JP2021190040A - Learning device for underwater target detection and underwater target detector - Google Patents

Abstract

To provide a system that realizes highly accurate target detection using many learning models that depend on observation conditions on an underwater target detector that uses a neural network.SOLUTION: A learning model aggregate 114 in which multiple learning models for observation data of underwater targets accumulated together with observation conditions 101 thereof is created, distance calculation means 109 that calculates the distance among calculation means 101 is provided, one or more learning models supported by model selection means 108 are selected based on the distance among the conditions of the learning model aggregate 114, and an underwater target is estimated by inference means 122 based on the extracted learning model.SELECTED DRAWING: Figure 1

Description

The present invention relates to a technique for detecting an underwater target using observation data. Among them, in particular, it relates to a technique for executing learning by machine learning used for detecting an underwater target.

The observation data includes sound waves and image data obtained by converting the observed sound waves.

Conventionally, Patent Document 1 is a conventional technique relating to an active sonar for detecting an underwater target. In Patent Document 1, a judge having specialized knowledge can send a plurality of continuous pulses to a target object and display a screen showing the intensity of the reflected echo on the vertical axis and the direction on the horizontal axis. The existence of the target was judged.

Here, in the detection of an underwater target using an active sonar, it is conceivable to substitute a neural network which is the result of machine learning. Specifically, the intensity of the reflected echo is shown as the distance on the vertical axis and the orientation on the horizontal axis, that is, the B scope screen is used as the input image, and the echo of the target contained in the screen is not a human but a neural network. Substitutes the detection of underwater targets. Patent Documents 2 and 3 are conventional techniques to which a neural network is applied in a sonar device for detecting an underwater target.

Japanese Unexamined Patent Publication No. 8-129065 Japanese Unexamined Patent Publication No. 2019-200175 Japanese Unexamined Patent Publication No. 2010-266280

With respect to the above Patent Documents 2 and 3, in the detection using the sound wave of the underwater target, the observed sonar data largely depends on the observation conditions, and a single model data is required for the data under all the conditions. It is difficult to maintain high detection accuracy. Therefore, it is conceivable to prepare a plurality of learning models according to individual conditions and switch the learning models according to the conditions.

It has been difficult to achieve higher detection accuracy corresponding to various observation conditions with the above-mentioned artificial intelligence (AI) -based underwater target detection system. Therefore, it is an object of the present invention to provide such an underwater target detection device. A further object of the present invention is to provide a method for constructing such an underwater target detection device.

In order to solve the above problems, in the present invention, a plurality of learning models for detecting an underwater target are prepared, and based on the distance between the observation conditions of the observation data and the observation conditions of the plurality of learning models. It selects one or more learning models.

More specifically, in a learning device for an underwater target for detecting an underwater target using observation data, a plurality of learning data used for detecting the underwater target are used as observation conditions for the underwater target. A storage unit that stores the observation data in association with each other, an input unit that accepts the observation data of the underwater target and the observation conditions of the observation data, the received observation conditions, and a plurality of observation conditions stored in the storage unit. Using a distance calculation unit that calculates each distance, a model selection unit that selects a learning model from the storage unit based on the distance calculated by the distance calculation unit, and a learning model selected by the model selection unit. It has a learning unit that learns about the detection of the underwater target, and uses the learning model learned by the learning unit to make inferences about detection of the underwater target received by the input unit. It is a learning device for underwater target detection that makes it possible to perform.

As another aspect of the present invention, in the underwater target detection device that detects an underwater target or an underwater target related to the navigation of a ship by using observation data, a plurality of learning data used for detecting the underwater target is obtained in the underwater. A storage unit that stores the observation conditions for the target object in association with each other, an input unit that receives the observation data of the underwater target object and the observation conditions of the observation data, the received observation conditions, and the storage unit that stores the observation data. Learning from the storage unit based on the distance calculation unit that calculates the distance of each of the plurality of observation conditions, the distance calculated by the distance calculation unit, and the navigation route that the underwater navigation object or the vessel plans to navigate. An underwater target having a model selection unit that selects a model and an inference unit that makes an inference regarding detection of the underwater target object received by the input unit based on the learning model selected by the model selection unit. Includes object detectors.

According to the present invention, it is possible to detect an underwater target that exhibits higher detection accuracy corresponding to various observation conditions to which AI is applied.

