WO2021220337A1

WO2021220337A1 - Unwanted wave learning device, unwanted wave learning method, unwanted wave detection device, and unwanted wave detection method

Info

Publication number: WO2021220337A1
Application number: PCT/JP2020/017948
Authority: WO
Inventors: 竜馬谷▲高▼; 將白石
Original assignee: 三菱電機株式会社
Priority date: 2020-04-27
Filing date: 2020-04-27
Publication date: 2021-11-04
Also published as: JPWO2021220337A1; JP7130169B2

Abstract

This unwanted wave learning device is configured to be provided with: an unwanted wave graph generation unit (1) that, by connecting one or more unwanted waves erroneously detected from a radio wave image indicating the result of observation by a radar, generates an unwanted wave graph in which the unwanted waves are connected and outputs the unwanted wave graph as teaching data; and a learning model generation unit (2) that performs learning of the unwanted wave graph by using the radio wave image and the teaching data outputted from the unwanted wave graph generation unit 1 and generates a learning model for outputting the unwanted wave graph when the radio wave image is imparted.

Description

Unnecessary wave learning device, unnecessary wave learning method, unnecessary wave detection device and unnecessary wave detection method

The present disclosure relates to an unnecessary wave learning device and an unnecessary wave learning method for generating a learning model, and an unnecessary wave detection device and an unnecessary wave detection method for searching a region where an unnecessary wave exists.

In addition to the target, the radio wave image showing the observation result of the radar may show unnecessary waves such as a sea clutter or an ionospheric clutter. In the conventional target detection device that detects a target from a radio wave image, in order to improve the target detection performance, it is possible to suppress unnecessary waves reflected in the radio wave image before performing the process of detecting the target from the radio wave image. be.
As a processing for suppressing unnecessary waves, a constant false alarm probability processing (CFAR: Constant False Alarm Rate) is known.

In the following Non-Patent Document 1, it is not necessary to detect the circumscribing rectangle of the unwanted wave region, which is the region where the unwanted wave included in the radio wave image exists, by using a convolutional neural network (CNN). Wave detection methods are disclosed. The CNN learns the circumscribing rectangle of the unnecessary wave region in advance by using the radio wave image and the teacher data indicating the circumscribing rectangle of the unnecessary wave region. Is output.
If the conventional target detection device performs CFAR on the detected circumscribed rectangle in the unwanted wave region, the unwanted wave contained in the circumscribed rectangle can be suppressed.

The target signal may be included in the circumscribed rectangle of the unwanted wave region detected by the unwanted wave detecting method disclosed in Non-Patent Document 1. When the conventional target detection device performs CFAR on the circumscribed rectangle containing the target signal, the target signal is suppressed together with the unnecessary wave, so that the target may not be detected.
In order to reduce the possibility that the target signal is included in the circumscribed rectangle in the unwanted wave region, teacher data showing the circumscribed rectangle divided as finely as possible is prepared, and the CNN disclosed in Non-Patent Document 1 is disclosed. However, there is a problem that it is necessary to learn the circumscribed rectangle of the unnecessary wave region using the teacher data. It can be difficult to provide teacher data showing finely divided circumscribed rectangles.

This disclosure is made to solve the above-mentioned problems, and it is possible to output an unnecessary wave graph in which unnecessary waves are connected without preparing teacher data showing finely divided circumscribing rectangles. It is an object of the present invention to obtain an unnecessary wave learning device and an unnecessary wave learning method capable of generating a possible learning model.

The unnecessary wave learning device according to the present disclosure generates an unnecessary wave graph in which unnecessary waves are connected by connecting one or more unnecessary waves erroneously detected from a radio wave image showing an observation result of a radar, and is unnecessary. The unnecessary wave graph is learned by using the unnecessary wave graph generator that outputs the wave graph as teacher data, the radio wave image, and the teacher data output from the unnecessary wave graph generator, and when the radio wave image is given, it is unnecessary. It is provided with a learning model generation unit that generates a learning model that outputs a wave graph.

According to the present disclosure, it is possible to generate a learning model capable of outputting an unnecessary wave graph in which unnecessary waves are connected without preparing teacher data showing finely divided circumscribed rectangles.

