A method for detecting marine targets based on differential circular system visual map

Abstract

The application discloses a marine target detection method based on a differential circular system visual graph, constructs a straight line finite penetration visual graph and a circular finite penetration visual graph from a radar echo phase difference sequence respectively, and obtains a differential circular system visual graph by difference; and the differential circular system visual graph is trained and judged based on a graph neural network, so that the stability and reliability of detection under a complex sea clutter background are improved.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, specifically relating to a method for detecting marine targets based on a differential circle system visual diagram. Background Technology

[0002] When performing surface target surveillance and early warning missions, maritime radars are inevitably affected by sea clutter interference generated by ocean background scattering. Influenced by changes in sea state and observation conditions, the amplitude statistics, spatiotemporal correlation, and transform domain characterization of sea clutter often exhibit significant fluctuations, accompanied by localized strong scattering and non-uniformity caused by reference cell contamination. Under high-resolution systems, clutter peaks and tail enhancements are more pronounced, making it easy for weak target echoes to be submerged by strong clutter. Conventional detection procedures struggle to achieve a stable trade-off between constant false alarm rate (CFAR) control and detectability. While the traditional CFAR method based on energy comparison is simple in structure and easy to implement, target energy is often weaker than sea clutter fluctuations under low signal-to-clutter ratio conditions, leading to a significant degradation in detection performance.

[0003] To enhance discrimination capabilities against complex sea clutter backgrounds, existing research has attempted to map echo sequences into graph structure signals based on phase evolution and structural differences, and to mine inter-node correlations through graph learning, forming a promising intelligent detection approach. Among these, the visual graph method, by transforming time series into graph structures and utilizing topological differences for detection, provides a new technical approach for detection against sea clutter backgrounds. However, existing visual graph construction and graph learning detection still have certain limitations: on the one hand, when using a single visual graph construction rule, although the adjacency structures corresponding to the target and sea clutter differ, they may still contain a large number of common structures, which can weaken the ability to highlight local differences; on the other hand, at the graph neural network learning level, traditional GCNs rely on the full graph Laplacian matrix and usually assume that training and testing are performed on the same fixed graph, which belongs to a transmission-based learning paradigm and is difficult to adapt to the expansion requirements of new graphs or new nodes. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a method for detecting maritime targets based on differential circular system views. A straight-line finite penetration view and a circular finite penetration view are constructed from radar echo phase difference sequences, and differential circular system views are obtained by differential analysis. The differential circular system views are trained and decided based on graph neural networks, thereby improving the stability and reliability of detection in complex sea clutter backgrounds.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Construct a normalized phase difference sequence for the range cell to be measured using the measured data from the marine radar; Step 2: Map the normalized phase difference sequence obtained in Step 1 into an undirected, unweighted graph structure for graph learning; Step 3: Using the differential circle finite penetration view (DLPVG) obtained in Step 2 as the structural input, and combining node features, a graph neural network is used to extract the graph-level embedding representation and output the binary classification probability of "target / clutter". Then, a threshold calibration is performed on the validation set to ensure that the detection meets the given false alarm rate constraint. Step 4: After the network completes training, the output probability of the network is thresholded using the clutter samples of the validation set. The obtained threshold is then used in the testing or online detection stage to realize target / clutter decision and detection result output under constant false alarm rate control.

[0006] Preferably, step 1 specifically comprises: Step 1-1: Acquire measured data from the marine radar, in the range cell to be measured. The slow-time echo sequence is extracted and its complex baseband form is expressed as follows:

[0007] in, and The first n The in-phase and quadrature components of each pulse The total number of available pulses for the distance unit to be measured; Step 1-2: Calculate the instantaneous phase from the complex echo sequence:

[0008] in, To obtain the argument operation of complex numbers, the following is adopted: The obtained phase is the main value interval ; Steps 1-3: To suppress absolute phase instability and highlight slow-time evolution differences, the phases are spaced out as follows: The difference is used to obtain the phase difference sequence:

[0009] in, For the phase difference time interval, when The time represents the phase difference between pulses; This indicates phase wrapping processing that maps the phase difference to the principal value interval. Indicates the first The instantaneous phase corresponding to each pulse; The phase difference calculation described above is equivalent to taking the argument of the conjugate product:

[0010] in Indicates complex conjugation; Steps 1-4: For the phase difference sequence Normalization is performed to obtain the normalized phase difference sequence. :

[0011] in, This represents the normalization operation. Subsequently, a sliding window is used to truncate the slow time series, with a window length of [value missing]. Step size is From a length of Normalized phase difference sequence From N win A short-time window sample:

[0012] in .

