JP2016192007A - Machine learning device, machine learning method, and machine learning program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To increase a detection accuracy.SOLUTION: A determination unit determines the presence or the absence of a target in a predetermined region of image data by applying a preset rule to feature information showing features extracted from the predetermined region. The acquisition unit acquires user information that a user has input by seeing an image shown in the image data. An information generation unit generates information combining the feature information and user information acquired by the acquisition unit. A model generation unit generates a learning model by a machine learning system on the basis of the information generated by the information generation unit. An update unit updates a rule in accordance with the learning model generated by the model generation unit.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a machine learning device, a machine learning method, and a machine learning program for discriminating a desired object from an image.

  When detecting a desired object from image data, a technique is known in which the object is detected according to a rule set by a user and the detection result of the object is machine-learned to improve detection accuracy.

  As such a technique, for example, in Patent Document 1, machine learning case data is collected based on image data under a user environment, and an identification function is generated to classify the image data. A technique capable of providing an appropriate classification suitable for the environment and learning the classification method is disclosed.

  Patent Document 2 discloses a method for performing highly accurate machine learning using appropriate learning data as evaluation data and selecting learning data from the learning data and the evaluation data every time data to be evaluated is given. A technique for extracting a second feature and selecting learning data suitable for use in machine learning based on the second feature is disclosed.

  Further, in Patent Document 3, the case database stores information on each case, and the learning device generates an inference rule for determining the category from the set of correct answer cases, and infers the category of the unknown case. The query generation unit generates a query for determining whether or not the inference result of the unknown case is correct, the user interface displays the query on the screen, and when the user inputs the determination result, the central controller A technique for recording a determined unknown case in a case database as a new correct answer case is disclosed.

JP 2010-92413 A JP 2005-182696 A JP 2002-222083 A

However, in the techniques disclosed in Patent Documents 1 to 3, when the rules set by the user are not suitable for detecting an object, the detection accuracy may not be improved even if machine learning is performed and suitable image data is selected. is there.
An object of the present invention is to provide a machine learning device, a machine learning method, and a machine learning program that solve the above-described problems.

  The present invention has been made to solve the above-described problem, and applies a predetermined rule to feature information indicating features extracted from a predetermined area of image data, and an object within the predetermined area. Is acquired by the determination unit that determines whether or not exists, the acquisition unit that acquires the user information that is input by allowing the user to visually recognize the image indicated by the image data, the feature information, and the acquisition unit An information generation unit that generates information that is combined with the user information, a model generation unit that generates a learning model by a machine learning method based on the information generated by the information generation unit, and the model generation unit An update unit that updates the rule according to the learning model.

  The present invention also applies a predetermined rule to feature information indicating features extracted from a predetermined area of image data to determine whether an object exists in the predetermined area, The user information selected by causing the user to visually recognize the image indicated by the image data is acquired, information that combines the feature information and the acquired user information is generated, and based on the generated information A machine learning method that generates a learning model by a machine learning method and updates the rule according to the generated learning model.

  Further, the present invention applies a predetermined rule to feature information indicating features extracted from a predetermined area of image data, and determines whether or not an object exists in the predetermined area. The user information selected by allowing the user to make a determination and viewing the image indicated by the image data, generating information combining the feature information and the acquired user information, and generating the generated information. Is a machine learning program for generating a learning model by a machine learning method and updating the rule according to the generated learning model.

  According to the present invention, detection accuracy can be improved.

It is a figure which shows an example of the basic composition of the machine learning apparatus 1 in this invention. It is a figure showing an example of composition of machine learning device 1 in one embodiment. It is a figure which shows an example of the image data which image | photographed the ground airfield from the sky. It is a figure which shows an example of rule information. It is a figure which shows an example of the detection result by the rule detection execution part 220. FIG. It is a figure which shows an example of the user's detection judgment result with respect to the detection result of the rule detection execution part 220 shown in FIG. It is a figure which shows an example of teacher data. It is a figure which shows an example of the classification result by the model classification implementation part 320. FIG. 6 is a diagram for explaining the concept of data string classification processing in a model classification execution unit 320. FIG. 6 is a diagram for explaining the concept of data string classification processing in a model classification execution unit 320. FIG. It is a flowchart which shows an example of a process of the machine learning apparatus 1 in one Embodiment.