It is a figure which showed the functional structure of Example 1. FIG. It is a figure which showed the list of observation data used in Example 1. FIG. It is a figure which showed the composition content of the example of the observation conditions used in Example 1. FIG. It is a figure which shows an example of the conversion method of the conditional distance in Example 1. FIG. It is a figure which shows an example of one expression form of the conditional space map used in Example 1. FIG. It is a figure which shows an example which described the learning model aggregate used in Example 1 on the conditional space map. It is a figure which cut out the content of the learning model aggregate used in Example 1 in a certain conditional space. It is a figure which shows the content of the condition data used in Example 1 of this invention. It is a figure which shows an example of the evaluation value of the conditional distance used in Example 1 of this invention. It is a figure which shows the inference performance data used in Example 1 of this invention. It is a figure which shows an example of the set of the multi-class learning model used in Example 1 of this invention. It is a figure which showed the functional structure of Example 2. FIG. It is a figure which shows the structure contents of the loop program used in Example 2. It is a system block diagram of Example 1. FIG. It is a flowchart which shows the processing outline of Example 1. FIG. It is a system block diagram of Example 2. It is a flowchart which shows the processing outline of Example 2.

Hereinafter, each embodiment of the present invention will be described. In each embodiment, an underwater target is detected for detecting an obstacle in the navigation of an underwater navigation object or a ship, or for searching using an underwater navigation object or a ship. Examples of underwater targets include schools of fish, mines, people, and other floating and drifting objects. These are just examples and do not exclude other objects. Further, in each embodiment, at least a part of the underwater target may appear on the water or the sea. The underwater navigation object or ship includes a moving body that navigates underwater or underwater, in addition to water and sea. In the following description, when the term "ship" is simply used, it means an underwater sailing object or a ship.

First, Example 1 of the present invention will be described. FIG. 1 is a functional configuration diagram of this embodiment. In FIG. 1, 100 is sonar data, 101 is observation conditions when the sonar data 100 is measured or observed, 102 is annotation data, 103 is model data, 104 is performance data, and 105 is a target position inference result. Further, 106 is a system interface, 107 is a model reading / writing means, 108 is a model selection means, 109 is a distance calculation means, 110 is a data division means, 111 is a learning process means, 112 is a preprocessing means, and 113 is an inference process means. show. Further, 114 is a learning model set, 115 is a condition map, 116 is condition data, 117 is duplicated Raw data, 118 is Raw data, 119 is a multi-class learning model, and 120 is inference performance data. Further, 121 is a learning means, 122 is a reasoning means, 123 is an output multidimensional score of a learning means, 124 is an output multidimensional score of a reasoning means, 125 is a score evaluation function, 126 is a score evaluation function, and 127 is a performance analysis. Show the means. The system also includes pointers to Raw data 118.

This system can be implemented in the underwater target detection system shown in FIG. Here, before explaining each function and processing of this system, the system configuration of this embodiment will be described with reference to FIG. In FIG. 14, this system is composed of an underwater target object detection learning device 10 and a terminal device 20 installed in a data center or the like, and an underwater target object detection device 34 provided on a ship 30 capable of navigating on water. Further, the ship 30 is provided with a sonar transmission / reception device 31, sensors 32, and a wireless device 33.

Here, the sonar transmission / reception device 31 has a function of converting the detected sound wave into an electric signal. Then, the sonar transmission / reception device 31 itself or another device converts the electric signal into sonar data 100 which is image data. The sonar transmission / reception device 31 includes a so-called active sonar.

Further, the sensors 32 have a function of detecting at least a part of the observation condition 101. For example, it can be realized by a GPS sensor that measures the position information of the sea area and a water temperature gauge that measures the underwater temperature profile which is the water temperature at each water depth. Of the observation conditions 101, the map information stored in advance by the underwater target detection device 34 may be used for the sea quality and the like.

Further, the underwater target object detection learning device 10 and the underwater target object detection device 34 are connected via the network 40. At this time, the underwater target detection device 34 is connected to the network 40 via the wireless device 33 installed on the ship 30. The network 40 may be a public line such as the Internet or a dedicated line.

By being connected in this way, the underwater target detection device 34 can notify the underwater target detection learning device 10 of the sonar data 100 measured or specified by the underwater target detection device 34 and the observation condition 101. Further, the learning result of the underwater target detection learning device 10 can be notified to the underwater target detection device 34, and the underwater target detection device 34 can make an inference regarding the detection of the underwater target. The data capacity of the above-mentioned sonar data 100, observation condition 101, and learning result may be large. In this case, it is desirable to exchange these via a storage medium 60 such as an HDD or a DVD instead of the network 40.