[0013] Preferably, step 2 specifically comprises: Step 2-1: Define the sequence sampling points as graph nodes; for the first... n sampling points Let the corresponding node be . The node set is:

[0014] in, N This indicates the total number of sampling points in the current short-time window sample. Indicates the number of samples within the current short time window. n One normalized phase difference sample value; To ensure graph locality and reduce computational complexity, only graphs satisfying the following conditions are considered. node pairs Check visibility and decide whether to connect the edges, where M Indicates the maximum penetration distance; Step 2-2: First, give the line visibility conditions of the standard visibility view VG; for any node pair and , For any of them n satisfy ,like:

[0015] That is, the middle point will not obscure the nodes. and The line of sight between them is called a node. and Satisfy the line visibility condition in the standard visible graph VG, and establish an undirected edge in the VG graph. Otherwise, set to zero; Steps 2-3: Construct a finite-penetration view LPVG in the VG; allow connections and At most, it can penetrate during the process There are occlusion points; the set of intermediate points that violate the line visibility condition is defined as:

[0016] When both conditions are met:

[0017] At that time, the judgment and In LPVG, the finite penetration visibility condition is satisfied; therefore, the adjacency matrix of LPVG is defined as:

[0018] in, Represents the LPVG adjacency matrix of the th i Line 1 j The elements of the column. Due to the node and It is evident that relations do not distinguish between direction and weight, therefore It is also an undirected and unweighted matrix; Steps 2-4: Introduce the concept of circular systems into the adjacency matrix construction to establish a CLPVG of the phase difference sequence; CLPVG no longer uses a straight line passing through two points as the visibility boundary, but instead constructs a family of circular arcs that pass through two points at the same time, and uses the circular arc curves to replace the original straight lines as the new visibility boundary. Take any two sampling points in the time series and , Construct the equations for the circular system:

[0019] in To adjust the nonlinear parameters of the geometric shape of the circular system; and These represent the sampling time or sampling number of the two endpoints in the normalized phase difference sequence, respectively. These represent the normalized phase difference values ​​corresponding to the two endpoints, respectively. For any sampling time between two points In the circular system Find the corresponding arc point on the top , That is, CLPVG in Nonlinear visibility boundary value at the location; The penetration parameter is set as follows Statistics in and The set of midpoints that extend beyond the boundary of the arc. :

[0020] When the following conditions are met:

[0021] At that time, determine the node and There exists a circular finite penetration visibility condition, which is connected by edges in a CLPVG; therefore, the adjacency matrix of a CLPVG is defined as:

[0022] Step 2-5: For the same normalized phase difference short-time window sample, obtain the LPVG adjacency matrix according to steps 2-3 and 2-4 respectively. and CLPVG adjacency matrix ,in :

[0023] The adjacency matrix is:

[0024] in, Represents the elements in the DLPVG adjacency matrix; Equivalent representation from the perspective of edge sets:

[0025] in, , and These represent the edge sets corresponding to DLPVG, LPVG, and CLPVG graphs, respectively. With node set with difference edge set Visual diagram of the difference circle system: .

[0026] Preferably, step 3 specifically comprises: Step 3-1: Truncate the normalized phase difference sequence obtained in Step 1 using a sliding window method to obtain multiple windows with a length of... L For short-time window samples, construct the following for each subsequence according to step 2: and The DLPVG adjacency matrix is ​​obtained by difference. This forms a sample image:

[0027] in, Indicates the first m The LPVG adjacency matrix corresponding to each graph sample. Indicates the first m The CLPVG adjacency matrix corresponding to each graph sample Indicates the first m The DLPVG adjacency matrix corresponding to each graph sample; Indicates the first m The set of nodes in a graph sample Indicates the first m The difference edge set in each graph sample; Each image sample is assigned a binary label. 1 represents the target and 0 represents clutter, and the training set, validation set and test set are divided for model training and performance evaluation; Step 3-2: For each image sample For each node Construct node feature vectors And form a node feature matrix. , as node attributes that serve as network input; Step 3-3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input the graph convolutional layer of the graph neural network; let the first... l -1 level nodes are represented as For nodes neighborhood Symmetric aggregation is performed to obtain neighborhood messages. :

[0028] Then, concatenate the self-representation with the neighborhood message, perform a linear transformation, and activate to obtain the updated node representation:

[0029] in, For nonlinear activation functions, For learnable parameter matrices, This indicates vector concatenation; Steps 3-4: Introduce batch normalization and random deactivation Dropout after the graph convolutional layer, and use inter-layer bypass stacking to achieve deep updates:

[0030] in This represents the feature matrix formed by stacking all nodes. Indicates batch normalization, Indicates the first l The node feature matrix output from layer -1 Indicates the first l The feature matrix of intermediate nodes obtained by layer graph convolution; Steps 3-5: Embed the nodes of the last layer By reading out operators, they are aggregated into a graph-level representation. It employs a dual-channel readout method combining global average pooling and global max pooling, and concatenates the data along the channel dimension.