  Hereinafter, a machine learning device, a machine learning method, and a machine learning program according to an embodiment of the present invention will be described with reference to the drawings.

(Basic configuration)
FIG. 1 is a diagram illustrating an example of a basic configuration of a machine learning device 1 according to the present invention.
The machine learning device 1 according to the present invention has a basic configuration of a determination unit 10, an acquisition unit 20, an information generation unit 30, a model generation unit 40, and an update unit 50.

  The determination unit 10 applies a rule set by the user to the feature information indicating the feature extracted from the detection area of the image data, and determines whether or not the object OB exists in the detection area. The determination unit 10 corresponds to a “rule detection execution unit 220” in an embodiment described later.

  The acquisition unit 20 acquires the detection detection result of the user described above. The acquisition unit 20 corresponds to a “teacher data generation unit 311” in an embodiment described later.

  The information generation unit 30 generates teacher data that combines the user detection determination result acquired by the acquisition unit 20 and the feature information indicating the feature extracted from the detection area A on the image data. The information generation unit 30 corresponds to a “teacher data generation unit 311” in an embodiment described later.

  The model generation unit 40 generates a learning model by a machine learning method based on the teacher data generated by the information generation unit 30. The model generation unit 40 corresponds to a “learning model generation unit 313” in an embodiment described later.

  The update unit 50 updates rules set by the user according to the learning model generated by the model generation unit 40. The update unit 50 corresponds to a “rule verification execution unit 233” in an embodiment described later.

  Thereby, the machine learning device 1 can update the rule for detection of the object OB scattered in the entire image data to a more appropriate rule based on the result of the machine learning. As a result, the machine learning device 1 can improve detection accuracy.

FIG. 2 is a diagram illustrating an example of the configuration of the machine learning device 1 according to an embodiment.
The machine learning device 1 in the present embodiment is a device that detects a desired object OB from image data. For example, the machine learning device 1 detects an aircraft as a desired object OB from image data obtained by photographing an airfield on the ground from above using an artificial satellite or an aircraft. FIG. 3 is a diagram illustrating an example of image data obtained by photographing a ground airfield from above. In the case of the example in FIG. 3, there are a plurality of aircraft that are objects to be detected OB on the image data. The operation of the machine learning device 1 will be described below with reference to the image data shown in FIG. 3 as an example.

  The machine learning device 1 includes an image display unit 100, a rule control unit 200, a machine learning unit 300, and a data management unit 400. Some or all of the image display unit 100, the rule control unit 200, the machine learning unit 300, and the data management unit 400 are programs stored in a storage device (not shown) such as a CPU (Central Processing Unit). This is a software function unit that functions by executing. Also, some or all of the functional units described above may be hardware functional units such as LSI (Large Scale Integration) and ASIC (Application Specific Integrated Circuit).

  The image display unit 100 displays an image based on the input information. The image display unit 100 is a liquid crystal display, an organic EL (Electroluminescence) display, or the like.

  The image display unit 100 includes an image display control unit 110. The image display control unit 110 controls the image display unit 100 to display an image. For example, the image display control unit 110 causes the image display unit 100 to display an execution result by the rule detection execution unit 220 described later and information for generating teacher data.

  The rule control unit 200 is a functional unit for detecting an aircraft from image data according to a rule set by a user. The rule control unit 200 includes a rule information management unit 210, a rule detection execution unit 220, and a rule verification unit 230.

  The rule information management unit 210 outputs information (hereinafter referred to as “rule information”) indicating a rule input by the user to the data management unit 400 and causes the data management unit 400 to manage the rule information. The rule information is a constraint condition that is referred to when detecting an aircraft, and is expressed in a program format, for example. For example, the rule control unit 200 may accept rule information input from a user via a user interface (not shown), or may acquire rule information from a storage device connected via a hardware interface.