It is a block diagram which shows the unnecessary wave learning apparatus which concerns on Embodiment 1. FIG. It is a hardware block diagram which shows the hardware of the unnecessary wave learning apparatus which concerns on Embodiment 1. FIG. It is a hardware block diagram of the computer when the unnecessary wave learning apparatus is realized by software, firmware, etc. It is a block diagram which shows the unnecessary wave detection apparatus which concerns on Embodiment 1. FIG. It is a hardware block diagram which shows the hardware of the unnecessary wave detection apparatus which concerns on Embodiment 1. FIG. It is a hardware block diagram of the computer when the unnecessary wave detection device is realized by software, firmware, etc. It is explanatory drawing which shows each function in the unnecessary wave learning apparatus shown in FIG. 1 and the unwanted wave detection apparatus shown in FIG. It is a flowchart which shows the unnecessary wave learning method which is the processing procedure of the unnecessary wave learning apparatus which concerns on Embodiment 1. FIG. It is explanatory drawing which shows the example of the unnecessary wave graph which does not satisfy the edge generation rule. It is explanatory drawing which shows an example of the minimum spanning tree. It is explanatory drawing which shows the problem of edge selection which occurs at the time of generation of the minimum spanning tree. It is explanatory drawing which shows the method of selecting an edge at the time of generating the minimum spanning tree in consideration of the power value of an unwanted wave. It is a conceptual diagram which shows the architecture of CNN. It is explanatory drawing which shows the design method of the correct answer feature map as the teacher data given at the time of learning of CNN. It is explanatory drawing which shows _{the generation rule Nj, k} ^* of a correct answer feature map. It is explanatory drawing which shows _{the generation rule Ej, k} ^* of a correct answer feature map. It is explanatory drawing which shows the generation example of a correct answer feature map N _j ^* , E _j ^*. It is a flowchart which shows the unnecessary wave detection method which is the processing procedure of the unnecessary wave detection apparatus which concerns on Embodiment 1. FIG. It is explanatory drawing which shows the processing content of the unnecessary wave detection apparatus which concerns on Embodiment 1. FIG. It is explanatory drawing which shows the peak detection of the output feature map _Nj ^T. It is explanatory drawing which shows the estimation of the edge e hat from the output feature map _Ej ^T. It is explanatory drawing which shows the estimation example of the unnecessary wave graph.

Hereinafter, in order to explain the present disclosure in more detail, a mode for carrying out the present disclosure will be described with reference to the attached drawings.

Embodiment 1.
FIG. 1 is a configuration diagram showing an unnecessary wave learning device according to the first embodiment.
FIG. 2 is a hardware configuration diagram showing the hardware of the unnecessary wave learning device according to the first embodiment.
The unnecessary wave learning device shown in FIG. 1 includes an unnecessary wave graph generation unit 1 and a learning model generation unit 2.

The unnecessary wave graph generation unit 1 is realized by, for example, the unnecessary wave graph generation circuit 11 shown in FIG.
The unnecessary wave graph generation unit 1 generates an unnecessary wave graph in which unnecessary waves are connected by connecting one or more unnecessary waves erroneously detected from a radio wave image showing an observation result of a radar.
The unnecessary wave graph generation unit 1 outputs the unnecessary wave graph as teacher data to the learning model generation unit 2.

The learning model generation unit 2 is realized by, for example, the learning model generation circuit 12 shown in FIG.
The learning model generation unit 2 learns the unnecessary wave graph using the radio wave image and the teacher data output from the unnecessary wave graph generation unit 1, and when the radio wave image is given, the learning model outputs the unnecessary wave graph. To generate. The learning model is realized by, for example, CNN.
The learning model generated by the learning model generation unit 2 is implemented as a learning model 32 in the unnecessary wave graph acquisition unit 31 of the unnecessary wave detection device shown in FIG.

In FIG. 1, it is assumed that each of the unnecessary wave graph generation unit 1 and the learning model generation unit 2, which are the components of the unnecessary wave learning device, is realized by the dedicated hardware as shown in FIG. That is, it is assumed that the unnecessary wave learning device is realized by the unnecessary wave graph generation circuit 11 and the learning model generation circuit 12.
Each of the unnecessary wave graph generation circuit 11 and the learning model generation circuit 12 includes, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), and an FPGA (Field-Programmable Gate). Array) or a combination of these is applicable.

The components of the unwanted wave learning device are not limited to those realized by dedicated hardware, but the unwanted wave learning device is realized by software, firmware, or a combination of software and firmware. It is also good.
The software or firmware is stored as a program in the memory of the computer. A computer means hardware that executes a program, and corresponds to, for example, a CPU (Central Processing Unit), a central processing unit, a processing unit, an arithmetic unit, a microprocessor, a microcomputer, a processor, or a DSP (Digital Signal Processor). do.

FIG. 3 is a hardware configuration diagram of a computer when the unnecessary wave learning device is realized by software, firmware, or the like.
When the unnecessary wave learning device is realized by software, firmware, or the like, a program for causing a computer to execute each processing procedure in the unnecessary wave graph generation unit 1 and the learning model generation unit 2 is stored in the memory 21. Then, the processor 22 of the computer executes the program stored in the memory 21.