[0031]

[0032] in, This represents the graph-level features obtained through global average pooling. This represents the set of nodes in a graph sample. This represents the graph-level features obtained by global max pooling; When the output dimension of the graph convolutional layer is 64, the dimension of the read-out graph-level vector is 64×2=128.

[0033] Preferably, step 4 specifically comprises: Step 4-1: Set the expected false alarm rate The number of clutter diagram samples in the validation set was counted. And calculate the allowable number of clutter false alarms:

[0034] in This is a rounding up operation; Step 4-2: After completing the training of the graph neural network binary classification model in Step 3, input all clutter map samples in the validation set to obtain the corresponding target probability output. Sort the sequences from largest to smallest to obtain the following sequence:

[0035] Take the first Using this value as the constant false alarm threshold, the actual false alarm rate on the validation set is approximately:

[0036] Step 4-3: For any image sample to be detected, input the corresponding DLPPG image sample within the short time window of the distance cell into the network to obtain the target probability. According to the threshold The following judgment rules are given:

[0037] If the value is greater than the threshold, the target is declared to exist and labeled as 1; otherwise, the target is declared to not exist and labeled as 0. The corresponding decision result and its confidence level are output.

[0038] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the above-described maritime target detection method.

[0039] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for detecting maritime targets.

[0040] A chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the aforementioned maritime target detection method.

[0041] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the above-described maritime target detection method.

[0042] The beneficial effects of this invention are as follows: 1. A finite penetration visibility diagram (LPVG) method for differential circular system for sea clutter phase difference sequences is proposed. By constructing LPVG and CLPVG respectively and performing differential operations, only the incremental connection of circular visibility relative to straight line visibility is retained, thereby weakening the common structure in the two types of diagrams, reflecting the nonlinear additional connections more centrally, and enhancing the topological separability of the target and sea clutter. 2. The finite penetration visibility of the circular arc introduces nonlinear degrees of freedom in the visibility boundary, and can be adjusted between noise resistance and structural sensitivity through parameters, so that the mapping results have a better robust characterization ability for complex nonlinear sea clutter phase evolution. 3. Combine the network output probability with the constant false alarm rate threshold calibration module, calibrate the detection threshold on the validation set and use it for subsequent decision-making, so that the detection process strictly meets the preset false alarm rate constraint; make decisions on the image samples corresponding to the echo data through the final detector form. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall flow of the algorithm proposed in this invention; Figure 2 This is a schematic diagram of the radar echo phase difference sequence; Figure 3 A schematic diagram of the construction rules / algorithm for a linear finite penetration visibility graph (LPVG), (a) abstract representation of the visibility relationship of a sequence, (b) LPVG graph structure; Figure 4A schematic diagram of the composition rules / algorithm for a circular finite-penetration visible graph (CLPVG), (a) abstract representation of the visible relationship of the sequence, (b) LPVG graph structure; Figure 5 A schematic diagram of the construction rules / algorithm for a finite penetration visibility diagram (DLPVG) of a differential circle system, (a) abstract representation of the visibility relationship of a sequence, (b) CLPVG graph structure; Figure 6 This is a schematic diagram of a graph neural network structure used for classifying DLPPG graph samples. Figure 7 The following are the detection probability curves for the measured data: (a) IPIX data, (b) CSIR data. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] This invention proposes a method for detecting maritime targets based on differential circular system views: a straight-line finite penetration view and a circular finite penetration view are constructed from the radar echo phase difference sequence, and differential circular system views are obtained by differential analysis; the differential circular system views are trained and decided based on graph neural networks, thereby improving the stability and reliability of detection in complex sea clutter backgrounds.

[0046] This invention aims to address the problems of unstable structural representation in existing visualization-based mapping methods under strong and non-stationary sea clutter backgrounds, and the difficulty of adapting traditional graph convolutional networks to learning new image samples with a strong dependence on fixed graph structures, thereby improving the stability and reliability of marine target detection under complex sea clutter conditions.

[0047] To achieve the above objectives, the technical solution adopted by the present invention mainly includes the following steps: Step 1: Construct a phase difference sequence from sea clutter measurement data. Step 1 specifically includes: Step 1-1: Acquire measured data from the marine radar, in the range cell to be measured. The extracted slow-time echo sequence (sampling across pulses within the same distance cell) can be represented in complex baseband form as follows:

[0048] in, and The first n The in-phase and quadrature components of each pulse This represents the total number of pulses available for this range cell. This complex baseband modeling is a common representation of coherent radar echoes (the signal is complex after demodulation to baseband).

[0049] Step 1-2: Calculate the instantaneous phase from the complex echo sequence:

[0050] in, To determine the argument of a complex number, the following method is used: The obtained phase is usually the principal value interval .