  FIG. 4 is a diagram illustrating an example of rule information. In the case of the example in FIG. 4, a rule is set in the rule information for each feature that is scheduled to be extracted by a rule detection execution unit 220 described later. A feature is a feature related to a shape such as length, width, area, and the like. For example, the rule number r001 is assigned a length feature, and when the length is 300 or more and 400 or less, the content that it may be determined that the object OB exists in the detection area A Is stored. The rule number is identification information that can identify each rule. Further, the rule number r002 stores a content indicating that the object OB can be determined to exist in the detection area A when the width feature is allocated and the width is 280 or more and 370 or less. Has been. In the example of FIG. 4, the number of rules is 15, but there is no limit to the number of rules. Therefore, the number of rules constituting the rule information may be adjusted as appropriate according to the number of features to be allocated. Also, the content of the rules may be changed as appropriate by the user. In such a case, the rule control unit 200 receives information indicating a rule change content input from the user via a user interface such as a mouse or a keyboard.

  The rule detection execution unit 220 detects an object OB (aircraft) from the image data according to the rules stored in the data management unit 400 with respect to the image data acquired by the representative data acquisition unit 231 described later.

  Specifically, the rule detection execution unit 220 first performs feature extraction processing on the entire range (or a partial range) of the image data. Next, the rule detection execution unit 220 refers to all the rules with the rule numbers r001 to r015 included in the rule information, determines whether or not the constraint conditions indicated in the rule contents are satisfied for each extracted feature, A detection result indicating that a region satisfying the constraint condition (rule content) is a detection region A in which the object OB exists is output. The rule detection execution unit 220 displays, for example, image information that displays only the detection region A in which the object OB exists as a detection result indicating that the detection region A in which the object OB exists. Output to 100.

  FIG. 5 is a diagram illustrating an example of a detection result by the rule detection execution unit 220. In the case of the example in FIG. 5, two areas where the object OB does not exist are included in the detection area A. That is, the rule detection execution unit 220 may output a detection result including erroneous detection. As an example of the cause of erroneous detection, the constraint condition indicated in the rule content is not suitable as an index for detecting the presence of the object OB.

  The rule verification unit 230 is a functional unit that verifies whether the rule set by the user is valid. The rule being valid means that the object OB can be detected from the image data with high accuracy. The rule verification unit 230 includes a representative data acquisition unit 231, a survey data generation unit 232, and a rule verification execution unit 233.

  The representative data acquisition unit 231 randomly selects representative image data from a plurality of image data managed by the data management unit 400. Hereinafter, the image data acquired by the representative data acquisition unit 231 is referred to as “representative data”.

  When the classification processing is performed by the model classification execution unit 320, which will be described later, the survey data generation unit 232 is data in which a desired object OB exists in the detection area A, or a desired target in the detection area A. A data string that is not classified into data in which no object OB exists is extracted from the n-dimensional space. That is, the survey data generation unit 232 extracts from the n-dimensional space a data string whose reliability is the same as a predetermined value or a data string whose reliability is within a predetermined range. The survey data generation unit 232 generates survey data that combines the extracted data strings. The survey data generation unit 232 is an example of an “extraction unit”.

  The rule verification execution unit 233 compares the classification result obtained by applying the learning model by the model classification execution unit 320, which will be described later, with the detection result obtained by applying the feature information of the data string constituting the survey data to the rule set by the user. Then, it is determined whether the classification result applying the learning model matches the detection result applying the rule. When the classification result to which the learning model is applied and the detection result to which the rule is applied do not match, the rule verification execution unit 233 changes the rule so that the detection result to which the rule is applied is brought closer to the classification result to which the learning model is applied. To do. For example, the rule verification execution unit 233 increases the threshold value set for each feature and relaxes the constraint condition (rule content) for each rule. The classification result to which the learning model is applied is an example of “model determination result”, and the detection result applied to the rule set by the user is an example of “rule determination result”.

  The machine learning unit 300 includes a teacher data management unit 310 and a model classification execution unit 320. The teacher data management unit 310 includes a teacher data generation unit 311, a feature information generation unit 312, and a learning model generation unit 313.

  The teacher data generation unit 311 presents the detection result from the rule detection execution unit 220 to the user, and causes the user to determine whether or not the detection result is an error. Specifically, the teacher data generation unit 311 causes the rule detection execution unit 220 described above to output a detection result (eg, FIG. 5) to the image display unit 100. As a result, the image display unit 100 displays the detection result by the rule detection execution unit 220. At this time, the user looks at the detection result of the rule detection execution unit 220 displayed by the image display unit 100 and determines whether or not the detection result is an error.