Next, the details of the learning device 10 for detecting an underwater target will be described. The underwater target detection learning device 10 can be realized by a so-called computer, and is composed of a processing unit 11 such as a CPU, a storage unit 12 such as a memory, and an input / output unit 13 such as an interface for inputting / outputting data. Then, each function is realized by the processing unit 11 performing an operation according to the program expanded in the storage unit 12. That is, the model selection unit 1080, the distance calculation unit 1090, the data division unit 1100, the learning process unit 1110, the preprocessing unit 1120, the learning unit 1210, and the performance analysis unit 1270 are used as a program or as a subject of calculation according to the program. , Will be realized. Here, each of these constituent requirements in the processing unit 11 has the following correspondence with each functional block shown in FIG. 1.

The model selection unit 1080 corresponds to the model selection means 108. The distance calculation unit 1090 corresponds to the distance calculation means 109. The data dividing unit 1100 corresponds to the data dividing means 110. The learning process unit 1110 corresponds to the learning process means 111. The pretreatment unit 1120 corresponds to the pretreatment means 112. The learning unit 1210 corresponds to the learning means 121. The performance analysis unit 1270 corresponds to the performance analysis means 127.

Further, it is desirable that the learning device 10 for detecting an underwater target is composed of a server, particularly among computers. In this case, the underwater target detection learning device 10 receives input from the terminal device 20 mounted as a computer, and outputs the processing result to the terminal device 20. Further, the underwater target detection learning device 10 is connected to the database 50. Various information shown in FIG. 1 is stored in the database 50.

Further, the underwater target detection device 34 can also be mounted as a computer. That is, it has a processing unit, a storage unit, and an input / output unit. The processing unit includes an inference process unit 1130 and an inference unit 1220. The inference process unit 1130 and the inference unit 1220 are realized as a program or as a subject of an operation according to the program, as in the configuration of the processing unit 11. Further, the inference process unit 1130 corresponds to the inference process means 113 of FIG. 1, and the inference unit 1220 also corresponds to the inference means 122.

Further, the input / output unit 13 of FIG. 14 corresponds to the system interface 106 and the model reading / writing means 107 of FIG. Some functions of the system interface 106 and the model reading / writing means 107 also correspond to the input / output unit of the underwater target detection device 34.

The underwater target object detection learning device 10 and the underwater target object detection device 34 may be integrated. In this case, it is desirable to mount the function of the underwater target object detection learning device 10 on the underwater target object detection device 34. Further, the division of functions between the underwater target object detection learning device 10 and the underwater target object detection device 34 is not limited to the content shown in FIG. For example, the underwater target detection device 34 may have a distance calculation unit 1090.

Next, the outline of the processing in this embodiment will be described. Then, after that, various information used in this embodiment and details of processing using each information will be described.

FIG. 15 is a flowchart showing an outline of the processing in this embodiment. In the following description, the main body of the process will be described as each functional block shown in FIG.

First, in step S1, the system interface 106 receives the sonar data 100, the observation condition 101, and the annotation data 102 which is the teacher data.

Next, in step S2, the distance calculation means 109 calculates the distance between the observation conditions input in step S1 and each observation condition of the learning model aggregate 114. This makes it possible for the distance calculation means 109 to specify the distance to the learning model associated with each observation condition of the learning model aggregate 114.

Next, in step S3, the model selection means 108 works differently in learning and inference. At the time of learning, one position inside the learning model set 114 that stores the learning model obtained as a result of learning is determined. At this time, the observation conditions are stored in the condition data 116, and the position of the condition map 115 is specified. At the time of inference, the learning model used for learning in step S4 is selected from the learning model aggregate 114 based on the distance of each learning model specified in step S2.

Next, in step S4, it is desirable that the process is executed only at the time of learning, and the learning process means 111 and the learning means 121 collaborate to execute the learning process. At the time of inference, it is desirable that step S4 skips this process and proceeds to step S5.

Finally, it is desirable that step S5 be executed only at the time of inference, and the inference process means 113 and the inference means 122 use the learning model selected in step S3 to make an underwater target for the sonar data 100. Performs inference processing for object detection. In step S5, it is desirable to skip this process during learning.

This is the end of the description of the outline of the process of this embodiment. Hereinafter, the details of the processing in this embodiment will be described with reference to FIG. 1 together with the explanation of the information to be used.

Here, a functional block for executing step S1 in FIG. 15 will be described. As this premise, the operation of the sonar transmission / reception device 31 of this embodiment will be described. The storage of the information described below means that the information is appropriately stored in the database 50, the storage unit 12, the storage unit of the underwater target detection device 34, etc. shown in FIG.