[0051] Steps 1-3: To suppress absolute phase instability and highlight slow-time evolution differences, the phases are spaced out as follows: The difference is used to obtain the phase difference sequence:

[0052] in, Phase difference time interval ( The time represents the phase difference between pulses, which is often used in engineering to characterize the carrier phase change between pulses. The above can also be equivalent to taking the argument of the conjugate product:

[0053] in This represents the complex conjugate. The two forms are equivalent in implementation; this chapter uses the latter to facilitate direct calculation from the complex echo and avoid the propagation of phase expansion errors. This is a phase wrapping operation that maps the phase difference to the principal value range.

[0054] Steps 1-4: Use a sliding window to extract the slow time series, with a window length of [value missing]. Step size is From a length of Normalized phase difference sequence From N win A short-time window sample:

[0055] in

[0056] Step 2: Normalize the phase difference sequence of a certain distance cell obtained in Step 1. The mapping is performed as an undirected, unweighted graph structure for graph learning. First, using sequential sampling points as nodes, adjacency matrices of the Limited Penetration Visibility graph (LPVG) are constructed based on the linear limited penetration visibility. Based on the finite penetration visibility of arcs, an adjacency matrix of a circular limited penetration visibility graph (CLPVG) is constructed. Both are values ​​taken from The undirected, unweighted adjacency matrix. After obtaining... and Then, the differential concept is further employed to highlight the local incremental connection of the visibility of arcs relative to the visibility of lines, defining a differential circle system visibility diagram of the sequence. Step 2 specifically includes: Step 2-1: Define the sequence sampling points as graph nodes. For the first... n sampling points Let the corresponding node be . The node set is:

[0057] To ensure graph locality and reduce computational complexity, only graphs satisfying:

[0058] node pairs Check visibility and decide whether to connect the edges, where M This indicates the maximum penetration distance (or local window length).

[0059] Step 2-2: First, give the line visibility conditions for the standard visibility graph (VG). For any pair of nodes... and For any of them n satisfy ,like:

[0060] That is, the middle point will not obscure the nodes. and The line of sight between them is called a node. and This is visible in the VG graph, and an undirected edge is constructed in the graph. Otherwise, set it to zero.

[0061] Steps 2-3: Introduce the concept of finite penetration into the VG, constructing a finite penetration visible view LPVG. This allows for connections... and At most, it can penetrate during the process There are several occlusion points. The set of intermediate points that violate the line visibility condition is defined as:

[0062] When both conditions are met:

[0063] At that time, it was believed and This is visible in LPVG. Therefore, the adjacency matrix of LPVG is defined as:

[0064] It is an undirected and unweighted matrix.

[0065] Steps 2-4: To enhance nonlinear expressive power and mitigate boundary effects, the concept of circular systems is introduced into the adjacency matrix construction to establish a CLPVG of the phase difference sequence. CLPVG no longer uses a straight line passing through two points as the visibility boundary, but instead constructs a family of arcs that simultaneously pass through two points, replacing the original straight line as the new visibility boundary.

[0066] Take any two sampling points in the time series and Construct the equations for the circular system:

[0067] in To adjust the nonlinear parameters of the circular system's geometry, The larger the arc, the closer it is to a straight line. The smaller the bend, the more obvious it is.

[0068] For any sampling time between two points In the circular system Find the corresponding arc point on the top . That is, CLPVG in The nonlinear visibility boundary value at that location.

[0069] Given a limited penetration parameter, statistics are presented in... and The set of midpoints that extend beyond the boundary of the arc:

[0070] When the following conditions are met:

[0071] At that time, it is considered that the node and There exists circular finite penetration visibility, which is achieved by connecting edges in a CLPVG. Therefore, the adjacency matrix of a CLPVG is defined as:

[0072] Steps 2-5: After obtaining... and Subsequently, to highlight the "local incremental connectivity of arc visibility relative to line visibility," a Difference Circular Limited Penetration Visibility graph (DLPVG) is defined:

[0073] The adjacency matrix is:

[0074] Equivalent representation from the perspective of edge sets:

[0075] With node set with difference edge set Visual diagram of the difference circle system:

[0076] Step 3: Using the finite penetration visible view (DLPVG) of the differential circle system obtained in Step 2 as the structural input, and combining node features, a graph neural network is used to extract graph-level embedding representations and output the binary classification probability of "target / clutter". Subsequently, threshold calibration is performed on the validation set to ensure that the detection meets the given false alarm rate constraint. Step 3 specifically includes: Step 3-1: Truncate the normalized phase difference sequence of a certain distance cell obtained in Step 1 using a sliding window method to obtain multiple sequences of length [missing information]. L For each subsequence, construct the subsequence according to step 2. and The DLPVG adjacency matrix is ​​obtained by difference. This forms a sample image:

[0077] Each image sample is assigned a binary label. (1 represents the target, 0 represents clutter), and the training set, validation set, and test set are divided for model training and performance evaluation.