  For example, when the user determines that a part or all of the detection result of the rule detection execution unit 220 is an error, the user inputs information indicating an error to the machine learning unit 300 via the user interface. In this case, the teacher data generation unit 311 causes the rule detection execution unit 220 to output the detection result to the image display unit 100 and then receives information input from the user. Hereinafter, information input from the user after the detection result of the rule detection execution unit 220 is output to the image display unit 100 will be referred to as “user detection determination result”. Note that the user's detection determination result is an example of “user information”.

  FIG. 6 is a diagram illustrating an example of a user detection determination result with respect to the detection result of the rule detection execution unit 220 illustrated in FIG. In the example of FIG. 6, the user performs an operation of designating the detection area A determined to be erroneous detection, and deletes the erroneous detection area A from the detection result of the rule detection execution unit 220. At this time, on the detection result of the rule detection execution unit 220, the teacher data generation unit 311 assigns an erroneous identifier (label) to the detection area A that has been designated for erroneous detection by the user. The detection area A that has been designated for erroneous detection is indicated by a broken line in the figure, and the detection area A that has not been designated for erroneous detection is indicated by a solid line in the figure.

  The teacher data generation unit 311 generates teacher data including a user detection determination result acquired by the above-described method and feature information indicating features extracted from the detection area A on the image data.

  FIG. 7 is a diagram illustrating an example of teacher data. Each data string constituting the teacher data stores a feature amount when a rule is applied for each feature and information indicating a user's detection determination result. Each data string is assigned a data number for each detection region A where the rule detection execution unit 220 determines that the object OB exists. That is, the data number corresponds to the number of detection areas A where the rule detection execution unit 220 determines that the object OB exists.

  For example, in the data string indicated by the data number d000001, a feature quantity 390 referred to when the constraint condition of the rule number r001 is applied, and a feature quantity 345 referred to when the constraint condition of the rule number r002 is applied. The feature quantity 55 referred to when the constraint condition of the rule number r015 is applied is omitted. In this data string, when an identifier (label) indicating erroneous detection is given to the detection area A corresponding to the data number d000001, information of “NO” is stored in the data string as a result of the user's detection determination. The If an identifier (label) indicating erroneous detection is not assigned to the detection area A, the data string indicated by the data number corresponding to the detection area A stores “YES” information as a user detection determination result. Is done.

  The feature information generation unit 312 extracts features from the detection area A in which the object OB is determined to be present by the rule detection execution unit 220 using, for example, an image processing technique such as secondary differential processing or Hough transform. Then, feature information indicating the feature is generated. In the present embodiment, the feature information generation unit 312 is not limited to the one that extracts features related to the shape described above. For example, the feature information generation unit 312 may extract features related to feature points and colors. Further, the feature information generation unit 312 may extract features from the detection area A designated by the user and generate feature information indicating the features.

  The learning model generation unit 313 generates a learning model for the teacher data generated by the teacher data generation unit 311 using a machine learning method such as a support vector machine (SVM). The learning model is, for example, a model (program) for causing the machine learning device 1 to automatically recognize a shape pattern of a desired object (aircraft in the present embodiment).

  The model classification execution unit 320 applies the learning model generated by the learning model generation unit 313, and calculates a reliability for each data string included in the teacher data generated by the teacher data generation unit 311. The reliability indicates the probability that the object OB exists in the detection area A. That is, the reliability represents the degree of conformity with the rule for determining whether or not the object OB exists in the detection area A, and the higher the degree of conformance with the rule, the higher the reliability value. The model classification execution unit 320 is an example of a “calculation unit”.

  The model classification execution unit 320 applies the learning model generated by the learning model generation unit 313, and sets a desired object OB in the detection area A for each data string constituting the survey data generated by the survey data generation unit 232. Is classified into two patterns, that is, a data string in which the object OB is determined to exist and a data string in which it is determined that the object OB does not exist on the survey data. To do. Hereinafter, the data string classification processing based on the reliability of the model classification execution unit 320 will be described in more detail with reference to FIGS.