In an active sonar device, which is a type of sonar transmitter / receiver 31, an active type sonar device generates sound waves for an underwater target, analyzes the intensity of the reflected sound waves based on the response time, and determines the direction and distance of the target. presume. The observation data, that is, sonar data, is converted into B-scope format image data, and a system is constructed in which the target is detected by the shape of the reflected wave of the target. When displaying an image in B-scope format, resolution conversion and enhancement processing for display are performed for projection on the display device. The sonar data 100 before this conversion is particularly called raw data. Raw data is mainly used when learning and inferring with AI. Data whose resolution has been converted may be used depending on the type of sonar data and the processing of AI.

The sonar data 100, the observation condition 101, and the annotation data 102 are input via the system interface 106 of the system, and learning by machine learning is executed by the learning means 121. As described above, the sonar data 100 is an image data, that is, a data format in which two-dimensional array data having a signal strength as a value is input in time series. The observation condition 101 includes position information of the sea area where the observation is being made, the depth of the target object, the sea quality (sand, mud, etc.), the water temperature, the speed of the measuring ship, the attitude of the sonar transmitter / receiver 31, and the like.

The annotation data 102 is teacher data for learning, and is information on a preset type, position, and posture of a target object. The annotation data 102 includes data obtained by converting latitude / longitude and depth information on the GPS system regarding the target into position information on two-dimensional image data, and sonar data displayed on the B scope by a skilled water surveyor. The position and posture of the underwater target are calculated for each frame and the results are used.

The machine learning model used in this system adopts a model format suitable for detecting objects in image data. An AI method that detects an object of any size with a fixed shape from an image has been released, and using one of them, it is possible to estimate the position of multiple objects and the image range of the object. In this case, a plurality of objects having different shapes are recognized as independent classes. A learning model that supports multiple classes is called multi-class support.

As the input / output of the other system interface 106 described above, there are model data 103, performance data 104, and target position inference result 105. The model data 103 indicates a learning model (learned data) generated by this system or a learning model itself stored in an external system. The performance data 104 is data related to the performance of the learning model held by this system, such as the detection accuracy of the target object, and is input / output to / from an external system (not shown). The target position inference result 105 is the inference result data of the target object for which a learning model is set in this system, new sonar data 100 is input, and inference is performed.

The block arranged on the right side of the system interface 106 in FIG. 1 is a functional block that executes internal processing of this system. The outline of the internal processing will be described below.

The model reading / writing means 107 is a functional block that executes input / output of the learning model, and inputs / outputs the learning model generated by the learning means 121 and the learning model included in the learning model aggregate 114 based on the observation condition 101. conduct.

Next, a functional block for executing step S2 in FIG. 15 will be described. The distance calculation means 109 calculates the distance from each observation condition included in the learning model aggregate 114 with respect to the observation condition 101 input via the system interface 106. As a result, the distance calculation means 109 specifies the distance of each learning model in which the calculated distance is stored in the learning model aggregate 114. Here, the distance calculation means 109 calculates the conditional distance for each learning model stored in the learning model aggregate 114, and stores the result as the conditional map 115.

The distance calculation means 109 further stores the input condition data 116 and reads out the input condition data 116 as necessary. In order to calculate the conditional distance, two conditional data, the input observation condition 101 and the observation condition of the learning model to be processed, are required. Here, FIG. 8 shows the condition data 116.

Next, a functional block for executing step S3 in FIG. 15 will be described. The model selection means 108 is a functional block that selects one or more learning models with suitable conditions from the learning model aggregate 114. The distance between the two conditions calculated by the distance calculation means 109, that is, the conditional distance is used for this condition determination. This conditional distance includes the distance of each of the above-mentioned learning models.

The evaluation functions 125 and 126 are functions for calculating one scalar value from a multidimensional score. In the case of detecting that the target is one of multiple classes, this evaluation function can use a function to extract the maximum value of the multidimensional score. In that case, the class corresponding to the maximum score will be estimated as the target class.

The performance analysis means 127 compares the inference result of the target object selected by the evaluation function 126 with the annotation data which is the teacher data, evaluates the degree of matching over the entire group of inference data, and calculates the inference accuracy. Is.

The inference performance data 120 is data used for selecting a learning model by the model selection means 108, which is data indicating the above inference accuracy. The inference performance data 120 is shown in FIG. The model selection means 108 considers the inference performance indicated by the inference performance data 120, and excludes the inference performance from the learning model to be selected when the inference performance does not satisfy the predetermined inference performance threshold value.

Next, step S4 in FIG. 15 will be described. The data dividing means 110 executes a process of dividing a data set of sonar data and annotation data that are inputs of the learning process means 111 and the inference process means 113. The time-series sonar data 100 is distributed to learning or inference in units of image frames for each time, and is divided into each data set.

The learning process means 111 operates the learning means 121 with respect to the sonar data set for learning to control a series of processes for generating a learning model. The generated learning model is stored in the learning model aggregate 114 based on the observation conditions and utilized.