[0078] Step 3-2: For each image sample For each node Construct node feature vectors And form a node feature matrix. , which are node attributes used as network inputs.

[0079] Step 3-3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input the graph convolutional layer of the graph neural network. Let the first... l -1 level nodes are represented as For nodes neighborhood Symmetric aggregation is performed to obtain neighborhood messages. Common aggregation methods include mean aggregation:

[0080] Then, concatenate the self-representation with the neighborhood message, perform a linear transformation, and activate to obtain the updated node representation:

[0081] in, For nonlinear activation functions, For learnable parameter matrices, This indicates vector concatenation.

[0082] Steps 3-4: To improve training stability and suppress overfitting, batch normalization and random deactivation (Dropout) can be introduced after the graph convolutional layer, and inter-layer bypass stacking can be used to achieve more stable deep layer updates.

[0083] in This represents the feature matrix formed by stacking all nodes. This indicates batch normalization.

[0084] Steps 3-5: Embed the nodes of the last layer By reading out operators, they are aggregated into a graph-level representation. It can employ a dual-channel readout method combining global average pooling and global max pooling, and then concatenate the data along the channel dimension.

[0085]

[0086] When the output dimension of the graph convolutional layer is 64, the dimension of the read-out graph-level vector is 64×2=128.

[0087] Step 4: After the network completes training, to strictly meet the preset false alarm rate constraint, the output probability of the network is thresholded using clutter samples from the validation set. The obtained threshold is then used in the testing or online detection phase to achieve target / clutter decision and detection result output under constant false alarm rate control. The specific steps include: Step 4-1: Set the expected false alarm rate The number of clutter diagram samples in the validation set was counted. And calculate the allowable number of clutter false alarms:

[0088] in This is a rounding up operation.

[0089] Step 4-2: Input all clutter map samples from the validation set into the network obtained in Step 3 to obtain the corresponding target probability output. Sort the set in descending order to obtain the sequence:

[0090] Take the first Using this value as the constant false alarm threshold, the actual false alarm rate on the validation set is approximately:

[0091] Step 4-3: For any image sample to be detected (a DLPPG image sample corresponding to a certain distance cell and a certain short time window), input it into the network to obtain the target probability. According to the threshold The following judgment rules are given:

[0092] If the value is greater than the threshold, the target is declared to exist and labeled as 1; otherwise, the target is declared to not exist and labeled as 0. The corresponding decision result and its confidence level are output.

[0093] Example: To verify the effectiveness of the algorithm, this example uses publicly available IPIX sea clutter measured radar data as the verification dataset. IPIX data is typically organized by different observation times and scene files, with each data file containing slow-time complex echo sequences for multiple range cells. To facilitate the explanation of the specific implementation process of this invention, the following uses the 19931107_135603_starea17 file from the IPIX dataset as an example to illustrate the detection algorithm flow. Its relevant radar operating parameters are shown in Table 1, with data using the HV polarization mode selected.

[0094] Table 1. Parameters related to the IPIX radar dataset

[0095] The present invention proposes a flowchart for a method for detecting maritime targets based on a differential circle system visual diagram. Figure 1 The specific implementation steps are as follows: Step 1: Obtain the normalized phase difference sequence from the measured echo data and construct the window sample. Step 1 specifically includes the following steps: Step 1-1: Obtain the IPIX radar measured sea clutter data file as input. The data includes four polarization channels: HH, HV, VH, and VV. The range cells for target echoes and clutter can be determined according to the data description table. For example, in the data file "19931107_135603_starea17", the target range cell is cell 9, and the affected cells are cells 8, 10, and 11. Furthermore, considering that the number of clutter range cells is much greater than that of target range cells, directly using all clutter range cells would easily cause a serious imbalance in the number of target / clutter samples and lead to overfitting during the training phase. In this embodiment, only the above two clutter range cells (cells 7 and 12) are selected for training and testing. The samples in cell 9 are labeled as target class, and the samples in cells 7 and 12 are labeled as clutter class. For the selected polarization channel (HV), in any range cell... Read in-phase and quadrature component sequences and Constructing complex baseband echoes:

[0096] Step 1-2: Calculate the instantaneous phase from the complex echo sequence:

[0097] The phase is spaced as The difference is used to obtain the phase difference sequence:

[0098] in Indicates complex conjugation.