  FIG. 8 is a diagram illustrating an example of a classification result obtained by the model classification execution unit 320. 9 and 10 are diagrams for explaining the concept of data string classification processing in the model classification execution unit 320. FIG.

  The model classification execution unit 320 plots the data strings one by one on the n-dimensional space where the number n of feature information (rule numbers) of the data strings constituting the survey data is the number of dimensions. In the case of the example in FIG. 9, the model classification execution unit 320 plots a data string on a two-dimensional space with two pieces of feature information, feature information A and B, as variables. At this time, as shown in FIG. 9, the model classification execution unit 320 classifies the two-dimensional space into two with the boundary line BD having the calculated reliability being a predetermined value (for example, 0.5) as a boundary.

  As shown in FIG. 10, the model classification execution unit 320 may classify the two-dimensional space into two by giving the boundary line BD a certain width. For example, the model classification execution unit 320 classifies the two-dimensional space into two with respect to an area indicating reliability within a predetermined range (for example, about 0.4 to 0.6). When the reliability is about 0.5, it indicates that it is difficult to determine whether the object OB exists in the detection area A. As a result, the machine learning device 1 is a data string that is clearly the result when the reliability is close to 0 or 1, and does not use such data and is difficult to determine in computer processing. On the other hand, it can be determined whether or not the object OB exists in the detection area A. As a result, the machine learning device 1 can incorporate feature information of a data string having a reliability of about 0.5, that is, a data string that is difficult to determine, into the teacher data. Accordingly, the machine learning device 1 generates a new learning model based on the teacher data including the feature information of the data string that is difficult to determine, and detects the object OB from the image data according to the newly generated learning model. The detection accuracy of the object OB can be increased.

  The model classification execution unit 320 determines a data string belonging to a space having a reliability higher than a predetermined value (predetermined range) among the classified spaces as data in which a desired object OB exists in the detection area A. Then, a data string belonging to a space whose reliability is smaller than a predetermined value (predetermined width) is determined as data in which a desired object OB does not exist in the detection area A.

  The model classification execution unit 320 outputs the classification result to the image display unit 100. For example, if the model classification execution unit 320 determines that a desired object exists in the detection area A, the model classification execution unit 320 outputs a classification result of “YES”. For example, when the model classification execution unit 320 determines that the desired object does not exist in the detection area A, the model classification execution unit 320 outputs a classification result of “NO”.

The data management unit 400 includes an image data management unit 410, a rule data management unit 420, and a machine learning data management unit 430.
The image data management unit 410 acquires image data from a storage device connected via a hardware interface, and stores the acquired image data in the image data storage unit 411. When the image data management unit 410 receives a request for acquiring image data from the rule control unit 200 or the machine learning unit 300, the image data management unit 410 outputs the image data stored in the image data storage unit 411 to the request source.

  The rule data management unit 420 causes the rule information storage unit 421 to store the rule information acquired by the rule information management unit 210. When the rule data management unit 420 receives a request for acquiring rule information from the rule control unit 200 or the machine learning unit 300, the rule data management unit 420 outputs the rule information stored in the rule information storage unit 421 to the request source.

  The machine learning data management unit 430 causes the teacher data storage unit 431 to store the teacher data generated by the teacher data generation unit 311. Further, the machine learning data management unit 430 stores the learning model generated by the learning model generation unit 313 in the model data storage unit 432. When the machine learning data management unit 430 receives a request for acquiring teacher data or a learning model from the rule control unit 200 or the machine learning unit 300, the machine learning data management unit 430 acquires the requested information from the storage unit corresponding to each information. The acquired information is output to the request source.

  The image data storage unit 411, the rule information storage unit 421, the teacher data storage unit 431, and the model data storage unit 432 described above are non-volatile such as ROM (Read Only Memory), flash memory, HDD (Hard Disk Drive), and the like. Storage media and volatile storage media such as RAM (Random Access Memory) and registers.

  FIG. 11 is a flowchart illustrating an example of processing of the machine learning device 1 according to an embodiment. The machine learning device 1 in the present embodiment performs the processing of this flowchart at a predetermined cycle.