Further, FIG. 2 shows an example in which each of the components of the unnecessary wave learning device is realized by dedicated hardware, and FIG. 3 shows an example in which the unnecessary wave learning device is realized by software, firmware, or the like. .. However, this is only an example, and some components in the unnecessary wave learning device may be realized by dedicated hardware, and the remaining components may be realized by software, firmware, or the like.

FIG. 4 is a configuration diagram showing an unnecessary wave detection device according to the first embodiment.
FIG. 5 is a hardware configuration diagram showing the hardware of the unwanted wave detection device according to the first embodiment.
The unnecessary wave detection device shown in FIG. 4 includes an unnecessary wave graph acquisition unit 31 and an unnecessary wave region search unit 33.

The unnecessary wave graph acquisition unit 31 is realized by, for example, the unnecessary wave graph acquisition circuit 41 shown in FIG.
The unnecessary wave graph acquisition unit 31 has a learning model 32 generated by the learning model generation unit 2 of the unnecessary wave learning device shown in FIG.
The unnecessary wave graph acquisition unit 31 acquires an unnecessary wave graph from the learning model 32 by giving a radio wave image to the learning model 32.
The unnecessary wave graph acquisition unit 31 outputs the acquired unnecessary wave graph to the unnecessary wave region search unit 33.

The unnecessary wave region search unit 33 is realized by, for example, the unnecessary wave region search circuit 42 shown in FIG.
The unnecessary wave area search unit 33 is an unnecessary wave area in which the unnecessary wave included in the radio wave image given to the learning model 32 exists from the unnecessary wave graph acquired by the unnecessary wave graph acquisition unit 31. To explore.

In FIG. 4, it is assumed that each of the unnecessary wave graph acquisition unit 31 and the unnecessary wave area search unit 33, which are the components of the unnecessary wave detection device, is realized by the dedicated hardware as shown in FIG. .. That is, it is assumed that the unnecessary wave detection device is realized by the unnecessary wave graph acquisition circuit 41 and the unnecessary wave region search circuit 42.
Each of the unnecessary wave graph acquisition circuit 41 and the unnecessary wave region search circuit 42 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof. do.

The components of the unwanted wave detector are not limited to those realized by dedicated hardware, but the unwanted wave detector is realized by software, firmware, or a combination of software and firmware. It is also good.
FIG. 6 is a hardware configuration diagram of a computer when the unwanted wave detection device is realized by software, firmware, or the like.
When the unnecessary wave detection device is realized by software, firmware, or the like, a program for causing a computer to execute each processing procedure in the unnecessary wave graph acquisition unit 31 and the unnecessary wave area search unit 33 is stored in the memory 51. Then, the processor 52 of the computer executes the program stored in the memory 51.

Further, FIG. 5 shows an example in which each of the components of the unnecessary wave detection device is realized by dedicated hardware, and FIG. 6 shows an example in which the unnecessary wave detection device is realized by software, firmware, or the like. .. However, this is only an example, and some components in the unwanted wave detection device may be realized by dedicated hardware, and the remaining components may be realized by software, firmware, or the like.

Next, the operation of the unnecessary wave learning device shown in FIG. 1 will be described.
FIG. 7 is an explanatory diagram showing the respective functions of the unnecessary wave learning device shown in FIG. 1 and the unnecessary wave detecting device shown in FIG.
FIG. 8 is a flowchart showing an unnecessary wave learning method which is a processing procedure of the unnecessary wave learning device according to the first embodiment.

First, the unnecessary wave graph generation unit 1 generates an unnecessary wave graph in which the unnecessary waves are connected by connecting one or more unnecessary waves erroneously detected from the radio wave image showing the observation result of the radar (FIG. Step 8 ST1).
The unnecessary wave graph generation unit 1 outputs the unnecessary wave graph as teacher data to the learning model generation unit 2.
An unwanted wave graph is a network structure in graph theory that includes a set of nodes and a set of edges. Nodes are the nodes of the graph or the vertices of the graph and are called unwanted wave plots. An edge is a branch of a graph or an edge of a graph that connects a node to another.

Hereinafter, the unnecessary wave graph generation process by the unnecessary wave graph generation unit 1 will be specifically described.
In CNN that realizes a learning model, not all unnecessary wave graphs can be used for learning, so it is necessary to generate unnecessary wave graphs that can be used for learning unnecessary wave graphs as teacher data.
The components of the unwanted wave graph G (V, E) are a set of node V = {v ₁ , v ₂ , ...} and edge E = {e ₁ , e ₂ , ...}. Node V is an unwanted wave plot erroneously detected by performing CFAR on a range Doppler map (hereinafter referred to as "RD (Ranger-Doppler) map"). The RD map is a two-dimensional map obtained from a radio wave image. Since the process itself for obtaining the RD map is a known technique, detailed description thereof will be omitted. In FIG. 7, the information related to the unwanted wave plot is referred to as plot information.
The edge E indicates the connection relationship between a certain unwanted wave plot and a certain unwanted wave plot.