[0099] Steps 1-4: To facilitate subsequent parameter tuning of the circular system, the length is... The phase difference sequence is normalized as needed to obtain a normalized phase difference sequence. Then, a sliding window was used to extract the slow time series, with a window length of [value missing]. Step size is From the normalized phase difference sequence, we obtain N win A short-time window sample:

[0100] This example uses a sliding segmentation method with a window length of 256 pulses and a step size of 20 pulses to obtain a sufficient number of target window and clutter window samples within a limited observation time. The target unit obtains a total of 6,000 windows, while the clutter unit obtains 12,000 windows.

[0101] Step 2: Using the normalized phase difference sequence within a certain distance cell and a certain sliding window obtained in Step 1 as input, here we re-denote it as... Each sampling point in the sequence Defined as a node in the graph, denoted as One-dimensional time series are mapped to an LPVG graph structure using the finite penetration visibility rule, and a corresponding difference circle system visibility diagram is constructed. This embodiment sets the following mapping parameters: finite penetration threshold. Circular system parameters Step 2 specifically includes the following steps: Step 2-1: First, construct a line-based finite penetration visibility diagram based on the line-based finite penetration visibility rule. Randomly select two sampling points in the phase difference sequence. and For its nodes and nodes Calculate the position of the line connecting the two endpoints in the middle. Linear height at:

[0102] Then, by traversing all intermediate points, the following is satisfied:

[0103] Add the midpoint to the set of occlusion points And count the number of occlusion points. .

[0104] When the finite penetration criterion is satisfied:

[0105] Write the LPVG edges into the adjacency matrix:

[0106] Otherwise, set to zero; Step 2-2: Then construct the CLPVG graph of the sequence. For each pair of candidate nodes... : With endpoints , Establish the equations of the circular system and pass in the parameters Obtain the corresponding circular arc boundary. Then sample all intermediate positions. Calculate the boundary value of the arc at that location. Traverse the intermediate points to find the points that satisfy the condition. Add the midpoint to the set and statistics Similarly, satisfying At that time, write the CLPVG connection:

[0107] Otherwise, set to zero; Steps 2-3: After obtaining and Then, apply the difference preservation rule to all elements: For any If satisfied Then set it to:

[0108] Set all others to zero. For the current sliding window sample, output the adjacency matrix. .

[0109] Step 3: Building upon Step 2, construct node features for each graph sample and utilize a graph neural network for training and discrimination. Step 3 specifically includes the following steps: Step 3-1: For each sliding window sample obtained in Step 1 and Step 2, construct a graph sample. , where the set of nodes The edge set is determined by the adjacency matrix. Confirmed. Based on the distance cell from which the sliding window sample originates, assign category labels to the image samples: image samples from the target distance cell (cell 9) are labeled as target class; image samples from the clutter distance cells (cells 7 and 12) are labeled as clutter class.

[0110] Step 3-2: For each node in the graph sample Then, construct the corresponding node feature vector. The node feature vector consists of the following three parts: 1. Amplitude characteristics: node The corresponding amplitude feature is taken as the normalized echo amplitude value at the same sliding window and the same pulse position in step 1, denoted as

[0111] 2. CLPVG node degree characteristics: Based on the adjacency matrix of a circular finite penetration view compute nodes Node degree in CLPVG:

[0112] 3. DLPVG node degree characteristics: Based on the adjacency matrix of a circular finite penetration view compute nodes Node degree in DLPVG:

[0113] By concatenating the above three items in order, we obtain the node. Feature vectors:

[0114] Perform the above operation on all nodes in the graph to form a node feature matrix:

[0115] Step 3-3: Represent each graph sample as a triple:

[0116] in Represented by the adjacency matrix A defined graph structure The node feature matrix, The corresponding category labels are used. All image samples are randomly divided into training, validation, and test sets in a 6:2:2 ratio for network training, parameter selection, and performance evaluation.

[0117] Steps 3-4: Train the graph samples using a graph neural network. During training, the difference circle system visual diagram is used. As structural input, the node feature matrix As an attribute input, the node features are updated layer by layer through multi-layer graph convolution operations.

[0118] In each layer of graph convolution, node features are aggregated based on the adjacency relationships in the visual diagram of the difference circle system to generate new node representations. After multiple layers of graph convolution, a graph-level convergence operation is performed on the node features, mapping the node-level representations to a single graph-level feature vector.

[0119] Subsequently, graph-level features are mapped to classification scores through fully connected layers, and the network parameters are optimized using a supervised cross-entropy loss function. The network parameters are updated using the backpropagation algorithm until training converges or reaches the preset number of training epochs.

[0120] Steps 3-5: After network training is complete, each graph sample in the test set is input into the trained graph neural network to obtain the corresponding output score or target presence probability. This output serves as the detection statistic for that sliding window sample, used for constant false alarm rate control and final decision in subsequent steps.

