present invention designs a concave hull detector with controllable false alarms based on the fast convex hull learning algorithm, and the specific steps are as follows:

[0128] Step 5.1: construct a training sample set ρ. Each training sample point in the training sample set is a sea clutter map feature vector ρ constructed for the subsequence. m ; and calculate the original convex hull S containing all sea clutter map feature vectors based on the training sample points n .

[0129] Step 5.2: Use the training samples to convex hull S n Specifically, by finding the redundant local space and its corresponding triangulation surface (represented by △) in the convex hull, and selecting appropriate internal training samples as internal segmentation points for internal segmentation, the spatial redundancy can be reduced.

[0130] The selection criteria for the internal cutting point are as follows: Find the cutting surface with the largest perimeter △ max and the longest edge ridge, and then find the corresponding training sample point as the inlier point P o , and then use P o Construct 4 new subdivision surfaces and add them to S n , and delete △ max To update S n , complete S n The internal section.

[0131] Step 5.3, after the internal section operation S n The shape of is more consistent with the feature vector distribution. However, when performing the introjection, in order to reduce the spatial redundancy, it is possible that the internal training sample points contained in the convex hull area will appear in the updated S after the introjection operation. n In the case of external points, such training sample points are called external points; therefore, after the internal sectioning operation is completed, it is necessary to n To perform external operations:

[0132] For S n The dividing surface △ k , find the subdivision surface closest to the external point, and construct a new external subdivision surface with the three edges of the external point and the subdivision surface to surround all training sample points; finally, delete the redundant subdivision surface to complete the update.

[0133] Step 5.4, set the number of iterations, repeat step 5.3, and obtain the final concave hull decision area Ω.

[0134] Step 5.5: After the convex hull is converted into the concave hull decision region Ω, in order to meet the requirement of constant false alarm rate, it is necessary to delete a specified number of training sample points on the surface of the generated concave hull decision region, as follows:

[0135] Step 5.5.1, based on the number Q of training sample points in the training sample set ρ and the preset false alarm rate P f , calculate the number of false alarm points: N f =Q×P f ;

[0136] Step 5.5.2, based on the principle of maximizing the volume loss of the concave hull decision area, delete one training sample point at a time from all training sample points (false alarm points) that satisfy the minimum concave hull decision area; based on the training sample set after deleting the training sample points, iteratively repeat steps 5.1 to 5.5 to complete the update of the training sample set and the concave hull decision area.

[0137] Step 5.6, when the specified false alarm rate (such as P fa =0.001), the iteration is stopped and the concave hull decision area Ω is fina as an object detector.

[0138] Step 6, for the sea surface echo amplitude sequence to be detected, process it according to the method of steps 1-4 to obtain the sea clutter map feature vector corresponding to each subsequence of the sequence;

[0139] Each sea clutter map feature vector is used as a sample point to be tested, and the sample point to be tested is detected to see whether it is in the concave hull decision area Ω of the target detector. fina To determine whether there is a target outside the concave hull decision area Ω fina If the ratio of the sample points to be tested other than the target to all the sample points to be tested is greater than the preset value, it is considered that there is a target in the sea surface echo amplitude sequence to be detected.

[0140] Figure 3 10 sets of IPIX radar measured sea clutter data are shown with a false alarm rate of 1e -3 , performance comparison of the proposed detector under HV polarization conditions with the original three-feature detector and the detector based on graph connectivity density. Figure 3 It can be seen that the target detection performance of the proposed detector is better than that of the comparison method under the condition of shorter observation time (0.256s).

[0141] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, a person skilled in the art should understand that the specific implementation modes of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A method for detecting faint targets on the sea surface based on weighted difference visibility map features, characterized in that: include: Collecting a sea surface echo amplitude sequence, dividing the sea surface echo amplitude sequence into subsequences and converting it into a phase difference sequence, and normalizing the phase difference sequence; The normalized phase difference sequence is used to construct a weighted difference visibility graph network; Feature extraction is performed on the weighted difference visibility graph network corresponding to each subsequence, and the sea clutter graph feature vector of the subsequence is constructed using the extracted features; Each sea clutter image feature vector is used as a training sample point to construct a training sample set; the target detector is trained using the training sample set; For the sea surface echo amplitude sequence to be detected, the sea clutter map feature vector corresponding to the sequence is used as the sample point to be tested. Through the relative position relationship between the sample point to be tested and the target detector, it is determined whether there is a target in the sea surface echo amplitude sequence to be detected.

