Learning model generation program, learning model generation device, and learning model generation method

Abstract

To provide a learning model generation program, a learning model generation apparatus, and a learning model generation method that enable the detection of wild animals that frequently appear in a specific area.SOLUTION: The method uses a first learning model generated by performing machine learning using multiple pieces of first training data to determine a second label corresponding to multiple pieces of evaluation data, specifies certain evaluation data of the multiple pieces of the evaluation data erroneously determined to correspond to labels other than the second label as erroneously determined data, generates multiple pieces of second training data by associating the second label with each of multiple pieces of evaluation data classified into each class for each second number class including a class corresponding to a similar label similar to the erroneously determined label and a first number class, and determines the second label corresponding to the multiple pieces of evaluation data using a second training model generated by performing the machine learning using the multiple pieces of second training data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a learning model generation program, a learning model generation device, and a learning model generation method. [Background technology]

[0002] In recent years, there has been an increase in damage to crops in rural areas near mountains caused by wild animals such as bears and boars (hereinafter simply referred to as wild animals), and there is also an increasing risk of wild animals harming people in urban areas. For this reason, various machine learning models (hereinafter simply referred to as learning models) capable of recognizing and detecting wild animals that have appeared have been proposed in recent years (see Patent Documents 1 to 7). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Chinese Patent Application Publication No. 112615984 [Patent Document 2] Chinese Patent Application Publication No. 112989892 [Patent Document 3] Patent Publication No. 2021-086441 [Patent Document 4] Japanese Patent Publication No. 2020-067913 [Patent Document 5] Patent Publication No. 2021-114930 [Patent Document 6] Japanese Patent Publication No. 2020-021111 [Patent Document 7] Japanese Patent Publication No. 2020-171248 Summary of the Invention [Problem to be solved by the invention]

[0004] Here, the learning model used for detecting such wild animals may be trained for the purpose of detecting many types of wild animals that appear in various regions, and therefore, such a learning model may not be able to detect with high accuracy, for example, wild animals that appear frequently in a particular region.

[0005] Therefore, an object of the present invention is to provide a learning model generation program, a learning model generation device, and a learning model generation method that enable wild animals that appear in a specific area to be detected with a high frequency. [Means for solving the problem]

[0006] In order to achieve the above object, a learning model generation program according to the present invention adds a first label corresponding to an object appearing in each of a plurality of image data to each of the plurality of image data, classifies the plurality of image data into a first number of classes for each of the image data to which the same first label is added, and associates a second label corresponding to each class with each of the plurality of image data classified into each of the first number of classes, thereby generating a plurality of first training data, generates a first learning model by performing machine learning using the plurality of first training data, determines the second label corresponding to a plurality of evaluation data by using the first learning model, identifies evaluation data among the plurality of evaluation data that has been erroneously determined to correspond to a label other than the second label corresponding to each of the evaluation data as erroneously determined data, and identifies the second label corresponding to the identified erroneously determined data as erroneously determined label. the second label corresponding to each of the plurality of evaluation data is determined using the second learning model; and, if a second cost indicating a result of the determination using the second learning model is greater than a first cost indicating a result of the determination using the first learning model, output information indicating that the second learning model has higher determination accuracy than the first learning model. [Effects of the Invention]