The unnecessary wave graph G (V, E) that can be used for learning CNN needs to have regularity so that the answer can be uniquely determined.
The unnecessary wave graph generation unit 1 prepares the following assumptions (1), assumptions (2), and assumptions (3) as edge generation rules of the unnecessary wave graph G (V, E).
FIG. 9 is an explanatory diagram showing an example of an unnecessary wave graph that does not satisfy the edge generation rule.
Assumption (1) is that there is a way to connect between the nodes that are any two unwanted wave plots, and that all the nodes are connected from the starting node to the ending node. It is a production rule. A road means that there is an edge between the two nodes.
The example on the left side of FIG. 9 shows an unwanted wave graph that does not satisfy assumption (1). An unnecessary wave graph in which there is no path between a certain node cannot be used as teacher data because one unnecessary wave graph is divided into a plurality of parts.

Assumption (2) is a generation rule that there are no extra edges that would be a cycle. A cycle means that there is a loop starting from a certain node, passing through two or more other nodes, and returning to the starting node.
The example in the middle of FIG. 9 shows an unwanted wave graph that does not satisfy assumption (2). An unnecessary wave graph having a cycle cannot be used as teacher data because the number of combinations connecting the nodes becomes enormous unless the order of the nodes is restricted. The order is the number of edges that a node has.
Assumption (3) is a production rule that the sum of the edge lengths is the minimum. It is considered that the degree of relevance between the nodes is higher when the nodes having a short distance from each other are connected than when the nodes having a long distance from each other are connected to each other.
The unwanted wave graph shown on the right side of FIG. 9 does not satisfy the assumption (3) because the sum of the edge lengths is large and the sum is not the minimum.

The unnecessary wave graph generation unit 1 generates a minimum spanning tree as an unnecessary wave graph that satisfies each of assumptions (1), assumptions (2), and assumptions (3). The minimum spanning tree is a graph in which all nodes are connected and does not have a cycle. The minimum spanning tree is a graph in which the sum of edge weights is minimized.
The minimum spanning tree is composed of a weighted edge set T and is expressed by the following equation (1).

In the formula (1), _{w e} is the weight of the edge e.

FIG. 10 is an explanatory diagram showing an example of the minimum spanning tree.
In FIG. 10, the unwanted wave graph composed of the edges represented by thick lines is the minimum spanning tree.
A typical optimization algorithm for finding the minimum spanning tree from an arbitrary set of nodes includes the Kruskal method, the Prim method, and the Borvka method.
Weight w _e edge e can be treated as the distance between nodes edge e is connected. However, when dealing with weight w _e edge e as the distance between nodes, as shown in FIG. 11, as an alternative of the edge e, there may occur multiple choices (1) to (3).
FIG. 11 is an explanatory diagram showing a problem of edge selection that occurs when a minimum spanning tree is generated. In FIG. 11, each of (1) to (3) shows an edge option. In FIG. 11, the horizontal axis represents Doppler and the vertical axis represents range.
In FIG. 11, since all of the options (1) to (3) have the same length, it is impossible to uniquely select the option (1) to (3) only by the distance between the nodes.
However, considering the distribution of unwanted waves on the RD map, as shown in FIG. 11, since only option (2) is an edge passing through the unwanted wave, it is appropriate to select option (2). It is believed that there is.

Unnecessary wave graph generation unit 1 is, for example, the weight w _e edges e that connects the node v _i and node v _j, is calculated from the following formula (2).

In the formula (2), Z (p) is the signal power value of the two-dimensional coordinates p in the RD _map, v i, _{v j} represents the coordinate of the node.
u is a parameter when the position between the nodes to which the edge e is connected is represented by 0 to 1 as shown in FIG.
FIG. 12 is an explanatory diagram showing a method of selecting an edge when generating a minimum spanning tree in consideration of the power value of an unwanted wave.
12, the weight w _e is a value obtained by multiplying the minus line integral value of the signal power value of the edge e.
In the selection method shown in FIG. 12, in the generation of minimum spanning tree, for selecting with priority weight w _e is small edge e, the edge e is chosen according to the smallest choice weight w _e is (2).
The unnecessary wave graph generation unit 1 outputs the generated minimum spanning tree G (V, T) as a correct unnecessary wave graph to the learning model generation unit 2 as teacher data.

The learning model generation unit 2 learns the unnecessary wave graph by using the radio wave image and the teacher data output from the unnecessary wave graph generation unit 1, and when the radio wave image is given, the learning model outputs the unnecessary wave graph. Is generated (step ST2 in FIG. 8).
Hereinafter, the learning model generation process by the learning model generation unit 2 will be specifically described.