[0121] Step 4: In Step 3, the network output score (or the probability of the target's presence) has been obtained for each sliding window image sample, denoted as... ,in m Number the sliding window samples. Step 4 performs constant false alarm rate (CFAR) control on the network output and provides the final detection results. Step 4 specifically includes: Step 4-1: In this embodiment, the network output score is first calculated on the clutter samples of the validation set. And set the target false alarm rate to .

[0122] Let the total number of clutter samples in the verification collection be The maximum allowed number of false alarms is then set as follows:

[0123] The scores of the verification clutter samples are sorted from largest to smallest, and denoted as:

[0124] The constant false alarm threshold is set as follows:

[0125] Step 4-2: For any sliding window sample in the test set m Take the network output score and give the detection judgment according to the following rules:

[0126] This completes the window-by-window decision output for all sliding window samples of the test data.

[0127] The effects of the present invention will be further explained below.

[0128] To demonstrate the improved effects of the method of the present invention, this embodiment... Under these conditions, the performance of the detector of this invention is compared with that of typical comparison methods, and a summary table of average detection accuracy (ACC) and ROC curves for all data files are provided (see Table 2 and ). Figure 7 .

[0129] Table 2. Summary of Average Detection Accuracy (ACC) Performance

Claims

1. A method for detecting maritime targets based on a differential circle system visual diagram, characterized in that, Includes the following steps: Step 1: Construct a normalized phase difference sequence for the range cell to be measured using the measured data from the marine radar; Step 2: Map the normalized phase difference sequence obtained in Step 1 into an undirected, unweighted graph structure for graph learning; Step 3: Using the differential circle finite penetration view (DLPVG) obtained in Step 2 as the structural input, and combining node features, a graph neural network is used to extract the graph-level embedding representation and output the binary classification probability of "target / clutter". Then, a threshold calibration is performed on the validation set to ensure that the detection meets the given false alarm rate constraint. Step 4: After the network completes training, the output probability of the network is thresholded using the clutter samples of the validation set. The obtained threshold is then used in the testing or online detection stage to realize target / clutter decision and detection result output under constant false alarm rate control.