  First, the representative data acquisition unit 231 randomly selects representative image data from a plurality of image data managed by the data management unit 400 (step S100). Next, the rule detection execution unit 220 detects the object OB (aircraft) from the image data according to the rules stored in the data management unit 400 for the image data acquired by the representative data acquisition unit 231 ( Step S102).

  Next, the feature information generation unit 312 extracts features from the detection area A where the rule detection execution unit 220 determines that the object OB exists, and generates feature information indicating the features (step S104). . Next, the teacher data generation unit 311 causes the rule detection execution unit 220 to output a detection result to the image display unit 100, and then receives a user detection determination result input from the user (step S106).

  Next, the teacher data generation unit 311 generates teacher data including the detection determination result of the user and feature information indicating the feature extracted from the detection area A on the image data (step S108). Next, the learning model generation unit 313 generates a learning model for the teacher data generated by the teacher data generation unit 311 using a machine learning method (step S110).

  Next, when the classification processing is performed by the model classification execution unit 320, the survey data generation unit 232 has data in which the desired object OB exists in the detection area A, or a desired data in the detection area A. A data string that is not classified into data in which the object OB does not exist is extracted from the n-dimensional space. In other words, the survey data generation unit 232 extracts a data string having a predetermined reliability from the teacher data referred to by the model classification execution unit 320 (step S112).

  Next, the survey data generation unit 232 generates survey data by combining the extracted data strings (step S114). Next, the model classification execution unit 320 applies the learning model generated by the learning model generation unit 313, and sets a desired value in the detection area A for each data string constituting the survey data generated by the survey data generation unit 232. It is determined whether or not the object OB exists, and on the survey data, a data string that is determined as the object OB is present and a data string that is determined as the object OB is not present 2 The patterns are classified (step S116).

  Next, the rule verification execution unit 233 uses the classification result obtained by applying the learning model by the model classification execution unit 320 and the detection result obtained by applying the feature information of the data string constituting the survey data to the rule set by the user. A comparison is made to determine whether or not the classification result to which the learning model is applied matches the detection result to which the rule is applied (step S118).

  When the classification result to which the learning model is applied and the detection result to which the rule is applied do not match (step S118; Yes), the rule verification execution unit 233 displays the detection result to which the rule is applied to the classification result to which the learning model is applied. The rule is changed so as to approach (step S120). When the classification result applying the learning model matches the detection result applying the rule (Step S118; No), the rule verification execution unit 233 holds the rule currently stored in the rule information storage unit 421 (Step S118). S122). Thereby, the machine learning device 1 ends the processing of this flowchart.

  As described above, according to the present embodiment, the machine learning device 1 applies the rule set by the user to the feature information indicating the feature extracted from the detection region of the image data, so that the target object is included in the detection region. It is determined whether or not an OB exists, and teacher data is generated by combining a user's detection determination result and feature information indicating features extracted from the detection area A on the image data, and based on the generated teacher data Thus, a learning model can be generated by a machine learning method, and rules set by the user can be updated according to the generated learning model. Thereby, the machine learning device 1 can update the rule for detection of the object OB scattered in the entire image data to a more appropriate rule based on the result of the machine learning. As a result, the machine learning device 1 can improve detection accuracy.

  Further, according to the present embodiment, the machine learning device 1 can automatically update the rule corresponding to the added image data even when new image data is added. The burden set for each can be reduced.

(Other Embodiments (Modifications))
Hereinafter, as another embodiment of the invention, it will be described that this embodiment can be used for a rule-based configuration corresponding to various cases.
For example, when it is necessary to review the priority order between rules, the machine learning device 1 has characteristics of rules that have not been applied due to the priority order as in the case of creating teacher data by the current method. Information may also be acquired and stored in teacher data. Thereby, the machine learning device 1 can verify whether or not the detection result can be obtained correctly when the rule is applied according to the current priority order using the survey data based on the determination result by the learning model. it can.

  Further, the machine learning device 1 may determine whether the rule is necessary or not by storing an accuracy result by rule detection. For example, when the accuracy does not increase due to the rule update, the machine learning device 1 is limited only by parameter adjustment, and a possibility of having contradictory rules is assumed. In such a case, the machine learning device 1 makes a determination based on a learning model by creating sampling data of a large number of patterns whose values are finely changed for feature information of two rules that may be contradictory. It becomes possible.