The learning model generation unit 2 constructs a CNN that learns an unnecessary wave graph in order to realize the learning model by CNN.
FIG. 13 is a conceptual diagram showing the architecture of CNN.
In the architecture shown in FIG. 13, the CNN is composed of two branches.
^{One branch is a node network φ t} (X) that estimates the node positions and node types in the unwanted wave graph given the RD map X. The RD map X is a two-dimensional map obtained from a radio wave image.
^{The other branch is an edge network Ψ t} (X) that estimates the edge position and edge type of the unwanted wave graph given the RD map X. R is the number of cells in the range direction in the RD map X, and V is the number of cells in the Doppler direction in the RD map X.
t = {1, 2, ..., T}, where t represents the stage number. 1 stage first (Stage t = 1), the node from the network phi ^t (X), calculates the loss values _f ^{N t = 1,} which will be described later, from the edge network Ψ ^t (X), described later loss value _f ^{E t = This} is the stage for calculating 1.
In the second and subsequent stages (Stage t = {2, ..., T}), the node network φ ^t (X) and the edge network Ψ ^t (X) are combined, and the node network φ ^t (X) and the edge network are combined. from the feature map is a bond results between Ψ ^{t (X),} a stage for calculating the respective loss value f _N ^t and loss values f _E ^t.

^{The learning model generation unit 2 learns each of the node network φ t} (X) and the edge network Ψ ^t (X) so that the loss function f shown in the following equation (3) is minimized.

f _N ^t is a loss value for the CNN estimation result corresponding to each of the node position and the node type in the unnecessary wave graph G (V, T) at the stage t of the CNN.
f _E ^t, in stage t of CNN, is the loss value for the estimation result of the CNN corresponding to respective positions and the edge of the types of edges in the unnecessary wave graph G (V, T).
N _j ^t, as shown in the following equation (6) is any one output characteristic map of the plurality of output characteristic map in a node network phi ^{t (X).}

E _j ^t, as shown in the following equation (7) is any one output characteristic map of the plurality of output characteristic map in the edge network [psi ^{t (X).}

p represents the position coordinates of the output feature map.
R'is the number of cells in the range direction in each output feature map, and V'is the number of cells in the Doppler direction in each output feature map.
Each of N _j ^* and E _j ^* is a correct answer feature map, and is generated from an unnecessary wave graph of the correct answer, which is teacher data.

J is the number of output characteristics map node network phi ^{t (X)} shown in FIG. 14, is set based on the type number or the like to be estimated.
FIG. 14 is an explanatory diagram showing a method of designing a correct answer feature map as teacher data given at the time of learning CNN.
If the type of estimation target is a sea clutter, the sea clutter is distributed in the range direction near the Doppler axis 0 [m / s] of the RD map.
If the type of estimation target is the ionospheric clutter, the ionospheric clutter is distributed in the Doppler direction at a distance of about 150 [km] or more of the range axis of the RD map.
In view of this fact, each of the sea clutter and the ionospheric clutter can be estimated by referring to the three regions as shown in FIG.
That is, the sea clutter can be estimated by referring to the region (1).
The ionospheric clutter existing on the positive side of the Doppler shaft can be estimated by referring to the region (2), and the ionospheric clutter existing on the negative side of the Doppler shaft can be estimated by referring to the region (3). be.

In the example of FIG. 14, J = 3, and the design is such that the unwanted wave is predicted for each type and region of the unwanted wave.
That is, the feature map for prediction of the sea clutter is N _{j = 1} ^t and E _{j = 1, 2} ^t , and the feature map for prediction of the ionospheric clutter existing on the plus side of the Doppler axis is N _{j = 2} ^t. And E _{j = 3,4} ^t , and the feature maps for prediction of the ionospheric clutter existing on the minus side are N _{j = 3} ^t and E _{j = 5,6} ^t .
Why the number of E _j ^t is 2J sheets is to represent the edge as a vector field.

The method of generating the correct feature maps _Nj ^* and _Ej ^{* will be described.}
The edge of the unnecessary wave graph, which is the teacher data, is generated as the minimum spanning tree by the unnecessary wave graph generation unit 1, and the position of the node of the minimum spanning tree depends on the setting of the hyperparameters of CFAR. Therefore, there is a problem that the position of the node has a large variance and the correct answer cannot be uniquely determined.
The learning model generation unit 2 expresses the unnecessary wave graph as two feature maps in order to deal with the above problem. That is, the learning model generation unit 2 represents the nodes of the unnecessary wave graph with a feature map based on a two-dimensional Gaussian distribution, and represents each point on the edge of the unnecessary wave graph with a feature map based on a two-dimensional vector.