The learning means 121 uses annotation data corresponding to the input observation data. The learning means 121 calculates the loss function using the annotation data 102 in its internal processing, and optimizes the learning model.

Further, the preprocessing means 112 executes preprocessing on the data to be handled when performing various processing such as learning. Preprocessing includes data shaping and the like.

Next, a functional block for executing step S5 in FIG. 15 will be described. The inference process means 113 operates the inference means 122 with respect to the above-mentioned sonar data set for inference to control a series of processes for calculating the inference result.

The inference means 122 outputs data about the candidate of the target. Specifically, it outputs a score that quantifies the positions of a finite number of target candidates and their certainty corresponding to the inference setting threshold value. If there are multiple targets, the multidimensional score corresponding to multiple classes is output. This is the end of the description of the functional block shown in FIG. 1, and the details of various information will be further described below with the processing contents. In particular, the details of the distance calculation in step S2 are also shown below.

First, the correspondence between the observation condition 101 and the sonar data 100 will be described with reference to FIG. Here, as two types of conditions, the identification information A1 to A4 of the sea area and the depths D1 to D4 will be described as an example. Reference numeral 201 is a type of sea area, and reference numeral 202 is for distinguishing the difference in depth, and the names of sonar data corresponding to these two conditions (reference numeral 203, etc.) are shown in correspondence with each other. In FIG. 2, sonar data corresponding to all conditions have been observed and are shown to be available. However, not all conditions have been acquired at the actual site, and some data (or missing data) has not been acquired (specifically, missing data is identified and exception handling is executed. ).

FIG. 3 shows four types of observation conditions: sea area information, depth, type of target, and size of target. The sea area information of a is composed of latitude, longitude, and sediment. The depth of b uses the depth of the target (here, it is assumed that depth1 etc. is replaced with a concrete numerical value in meters). The type of target object in c is data storing a name or an identification code for distinguishing them. The size of d is stored as a numerical value in centimeters for the size of the target.

The details of the sea area information in FIG. 2 correspond to the observation condition a in FIG. 3, and the depth in FIG. 2 corresponds to the details b in the depth in FIG.

Here, the information regarding the distance calculation in step S2 and the details of the distance calculation will be described with reference to FIGS. 4, 5 and 9. FIG. 4 shows an example of a method for converting the above four types of observation conditions into distances, that is, a method for converting conditional distances. Each condition is associated with a scalar quantity (distance). In the conversion table to scalar quantities X1, X2, X3, X4, each of a, b, c, d corresponds to the sea area (A), depth (D), target type (O), and target size (S). be. If the types of each condition are regarded as independent dimensions, a multidimensional space with each of X1 to X4 shown in FIG. 4 as a coordinate axis can be assumed. FIG. 4 can be regarded as a method of calculating the distance along each coordinate axis. It is necessary to separately define the method for measuring the distance in the above-mentioned multidimensional space.

FIG. 9 is an example showing the result of the conditional distance calculated by using the conversion method of the conditional distance of FIG. 4, that is, the evaluation value. The numerical values of the conditional distances shown here are values based on the sea area A1 and the depth D1. The numerical range of the conditional distance is 0 to 330. It is not always necessary to calculate the conditional distance using all the conditions input in the observation condition 101. Any method may be used to calculate the conditional distance corresponding to the application of the learning model.

Here, an example of the calculation method of the conditional distance used in the calculation of FIG. 9 is shown in Equation 1.
Conditional distance (d1, d2) = k1 * ｜ X1 (d1) -X1 (d2) ｜ ＋ k2 * ｜ X2 (d1) --X2 (d2) ｜ (However, k1 = 1, k2 = 1) ・ ・ ・ ( Number 1)
For the two observation points d1 and d2, the conditional distance (d1, d2) is a calculation formula that depends only on the two conditions, the distance X1 between the sea area conditions, and the distance X2 between the depths. Here, k1 and k2 are the weighting coefficients of each condition type (dimension).

Observation data and a learning model that is the learning result can be associated with a multidimensional space using the dimensions of each of the above conditions, and these are called condition maps.

FIG. 5 is an example of a condition map, and observation data is associated with each observation point. It is plotted at the position corresponding to the scalar quantity X1 and X2 of each dimension. If there is a condition that has not been observed, that point will be a defect.

Next, the selection conditions in the model selection in step S3 will be described. The model selection means 108 selects one or more learning models based on the calculated distance according to the selection conditions stored in advance. This selection condition includes, for example, "neighborhood" and "whole".