2. The method for detecting maritime targets based on a differential circle system visual diagram according to claim 1, characterized in that, Step 1 specifically involves: Step 1-1: Acquire measured data from the marine radar, in the range cell to be measured. The slow-time echo sequence is extracted and its complex baseband form is expressed as follows: in, and The first n The in-phase and quadrature components of each pulse The total number of available pulses for the distance unit to be measured; Step 1-2: Calculate the instantaneous phase from the complex echo sequence: in, To obtain the argument operation of complex numbers, the following is adopted: The obtained phase is the main value interval ; Steps 1-3: To suppress absolute phase instability and highlight slow-time evolution differences, the phases are spaced out as follows: The difference is used to obtain the phase difference sequence: in, For the phase difference time interval, when The time represents the phase difference between pulses; This indicates phase wrapping processing that maps the phase difference to the principal value interval. Indicates the first The instantaneous phase corresponding to each pulse; The phase difference calculation described above is equivalent to taking the argument of the conjugate product: in Indicates complex conjugation; Steps 1-4: For the phase difference sequence Normalization is performed to obtain the normalized phase difference sequence. : in, This indicates a normalization operation; subsequently, a sliding window is used to truncate the slow time series, with a window length of [value missing]. Step size is From a length of Normalized phase difference sequence From N win A short-time window sample: in 。 3. The method for detecting maritime targets based on a differential circle system visual diagram according to claim 2, characterized in that, Step 2 specifically involves: Step 2-1: Define the sequence sampling points as graph nodes; for the first... n sampling points Let the corresponding node be . The node set is: in, N This indicates the total number of sampling points in the current short-time window sample. Indicates the number of samples within the current short time window. n One normalized phase difference sample value; To ensure graph locality and reduce computational complexity, only graphs satisfying the following conditions are considered. node pairs Check visibility and decide whether to connect the edges, where M Indicates the maximum penetration distance; Step 2-2: First, give the line visibility conditions of the standard visibility view VG; for any node pair and , For any of them n satisfy ,like: That is, the middle point will not obscure the nodes. and The line of sight between them is called a node. and Satisfy the line visibility condition in the standard visible graph VG, and establish an undirected edge in the VG graph. Otherwise, set to zero; Steps 2-3: Construct a finite-penetration view LPVG in the VG; allow connections and At most, it can penetrate during the process There are occlusion points; the set of intermediate points that violate the line visibility condition is defined as: When both conditions are met: At that time, the judgment and In LPVG, the finite penetration visibility condition is satisfied; therefore, the adjacency matrix of LPVG is defined as: in, Represents the LPVG adjacency matrix of the th i Line 1 j Column elements; due to nodes and It is evident that relations do not distinguish between direction and weight, therefore It is also an undirected and unweighted matrix; Steps 2-4: Introduce the concept of circular systems into the adjacency matrix construction to establish a CLPVG of the phase difference sequence; CLPVG no longer uses a straight line passing through two points as the visibility boundary, but instead constructs a family of circular arcs that pass through two points at the same time, and uses the circular arc curves to replace the original straight lines as the new visibility boundary. Take any two sampling points in the time series and , Construct the equations for the circular system: in To adjust the nonlinear parameters of the geometric shape of the circular system; and These represent the sampling time or sampling number of the two endpoints in the normalized phase difference sequence, respectively. These represent the normalized phase difference values ​​corresponding to the two endpoints, respectively. For any sampling time between two points In the circular system Find the corresponding arc point on the top , That is, CLPVG in Nonlinear visibility boundary value at the location; The penetration parameter is set as follows Statistics in and The set of midpoints that extend beyond the boundary of the arc. : When the following conditions are met: At that time, determine the node and There exists a circular finite penetration visibility condition, which is connected by edges in a CLPVG; therefore, the adjacency matrix of a CLPVG is defined as: Step 2-5: For the same normalized phase difference short-time window sample, obtain the LPVG adjacency matrix according to steps 2-3 and 2-4 respectively. and CLPVG adjacency matrix ,in : The adjacency matrix is: in, Represents the elements in the DLPVG adjacency matrix; Equivalent representation from the perspective of edge sets: in, , and These represent the edge sets corresponding to DLPVG, LPVG, and CLPVG graphs, respectively. With node set with difference edge set Visual diagram of the difference circle system: 。 4. The method for detecting maritime targets based on a differential circle system visual diagram according to claim 3, characterized in that, Step 3 specifically involves: Step 3-1: Truncate the normalized phase difference sequence obtained in Step 1 using a sliding window method to obtain multiple windows with a length of... L For short-time window samples, construct the following for each subsequence according to step 2: and The DLPVG adjacency matrix is ​​obtained by difference. This forms a sample image: in, Indicates the first m The LPVG adjacency matrix corresponding to each graph sample. Indicates the first m The CLPVG adjacency matrix corresponding to each graph sample Indicates the first m The DLPVG adjacency matrix corresponding to each graph sample; Indicates the first m The set of nodes in a graph sample Indicates the first m The difference edge set in each graph sample; Each image sample is assigned a binary label. 1 represents the target and 0 represents clutter, and the training set, validation set and test set are divided for model training and performance evaluation; Step 3-2: For each image sample For each node Construct node feature vectors And form a node feature matrix. , as node attributes that serve as network input; Step 3-3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input the graph convolutional layer of the graph neural network; let the first... l -1 level nodes are represented as For nodes neighborhood Symmetric aggregation is performed to obtain neighborhood messages. : Then, concatenate the self-representation with the neighborhood message, perform a linear transformation, and activate to obtain the updated node representation: in, For nonlinear activation functions, For learnable parameter matrices, This indicates vector concatenation; Steps 3-4: Introduce batch normalization and random deactivation Dropout after the graph convolutional layer, and use inter-layer bypass stacking to achieve deep updates: in This represents the feature matrix formed by stacking all nodes. Indicates batch normalization, Indicates the first l The node feature matrix output from layer -1 Indicates the first l The feature matrix of intermediate nodes obtained by layer graph convolution; Steps 3-5: Embed the nodes of the last layer By reading out operators, they are aggregated into a graph-level representation. It employs a dual-channel readout method combining global average pooling and global max pooling, and concatenates the data along the channel dimension. in, This represents the graph-level features obtained through global average pooling. This represents the set of nodes in a graph sample. This represents the graph-level features obtained by global max pooling; When the output dimension of the graph convolutional layer is 64, the dimension of the read-out graph-level vector is 64×2=128.

5. A method for detecting maritime targets based on a differential circle system visual diagram according to claim 4, characterized in that, Step 4 specifically involves: Step 4-1: Set the expected false alarm rate The number of clutter diagram samples in the validation set was counted. And calculate the allowable number of clutter false alarms: in This is a rounding up operation; Step 4-2: After completing the training of the graph neural network binary classification model in Step 3, input all clutter map samples in the validation set to obtain the corresponding target probability output. Sort the sequences from largest to smallest to obtain the following sequence: Take the first Using this value as the constant false alarm threshold, the actual false alarm rate on the validation set is approximately: Step 4-3: For any image sample to be detected, input the corresponding DLPPG image sample within the short time window of the distance cell into the network to obtain the target probability. According to the threshold The following judgment rules are given: If the value is greater than the threshold, the target is declared to exist and labeled as 1; otherwise, the target is declared to not exist and labeled as 0. The corresponding decision result and its confidence level are output.

6. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

8. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the method as described in any one of claims 1 to 5.

Images