The correct feature map _Nj ^* is expressed by the following equation (8).

In the equation (8), N _{j and k} ^* are correct heat maps of the kth unnecessary wave in the estimation target j, and are expressed by the following equation (9). The estimation target j is a sea clutter, an ionospheric clutter existing on the positive side of the Doppler shaft, or an ionospheric clutter existing on the negative side of the Doppler shaft.

In equation (10), v _{j, k, v} are the coordinates of the correct node v for the kth unnecessary wave in the estimation target j.
As shown in FIG. 15, the learning model generation unit 2 generates _{a Gaussian distribution in which v j, k, and v} are the vertices of the distribution.
Then, the learning model generation unit 2 generates a two-dimensional mixed Gaussian distribution by superimposing _{all v j, k, and v.} The two-dimensional mixed Gaussian distribution is a correct feature map _{Nj, k} ^* for the kth unnecessary wave in the estimation target j.
FIG. 15 is an explanatory diagram showing _{the generation rules Nj, k} ^* of the correct answer feature map.

The correct feature map _Ej ^* is expressed by the following equation (11).

In equation (11), E _{j, k, e} ^* are correct vector fields for the edge e of the kth unnecessary wave in the estimation target j. e is an edge between the node _{v 1} and node _{v 2.} A vector field is generated by the signal power value z on the edge e in the RD map.
Further, _{n j,} as shown in FIG. 16, in the region e, which can be generated by setting the appropriate width σ to the edge e, _{E j} is defined by the following equation _{^{(12), k, e *}} ( p) is the number of coordinates p that are not 0. FIG. 16 is an explanatory diagram showing _{the generation rules Ej, k} ^* of the correct answer feature map.

In the formula (12), || ▽ Z ( p) || , the gradient strength of the signal power value z (p) in the coordinates _{p, e j, k, vi} is the edge e that is connected to the node _{v i} direction of the unit _{vector, v j, k, vi} are the coordinates of the node _{v i.}
“P on e” is a symbol representing an operation of determining whether or not the coordinate p is on the edge e, and is determined by using the region e bar shown in FIG. In the text of the specification, because of the electronic application, it is not possible to add a "-" symbol above the characters, so it is written as an e-bar.

The above determination is expressed by the following equation (13).

FIG. 17 is an explanatory diagram showing an example of generating the correct answer feature maps _Nj ^* and _Ej ^*.
The learning model generation unit 2 generates a learning model in which the unnecessary wave graph is learned by learning the unnecessary wave graph using _{the correct answer feature maps Nj} ^* and _Ej ^*.
The learning model has learned parameters optimized by learning each of the correct feature map _Nj ^* ^{of the node network φ t} (X) and the correct feature map E _j ^* of the edge network Ψ ^{t (X).} ing.
The learning model generated by the learning model generation unit 2 is implemented as a learning model 32 in the unnecessary wave graph acquisition unit 31 of the unnecessary wave detection device shown in FIG.

FIG. 18 is a flowchart showing an unnecessary wave detection method which is a processing procedure of the unnecessary wave detection device according to the first embodiment.
FIG. 19 is an explanatory diagram showing the processing content of the unwanted wave detection device according to the first embodiment.
The unnecessary wave graph acquisition unit 31 has a learning model 32 generated by the learning model generation unit 2 of the unnecessary wave learning device shown in FIG.
The unnecessary wave graph acquisition unit 31 acquires an unnecessary wave graph from the learning model 32 by giving a radio wave image to the learning model 32 (step ST11 in FIG. 18).
The unnecessary wave graph acquisition unit 31 outputs the acquired unnecessary wave graph to the unnecessary wave region search unit 33.

Hereinafter, the unnecessary wave graph acquisition process by the unnecessary wave graph acquisition unit 31 will be specifically described.
As shown in FIG. 19, the unnecessary wave graph acquisition unit 31 estimates the unnecessary wave graph of each unnecessary wave region in the RD map by giving the RD map obtained from the radio wave image to the learning model 32.
Node network phi ^{t (X)} output feature map N _j ^T in, because they are represented as a heat map, coordinates of a specific node is not determined to a point. Output feature map _N ^{j T} is the output characteristic map in the final stage of the CNN.
Since the unnecessary wave graph cannot be estimated unless the coordinates of the specific node are determined at one point, the unnecessary wave graph acquisition unit 31 needs to estimate the specific node.
_{Since the output feature map Ej} ^T in the edge network Ψ ^t (X) is expressed as a vector field, no specific edge is determined. Output feature map _E ^{j T} is an output characteristic map in the final stage of the CNN.
If the specific edge is not fixed to one, the unnecessary wave graph cannot be estimated. Therefore, the unnecessary wave graph acquisition unit 31 needs to estimate the specific edge.