First, the neighborhood will be described. When using a neighborhood, the model selection means 108 selects the learning model with the shortest calculated distance. Further, in this case, as a neighborhood condition, the model selection means 108 may select a plurality of learning models in ascending order of distance. Further, as a neighborhood condition, the model selection means 108 may be configured to select one or more learning models within a predetermined distance. At this time, if there is no learning model within a predetermined distance, it is desirable that the model selection means 108 selects one or more learning models in ascending order of distance. This makes it possible to select a learning model that pinpoints the observation conditions 101 observed at that time.

Next, the whole will be described. Here, the model selection means 108 calculates a representative value of the distance (nearest neighbor distance) of each learning model of the learning model assembly 114 to the nearest model, and uses this. For this purpose, the model selection means 108 selects so that the representative value of the nearest neighbor distance of the selected learning model is equal to the entire learning model of the learning model aggregate 114. Here, "equivalent" includes the same value and within a predetermined difference. In addition, the representative value includes an average value, a median value, and the like. By using all of the above, it becomes possible to select a more general-purpose learning model.

As described above, the selection condition may be any condition as long as it is a condition for selecting one or more, that is, a single or a plurality of learning models.

Therefore, the model selection means 108 can also use a plurality of selection conditions in combination. For example, the model selection means 108 selects a learning model having the shortest distance and a plurality of learning models having the same representative value. This makes it possible to select a learning model that is well-balanced between the neighborhood and the whole.

Further, in each of the above-mentioned selection conditions, an upper limit of the number of learning models to be selected may be set, or the number may be set to a fixed number.

The fact that the three output multidimensional scores 123 are output from the learning means 121 in FIG. 1 indicates that learning is being performed for the three learning models. That is, it indicates that the model selection means 108 has selected three learning models. Further, the fact that the inference means 122 outputs the three multidimensional scores 124 also indicates that the model selection means 108 has selected the three learning models.

Next, information regarding the learning process in step S4 will be described.

FIG. 6 is a plot of the learning model on the condition map 115. That is, it is the result of learning by the learning process means 111 and the learning means 121. In this figure, the learning model is plotted at the position of the condition where the observation is not performed. They are learning models 504,505,506. These are the results of training using the union of two or more training data. The learning model 504 is the result of using the union of the sea area A1 and the sea area A2 at the depth D1, and the same applies to the learning models 505 and 506.

FIG. 7 is a diagram showing the contents of the learning model aggregate. Here, the names of the learning models corresponding to each condition are described for the dimensions of the two conditions. The learning model from the union of the observation data shown in the explanation of FIG. 6 is also added to the data contents of the learning model aggregate. As shown in this example, the learning model set can be a mechanism for expanding the data contents by operation. Regarding the name of the learning model, as an example, M1111 is a learning model corresponding to the sea area A1 and the depth D1, and as another example, M12-1-1-1 is the union of the sea areas A1 and A2, and the condition of the depth D1. The corresponding learning model.

Finally, the estimation process of step S5 in FIG. 15 will be described. A process in which a new observation condition 101 and its sonar data 100 are input to the system of this embodiment and inference is performed using the generated learning model aggregate will be described.

The sonar data 100 and the observation condition 101 are input to the system interface 106, and the model selection means 108 reads out one or more learning models optimal for the input observation conditions from the training model aggregate 114 and temporarily inputs them to the multiclass learning model 119. Hold on. The inference process means 113 operates the inference means 122 for each frame of the input sonar data, outputs a multidimensional score 124, and calculates the inference result by the evaluation function 126. The calculated inference result is output to the target position inference result 105 by the inference process means 113 via the system interface 106.

An example of the configuration of the multi-class learning model 119 is shown in FIG. In this example, three learning models M1, M2, and M3 are used. In the above process, the inference means 122 makes inferences using the three learning models included in this multi-class learning model, and outputs the score values corresponding to each model as a multidimensional score. The process of the evaluation function 126 uses the above-mentioned process of calculating the maximum value. As a result, one of the three selected learning models becomes a candidate for the target, and the estimated position and the certainty of the target are calculated.

As described above, according to the first embodiment, a learning model is generated for each observation condition 101 of the sonar data 100, a set of learning models is constructed by them, and the newly input sonar data is optimized according to the observation conditions. You can select one or more learning models. Therefore, even for the sonar data 100 of the underwater target whose learning model depends on the observation condition 101, the underwater target can be detected with high accuracy according to the observation condition of the input point.

Next, Example 2 of the present invention will be described. In this embodiment, the underwater target detection device 34'installed on the ship 30 executes model selection. FIG. 16 shows a system configuration including the underwater target detection device 34'. The underwater target detection device 34'is connected to the sonar transmission / reception device 31, the sensors 32, and the wireless device 33, similarly to the underwater target detection learning device 10 of the first embodiment. Further, although not shown in FIG. 16, information can be exchanged with the underwater target detection learning device 10 via the network 40 and the storage medium 60, as in the first embodiment.