2. The method for detecting faint targets on the sea surface based on weighted difference visibility graph features according to claim 1, characterized in that: The step of dividing the sea surface echo amplitude sequence into subsequences and converting them into a phase difference sequence includes: The sea surface echo amplitude sequence received by the radar is divided into subsequences of length N. After division, the m-th subsequence x at the nth time is obtained. m (n); the phase difference sequence is calculated as follows: Where τ represents the time interval, the superscript * represents the complex conjugate, and arg(·) represents the phase of the complex variable.

3. The method for detecting faint targets on the sea surface based on weighted difference visibility graph features according to claim 1, characterized in that: The method of constructing a weighted difference visibility graph network using the normalized phase difference sequence includes: Step 2.1: Build the visibility graph network VG Based on the visibility graph theory, the visibility criterion for the phase difference sequence x(n), n=1,2,...,N is defined as follows: x(k)<x(j)+(x(i)-x(j))(ik) / (ji),i<k<j Where i, k, j represent the data points in the phase difference sequence, and x(i), x(j), and x(k) represent the amplitudes at the corresponding data points in the sequence; For any two data points i and j in the sequence, if any point k between them and its amplitude satisfy the above formula, then the data points i and j are regarded as two mutually visible nodes in the visibility graph network, and there is an edge between the two nodes; The normalized phase difference sequence of each subsequence is converted into a visibility graph network, where each data point in the phase difference sequence is a node in the network; Step 2.2: Construct the horizontal visibility graph network HVG Two data points i and j are considered as two connected nodes in the horizontal visibility graph network if the following geometric criteria are met in the sequence x(a): x(i),x(j)>x(k) (for all k,i<k<j) According to the above definition, the visibility map network established for the phase difference sequence is converted into a horizontal visibility map network network; Step 2.3, construct the difference visibility graph network DVG The difference visibility graph network is defined as follows: E(DVG)=E(VG) / E(HVG), V(DVG)=V(VG)=V(HVG) Wherein, E(DVG), E(VG) and E(HVG) represent the sets of edges of the DVG network, VG network and HVG network respectively, and V(DVG), V(VG) and V(HVG) represent the sets of nodes of the DVG network, VG network and HVG network respectively; Through the above formula, based on the VG network and the HVG network, the phase difference sequence is represented by the DVG network; Step 2.4, construct the weighted difference visibility graph network WDVG On the basis of the DVG network, the arctan function is used to assign weights to the edges between two nodes with a connection relationship in the DVG network to characterize the changes between the quantity data points with a link relationship in the phase difference sequence, thereby obtaining a weighted difference visibility graph network; the definition of the weight is as follows: W ij =arctan(x(j)-x(i)) / (j-i) Among them, W ij Represents the weight of the edge between node i and node j in the DVG network; the weighted adjacency matrix WA is defined as follows: Among them, W ij represents the element in the i-th row and j-th column of the adjacency matrix; if W ij =W ji ≠0, indicating that node i is connected to node j, and the edge weight is W ij ; If W ij =W ji =0, indicating that there is no edge between the two nodes.

4. The method for detecting faint targets on the sea surface based on weighted difference visibility graph features according to claim 1, characterized in that: The feature extraction of the weighted difference visibility graph network corresponding to each subsequence includes: (1) Weighted graph weight peak height WPH WPH=max(W ij ) Among them, W ij represents the element in the i-th row and j-th column of the weighted adjacency matrix WA; (2) Weighted Graph Complexity (WGC) WGC=4κ(1-κ) k=l max -2cos(π / (N+1)) / N-1-2cos(π / (N+1)) Among them, λ max represents the maximum eigenvalue of the weighted adjacency matrix WA; (3) Weighted Graph Entropy (WGE) Among them, k i represents the degree of node i in WDVG, and p(k) represents the distribution of the degree.