Unnecessary wave graph acquiring unit 31 detects the peak of the output characteristic map N _j ^T in estimation object j.
As shown in the lower left graph in FIG. 20, it is assumed that the output feature map _Nj ^{T has a plurality of mixed Gaussian distributions with large overlap.}
Figure 20 is an explanatory view showing a peak detection output characteristic map N _j ^T.
For example, NMS (Non-Maximum Supression) can be used to detect the peak of one Gaussian distribution while leaving one Gaussian distribution among a plurality of mixed Gaussian distributions having a certain overlap or more. NMS is a known method used for object detection in optical images. That is, NMS is a method of deleting the other detection windows while leaving only one detection window when the plurality of detected detection windows overlap.
The degree of overlap can be measured using IoU (Intersection over Union).
To detect the peak, the signal power value of the cell at the coordinate p is compared with the signal power value of the eight cells existing around the coordinate p, and the signal power value of the cell at the coordinate p is the signal power value of the eight cells. If it is larger than all of the signal power values, it is assumed that the cell at coordinate p is the peak v hat. In the text of the specification, the symbol "^" cannot be added above the characters due to the electronic application, so it is written like a v-hat.
The unnecessary wave graph acquisition unit 31 sets the detected peak v hat as the final estimated V peak hat = {v ₁ hat, v ₂ hat, ...}.

The unnecessary wave graph acquisition unit 31 obtains a plurality of edge e-hats from the _{output feature map Ej} ^T.
Since the output feature map _Ej ^T is expressed as a vector field, the unnecessary wave graph acquisition unit 31 considers the direction of the vector in each cell as shown in the following equation (14), and the signal power value. accordingly each edge e hat _{= {v j, vp} _{hat, v j, vq} hat} connection strength _L ei _{{v j, vp} _{hat, v j, vq} hat} is calculated.

Unnecessary wave graph acquisition unit 31, as shown in FIG. 21, a plurality of edge e in the hat connection strength L _ei, edge e hat the final estimated edge T hat according to the threshold λ or more connection strength L _ei = {E ₁ hat, e ₂ hat, ...}. The threshold value λ may be stored in the internal memory of the unnecessary wave graph acquisition unit 31 or may be given from the outside of the unnecessary wave detection device.
FIG. 21 is an explanatory diagram showing the estimation of the edge e-hat from the _{output feature map Ej} ^T.
Since _{each j in the} _{output feature map Nj} ^T and the output feature map ^{Ej T} indicates an estimation target, as shown in FIG. 22, from each of the output feature map N _j ^T and the output feature map E _j ^T. , The unwanted wave graph in the unwanted wave region can be estimated.
FIG. 22 is an explanatory diagram showing an estimation example of an unnecessary wave graph.

The unnecessary wave region search unit 33 searches for an unnecessary wave region from the unnecessary wave graph for each type acquired by the unnecessary wave graph acquisition unit 31 (step ST12 in FIG. 18).
That is, the unnecessary wave region search unit 33 searches for a region in which the distance L from the estimated edge T hat in each type of unwanted wave graph is _{within the threshold Th L} as an unnecessary wave region. The threshold value Th _L may be stored in the internal memory of the unnecessary wave region search unit 33, or may be given from the outside of the unnecessary wave detection device.
Here, the unnecessary wave region search unit 33 searches the region where the distance L from the estimated edge T hat in the unnecessary wave graph is _{within the threshold Th L} as the unnecessary wave region. However, this is only an example, and the unnecessary wave region search unit 33 may search the region obtained by performing graph cut for each type of unnecessary wave graph as the unnecessary wave region.
Graph cut is an algorithm that solves the minimum cutting problem for finding the boundary between the foreground area and the background area with reference to the foreground area and the background area roughly specified in advance. The unwanted wave region search unit 33 sets the unwanted wave graph in the foreground region, sets a region separated from the foreground region by an appropriate distance or more as the background region, and then applies the graph cut.
The unnecessary wave region search unit 33 suppresses unnecessary waves by performing CFAR on the searched unwanted wave region.

In the above-described first embodiment, by connecting one or more unnecessary waves erroneously detected from the radio wave image showing the observation result of the radar, an unnecessary wave graph in which the unnecessary waves are connected is generated, and the unnecessary wave graph is generated. The unnecessary wave graph is learned by using the unnecessary wave graph generation unit 1 that outputs the above as teacher data, the radio wave image, and the teacher data output from the unnecessary wave graph generation unit 1, and when the radio wave image is given, it is unnecessary. The unnecessary wave learning device is configured to include a learning model generation unit 2 that generates a learning model that outputs a wave graph. Therefore, the unnecessary wave learning device can generate a learning model capable of outputting an unnecessary wave graph in which unnecessary waves are connected without preparing teacher data showing finely divided circumscribed rectangles.