Further, the underwater target detection device 34'is connected to a terminal device 36 or a database 35 capable of inputting / outputting information to the underwater target detection device 34'. The terminal device 36 and the database 35 may be integrated with the underwater target detection device 34'.

Next, the details of the underwater target detection device 34'will be described. The underwater target detection device 34'is realized by a computer in the same manner as the underwater target detection learning device 10 and the underwater target detection device 34 of the first embodiment. That is, the underwater target detection device 34'has a processing unit 341, a storage unit 342, and an input / output unit 343. Therefore, each function of the underwater target detection device 34'is realized by the processing unit 341 performing an operation according to a program developed in the storage unit 342.

Here, a program that realizes each function of the underwater target detection device 34'will be described. The underwater target detection device 34'has the following configurations in addition to the inference process unit 1130 and the inference unit 1220 included in the underwater target object detection device 34 of the first embodiment.

First, the underwater target detection device 34'has a model selection unit 1080, a distance calculation unit 1090, a data division unit 1100, and a preprocessing unit 1120. These are also possessed by the learning device 10 for detecting an underwater target according to the first embodiment. In addition to these, the underwater target detection device 34'has a route processing unit 14021. Other than the route processing unit 14021, the same functions as in the first embodiment are exhibited. The function of the route processing unit 14021 will be described later with reference to FIGS. 12 and 13.

Next, various information stored in the database 35 will be described. The database 35 has annotation data 102, performance data 104, target position inference result 105, learning model aggregate 114, condition map 115, and condition data 116. These are the same as the information stored in the database 50 of the first embodiment. In addition to this, in the second embodiment, the database 35 stores the root program 1403.

Here, the functional configuration of the second embodiment will be described with reference to FIG. The functional configuration of FIG. 12 realizes the function of the underwater target detection device 34'shown in FIG. Here, FIG. 12 has the following differences as compared with FIG. 1. A route processing means 1402 and a route program 1403 have been added. Here, the route processing means 1402 corresponds to the route processing unit 14021 of FIG. Further, in FIG. 12, the learning process means 111, the duplicate data 117, the raw data 118, the inference performance data 120, the learning means 121, the output multidimensional score 123, the score evaluation function 125, and the performance analysis means 127 are omitted from FIG. ing.

Next, the processing of this embodiment will be described with reference to the flowchart of FIG. Due to the difference in the configuration described above, in this embodiment, the same processing as in steps S1, S2 and S5 of FIG. 15 is executed. Here, the description of these steps will be omitted, and steps S3-1 and S3-2, which are different from those of the first embodiment, will be described.

Further, in FIG. 17, (a) steps S1 to S3-1 show preparatory processing, and (b) steps S3-2 and step S5 show processing at the time of navigation. Therefore, regarding (a), the learning device 10 for detecting an underwater target may be executed.

As in the case of the first embodiment, the processing subject will be described using the functional block shown in FIG.

First, in step S3-1, the route processing means 1402 creates a route program 1403 that associates the navigation route, the observation conditions, and the learning model. To this end, the model selection means 108 extracts a learning model based on the distance calculated in step S2. This extraction performs the same processing as the model selection in step S3 of the first embodiment.

However, it is input information such as the route processing means 1402 and the observation conditions according to the planned navigation route. Then, the route processing means 1402 extracts each learning model for each navigation route in the order of unit time.

Then, the route processing means 1402 creates a route program 1403 in which the extracted learning model, the observation conditions corresponding to the extracted learning model, and the travel route are associated with each other, and stores the route program 1403 in the database 35. An example of this route program 1403 is shown in FIG. In FIG. 13, the route indicating the navigation route (underwater route information), the sea area (A) and depth (D) indicating the observation conditions, the type (O) of the underwater target, and the model indicating the learning model are associated with each other. There is. In FIG. 13, one observation condition and one learning model are associated with each navigation route, but a plurality of observation conditions and a learning model may be associated with each other.

Further, the route processing means 1402 sorts and records the route program 1403 in chronological order in the order of the scheduled unit time of the navigation route. Further, the route processing means 1402 may record the scheduled navigation time in the route program 1403.

Next, at the time of navigation, in step S3-2, the model selection means 108 appropriately selects a learning model corresponding to the route or the navigation time from the route program 1403. That is, the model selection means 108 selects a learning model according to the planned navigation route from the route program 1403 in the order sorted in chronological order.