  The machine learning device 1 described above has a computer system inside. The operation of each processing unit described above is stored in a computer-readable recording medium in the form of a program, and the above processing is performed by the computer reading and executing this program. Here, the computer-readable recording medium means a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, a semiconductor memory, or the like. Alternatively, the computer program may be distributed to the computer via a communication line, and the computer that has received the distribution may execute the program.

  The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

  Although several embodiments of the present invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and the equivalents thereof.

DESCRIPTION OF SYMBOLS 1 ... Machine learning apparatus, 100 ... Image display part, 110 ... Image display control part, 200 ... Rule control part, 210 ... Rule information management part, 220 ... Rule detection execution part, 230 ... Rule verification part, 231 ... Representative data acquisition , 232 ... Survey data generation unit, 233 ... Rule verification execution unit, 300 ... Machine learning unit, 310 ... Teacher data management unit, 311 ... Teacher data generation unit, 312 ... Feature information generation unit, 313 ... Learning model generation unit, 320 ... Model classification execution unit, 400 ... Data management unit, 410 ... Image data management unit, 411 ... Image data storage unit, 420 ... Rule information management unit, 421 ... Rule information storage unit, 430 ... Machine learning data management unit, 431 ... teacher data storage unit, 432 ... model data storage unit

Claims (7)

A determination unit that applies a predetermined rule to feature information indicating a feature extracted from a predetermined region of image data and determines whether or not an object exists in the predetermined region;
An acquisition unit for acquiring user information input by allowing a user to visually recognize an image indicated by the image data;
An information generation unit that generates information that combines the feature information and the user information acquired by the acquisition unit;
A model generation unit that generates a learning model by a machine learning method based on the information generated by the information generation unit;
An update unit for updating the rule according to the learning model generated by the model generation unit;
A machine learning device comprising: The information generation unit generates a plurality of information that combines the feature information and the user information acquired by the acquisition unit,
A calculation unit that calculates an index indicating the degree to which the feature information included in the information generated by the information generation unit conforms to the rule;
An extraction unit that extracts information within a predetermined range of the index calculated by the calculation unit from the plurality of pieces of information generated by the information generation unit;
The determination unit determines whether or not an object exists in the predetermined region by applying the learning model to feature information included in the information extracted by the extraction unit,
The update unit applies the learning model to the feature information included in the information extracted by the extraction unit, and determines whether or not an object exists in the predetermined area to the determination unit. The rule is updated based on the determined model determination result and the rule determination result in which the determination unit determines whether an object exists in the predetermined region by applying the rule. To
The machine learning device according to claim 1. The update unit updates the rule so that the rule determination result matches the model determination result when the model determination result and the rule determination result do not match.
The machine learning device according to claim 2. The predetermined range is a range excluding a range where the index is equal to or greater than a first predetermined value and a range equal to or less than a second predetermined value which is lower than the first predetermined value.
The machine learning device according to claim 2 or 3. The calculation unit calculates an index for each of the plurality of feature information when the information generated by the information generation unit includes a plurality of feature information,
The extraction unit plots information having an index calculated by the calculation unit on a space having a number of dimensions corresponding to the number of the plurality of feature information, and indicates that the index is within the predetermined range Extracting the information from the region,
The machine learning device according to claim 2. Applying a predetermined rule to the feature information indicating the feature extracted from the predetermined area of the image data, it is determined whether or not an object exists in the predetermined area,
Obtaining user information selected by allowing the user to visually recognize the image indicated by the image data;
Generating information that combines the feature information and the acquired user information;
Based on the generated information, a learning model is generated by a machine learning method,
Updating the rules according to the generated learning model;
Machine learning method. On the computer,
Applying a predetermined rule to the feature information indicating the feature extracted from the predetermined region of the image data, it is determined whether or not an object exists in the predetermined region,
Causing the user to obtain user information selected by visually recognizing the image indicated by the image data;
Generating information that combines the feature information and the acquired user information;
Based on the generated information, a learning model is generated by a machine learning method,
In order to update the rule according to the generated learning model,
Machine learning program.

Images