In the present disclosure, it is possible to modify any component of the embodiment or omit any component of the embodiment.

The present disclosure is suitable for an unwanted wave learning device and an unwanted wave learning method for generating a learning model.
The present disclosure is suitable for an unnecessary wave detection device and an unnecessary wave detection method for searching a region where an unnecessary wave exists.

1 Unnecessary wave graph generation unit, 2 Learning model generation unit, 11 Unnecessary wave graph generation circuit, 12 Learning model generation circuit, 21 Memory, 22 Processor, 31 Unnecessary wave graph acquisition unit, 32 Learning model, 33 Unnecessary wave area search unit, 41 Unnecessary wave graph acquisition circuit, 42 Unnecessary wave area search circuit, 51 Memory, 52 Processor.

Claims

By connecting one or more unnecessary waves that are erroneously detected from the radio wave image showing the observation result of the radar, an unnecessary wave graph in which the unnecessary waves are connected is generated, and the unnecessary wave graph is output as teacher data. Wave graph generator and
A learning model generation that learns an unnecessary wave graph using the radio wave image and the teacher data output from the unnecessary wave graph generation unit, and generates a learning model that outputs an unnecessary wave graph when a radio wave image is given. Unnecessary wave learning device equipped with a unit.
The unnecessary wave learning device according to claim 1, wherein the unnecessary wave graph generation unit generates a minimum spanning tree as the unnecessary wave graph.
The learning model generation unit generates a correct answer feature map showing the distribution of unnecessary waves from the teacher data, and learns an unnecessary wave graph by using the radio wave image and the correct feature map. Item 1. The unnecessary wave learning device according to Item 1.
The learning model generation unit generates a correct answer feature map showing the distribution of unnecessary waves for each type of unnecessary waves included in the teacher data, and uses the radio wave image and the correct answer feature map of each type. The unnecessary wave learning device according to claim 1, wherein the unnecessary wave graph is learned for each type of unnecessary wave.
The learning model generation unit is characterized in that the node of the unnecessary wave graph is a mixed Gaussian distribution and the edge of the unnecessary wave graph is a vector field to generate the correct answer feature map for each type of unnecessary wave. The described unwanted wave learning device.
The unwanted wave graph generator generates an unwanted wave graph in which the unwanted waves are connected by connecting one or more unwanted waves erroneously detected from the radio wave image showing the observation result of the radar, and the unwanted wave graph is generated. Is output as teacher data,
The learning model generation unit learns the unnecessary wave graph using the radio wave image and the teacher data output from the unnecessary wave graph generation unit, and when the radio wave image is given, the learning model outputs the unnecessary wave graph. Unnecessary wave learning method to generate.
It has a learning model generated by the learning model generation unit of the unnecessary wave learning device according to any one of claims 1 to 5, and by giving a radio wave image to the learning model, the said The unnecessary wave graph acquisition unit that acquires the unnecessary wave graph from the learning model,
From the unnecessary wave graph acquired by the unnecessary wave graph acquisition unit, the unnecessary wave area search unit that searches for the unnecessary wave region in which the unnecessary wave included in the radio wave image given to the learning model exists. Unwanted wave detector with and.
The unnecessary wave detection device according to claim 7, wherein the unnecessary wave graph acquisition unit acquires a feature map showing the distribution of unnecessary waves from the learning model, and generates an unnecessary wave graph from the feature map.
By detecting the peak of the feature map, the unnecessary wave graph acquisition unit estimates each of a plurality of nodes in the feature map, estimates the edge connecting the plurality of nodes, and the estimated node and the said node. The unwanted wave detection device according to claim 8, wherein the unwanted wave graph is generated from the estimated edge.
9. The unwanted wave region search unit searches for a region in the unwanted wave graph acquired by the unwanted wave graph acquisition unit within a threshold value of the distance from the estimated edge as the unwanted wave region. The unwanted wave detector described.
The seventh aspect of claim 7, wherein the unnecessary wave region search unit searches for a region obtained by performing a graph cut on the unnecessary wave graph acquired by the unnecessary wave graph acquisition unit as the unnecessary wave region. Unwanted wave detector.
The unnecessary wave graph acquisition unit gives a radio wave image to the learning model generated by the learning model generation unit of the unnecessary wave learning device according to any one of claims 1 to 5. Obtain an unnecessary wave graph from the learning model and
From the unnecessary wave graph acquired by the unnecessary wave graph acquisition unit, the unnecessary wave region search unit obtains an unnecessary wave region which is a region in which the unnecessary wave included in the radio wave image given to the learning model exists. Unnecessary wave detection method to search.