Then, at the time of navigation, the inference process in step S5 is executed using the selected learning model.

The process of the second embodiment may also exhibit its function in cooperation with the learning device 10 for detecting an underwater target, as in the first embodiment.

According to the second embodiment, the learning corresponding to the observation conditions of each navigating point is held by holding the necessary learning model appropriately selected according to the time series of the observation conditions corresponding to the preset navigation route. Read the model and estimate the target. Therefore, in this embodiment, it is possible to detect an underwater target object with high accuracy according to the navigation route.

In each of the above embodiments, the system with learning can prevent the operation cost from increasing due to the maintenance of the learning data and the learning model.

Further, in each embodiment, although the target is a ship, it can also be applied to a moving body or a means for detecting an underwater target such as an aircraft.

100 ... sonar data, 101 ... observation condition input, 102 ... annotation data, 105 ... target position inference result, 114 ... learning model aggregate, 108 ... model selection means, 109 ... distance calculation means, 115 ... condition map, 121 ... learning Means, 122 ... Inference means, 119 ... Multi-class learning model

Claims (11)

In a learning device for underwater targets for detecting underwater targets using observation data
A storage unit that stores a plurality of learning data used for detecting the underwater target in association with observation conditions for the underwater target, respectively.
An input unit that accepts the observation data of the underwater target and the observation conditions of the observation data,
A distance calculation unit that calculates the distance between the accepted observation conditions and the plurality of observation conditions stored in the storage unit, and the distance calculation unit.
A model selection unit that selects a learning model from the storage unit based on the distance calculated by the distance calculation unit.
It has a learning unit that learns about the detection of the underwater target using the learning model selected by the model selection unit.
An underwater target detection learning device capable of making inferences regarding detection of the underwater target received by the input unit using the learning model learned by the learning unit. In the learning device for detecting an underwater target according to claim 1,
The model selection unit is a learning device for detecting an underwater target that selects a learning model associated with an observation condition having the shortest distance calculated by the distance calculation unit. In the learning device for detecting an underwater target according to claim 1,
The model selection unit is a learning device for detecting an underwater target that selects a plurality of learning models based on the distance calculated by the distance calculation unit. In the learning device for detecting an underwater target according to claim 3,
The model selection unit
The first representative value of the distances of the plurality of observation conditions stored in the storage unit is calculated.
A plurality of observation conditions that have a second representative value equivalent to the first representative value are specified, and a plurality of observation conditions are specified.
A learning device for detecting an underwater target that selects a learning model associated with a plurality of specified observation conditions. In the learning device for detecting an underwater target according to any one of claims 3 or 4.
The model selection unit is a learning device for detecting an underwater target that further selects a learning model associated with the observation condition having the shortest distance calculated by the distance calculation unit. In an underwater target detection device that detects underwater targets or underwater targets related to ship navigation using observation data.
A storage unit that stores a plurality of learning data used for detecting the underwater target in association with the observation conditions for the underwater navigation object or the underwater target, respectively.
An input unit that accepts the observation data of the underwater target and the observation conditions of the observation data,
A distance calculation unit that calculates the distance between the accepted observation conditions and the plurality of observation conditions stored in the storage unit, and the distance calculation unit.
A model selection unit that selects a learning model from the storage unit based on the distance calculated by the distance calculation unit and the navigation route that the ship plans to navigate.
An underwater target detection device having an inference unit that makes inferences regarding detection of the underwater target received by the input unit based on the learning model selected by the model selection unit. In the underwater target detection device according to claim 6,
Further, a learning model is extracted based on the distance calculated by the distance calculation unit, and a route program in which the extracted learning model is associated with the observation conditions corresponding to the learning model and the navigation route is created. The storage unit has a route processing means for storing the route program.
The model selection unit is an underwater target detection device that selects the learning model from the route program. In the underwater target detection device according to claim 7,
The route processing means is an underwater target detection device that extracts a learning model associated with the observation condition with the shortest distance calculated by the distance calculation unit. In the underwater target detection device according to claim 7,
The route processing means is an underwater target detection device that extracts a plurality of learning models based on the distance calculated by the distance calculation unit. In the underwater target detection device according to claim 9,
The route processing means is
The first representative value of the distances of the plurality of observation conditions stored in the storage unit is calculated.
A plurality of observation conditions that have a second representative value equivalent to the first representative value are specified, and a plurality of observation conditions are specified.
An underwater target detection device that extracts a learning model associated with a plurality of specified observation conditions. In the underwater target detection device according to any one of claims 9 or 10.
The route processing means is an underwater target detection device that further extracts a learning model associated with the observation condition with the shortest distance calculated by the distance calculation unit.

Images