5. The method for detecting faint targets on the sea surface based on weighted difference visibility graph features according to claim 1, characterized in that: The extracted features are used to construct the sea clutter map feature vector of the subsequence, specifically: r m =[ρ1(x m ),ρ2(x m ),ρ3(x m )] T Among them, x m represents the mth sea surface echo subsequence, ρ1(x m ),ρ2(x m ),ρ3(x m ) means for the subsequence x m The vector composed of the WPH features extracted in the corresponding WDVG, the vector composed of the WGC features extracted, and the vector composed of the WGE features extracted. The superscript T represents the transposition operation.

6. The method for detecting faint targets on the sea surface based on weighted difference visibility graph features according to claim 1, characterized in that: The method uses each sea clutter image feature vector as a training sample point to construct a training sample set; Use the training sample set to train the target detector, including: Step 5.1: construct a training sample set ρ. Each training sample point in the training sample set is a sea clutter map feature vector ρ constructed for the subsequence. m ; and calculate the original convex hull S containing all sea clutter map feature vectors based on the training sample points n ; Step 5.2: Use the training samples to convex hull S n Specifically, the redundant local space and its corresponding triangulation surface are found in the convex hull, and the training samples are selected as the inline points for internal segmentation. Step 5.3, when performing the internal sectioning, it may cause the internal training sample points contained in the convex hull area to appear in S after the internal sectioning operation. n In the case of external points, such training sample points are called external points; after the internal sectioning operation is completed, S n To perform external operations: For S n , find the closest subdivision surface to the external point, and construct a new external subdivision surface with the three edges of the external point and the subdivision surface to surround all training sample points; finally, delete the redundant subdivision surface to complete the update; Step 5.4, set the number of iterations, repeat step 5.3, and obtain the final concave hull decision region Ω; Step 5.5, deleting a specified number of training sample points on the generated concave hull decision area surface; Step 5.6: When the specified false alarm rate requirement is met, stop the iteration and convert the concave hull decision area Ω fina as an object detector.

7. The method for detecting faint targets on the sea surface based on weighted difference visibility graph features according to claim 1, characterized in that: The selection criteria of the internal sectioning point are as follows: find the sectioning surface with the largest perimeter △ max and the longest side, and then find the corresponding training sample point as the inline point P o , and then use P o Construct 4 new subdivision surfaces and add them to S n , and delete △ max To update S n , complete S n The internal section.

8. The method for detecting faint targets on the sea surface based on weighted difference visibility graph features according to claim 1, characterized in that: The specific method of deleting a specified number of training sample points on the surface of the generated concave hull decision area is as follows: Step 5.5.1, based on the number Q of training sample points in the training sample set ρ and the preset false alarm rate P f , calculate the number of false alarm points: N f =Q×P f ; Step 5.5.2, based on the principle of maximizing the volume loss of the concave hull decision area, delete one training sample point at a time from all training sample points that satisfy the minimum concave hull decision area; According to the training sample set after deleting the training sample points, iteratively repeat steps 5.1 to 5.5 to complete the update of the training sample set and the concave hull decision area.

9. The method for detecting faint targets on the sea surface based on weighted difference visibility graph features according to claim 1, characterized in that: The method comprises: taking the sea clutter map feature vector corresponding to the sequence as the sample point to be tested, and determining whether there is a target in the sea surface echo amplitude sequence to be detected according to the relative position relationship between the sample point to be tested and the target detector, including: For the sea surface echo amplitude sequence to be detected, subsequence division, phase difference sequence conversion, normalization processing, weighted difference visibility graph network construction, feature extraction, sea clutter map feature vector construction are carried out in sequence and used as the sample points to be tested; If the ratio of the sample points to be tested that fall outside the concave hull decision area to all sample points to be tested is greater than a preset value, it is considered that there is a target in the sea surface echo amplitude sequence to be detected.

10. A computer-readable storage medium, wherein a computer program is stored in the medium; characterized in that: When the computer program is executed by a processor, the method for detecting faint sea surface targets based on weighted difference visibility map features according to any one of claims 1 to 9 is implemented.

Images