[0007] According to the learning model generation program, learning model generation device, and learning model generation method of the present invention, it becomes possible to detect wild animals appearing in a specific area with a high frequency. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram showing an example of the configuration of an information processing device 1 according to the first embodiment. [Figure 2] FIG. 2 is a diagram illustrating an outline of the learning model generation process in the first embodiment. [Figure 3] FIG. 3 is a diagram illustrating an outline of the learning model generation process in the first embodiment. [Figure 4] FIG. 4 is a flowchart illustrating details of the learning model generation process according to the first embodiment. [Figure 5] FIG. 5 is a flowchart illustrating details of the learning model generation process in the first embodiment. [Figure 6] FIG. 6 is a flowchart illustrating details of the learning model generation process in the first embodiment. [Figure 7] FIG. 7 is a flowchart illustrating details of the learning model generation process in the first embodiment. [Figure 8] FIG. 8 is a flowchart illustrating details of the learning model generation process according to the first embodiment. [Figure 9] FIG. 9 is a flowchart illustrating details of the learning model generation process according to the first embodiment. [Figure 10] FIG. 10 is a diagram illustrating a specific example of the image data DT0. [Figure 11] FIG. 11 is a diagram illustrating a specific example of the similarity list LS. [Figure 12] FIG. 12 is a diagram illustrating a specific example of the process of S32. [Figure 13] FIG. 13 is a diagram illustrating a specific example of the process of S32. [Figure 14] FIG. 14 is a diagram illustrating a specific example of the process of S32. [Figure 15] FIG. 15 is a diagram illustrating a specific example of the cost in each learning model MD. [Figure 16] FIG. 16 is a diagram illustrating a specific example of the process of S41. [Figure 17] FIG. 17 is a diagram illustrating details of the learning model generation process in the first embodiment. [Figure 18] FIG. 18 is a diagram illustrating a specific example of the cost in each learning model MD. [Figure 19] FIG. 19 is a diagram illustrating details of the learning model generation process in the first embodiment. [Figure 20] FIG. 20 is a diagram illustrating details of the learning model generation process in the first embodiment. [Figure 21] FIG. 21 is a diagram illustrating a specific example of the cost in each learning model MD. [Figure 22] FIG. 22 is a diagram illustrating a specific example of the process of S41. [Figure 23] FIG. 23 is a diagram illustrating details of the learning model generation process in the first embodiment. [Figure 24] FIG. 24 is a diagram illustrating a specific example of the cost in each learning model MD. [Figure 25] FIG. 25 is a diagram illustrating a specific example of the process of S41. DETAILED DESCRIPTION OF THE INVENTION

[0009] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. However, the technical scope of the present invention is not limited to these preferred embodiments.

[0010] [Configuration Example of Information Processing Device 1 in First Embodiment] First, a description will be given of a configuration example of an information processing device 1 (hereinafter also referred to as a learning model generation device 1) in the first embodiment. Fig. 1 is a diagram showing a configuration example of the information processing device 1 in the first embodiment.

[0011] The information processing device 1 is a computer device, such as a general-purpose PC (Personal Computer). The information processing device 1 performs a process (hereinafter also referred to as a learning model generation process) to generate a learning model for detecting wild animals (hereinafter also simply referred to as wild animals) that appear in a specific area.

[0012] The information processing device 1 has the hardware configuration of a general-purpose computer device, and for example, as shown in Fig. 1, includes a CPU 101 which is a processor, a memory 102, a communication interface 103, and a storage medium 104. Each part is connected to each other via a bus 105.

[0013] The storage medium 104 has, for example, a program storage area (not shown) that stores a program (not shown) for performing a learning model generation process.

[0014] The storage medium 104 also includes a storage unit 110 (hereinafter also referred to as a storage area 110) that stores information used when performing the learning model generation process. The storage medium 104 may be, for example, a hard disk drive (HDD) or a solid state drive (SSD).

[0015] The CPU 101 executes a program loaded from the storage medium 104 into the memory 102 to perform a learning model generation process.

[0016] The communication interface 103 communicates with the worker terminal 2 via a network NW such as the Internet. The worker terminal 2 may be, for example, a PC (Personal Computer), and may be a terminal through which a worker (hereinafter simply referred to as a worker) who generates a learning model inputs necessary information.

[0017] [Outline of the first embodiment] Next, an outline of the learning model generation process in the first embodiment will be explained. Figures 2 and 3 are diagrams for explaining the outline of the learning model generation process in the first embodiment.

[0018] The information management unit 111 of the information processing device 1 stores, for example, various pieces of information input by a worker via the worker terminal 2 in the storage medium 104. Specifically, when the information management unit 111 receives input of image data DT0 used to generate a learning model MD1, a learning model MD2, etc. (hereinafter, these will also be collectively referred to simply as learning models MD), the information management unit 111 stores the received image data DT0 in the storage medium 104.

[0019] The label adding unit 112 of the information processing device 1 adds, for example, to each of the plurality of image data DT0 stored in the storage medium 104, a label (hereinafter also referred to as a first label) corresponding to an object appearing in each of the plurality of image data DT0. Specifically, the label adding unit 112 adds, for example, to each of the plurality of image data DT0 stored in the storage medium 104, a first label designated by the worker via the worker terminal 2. Note that the first label may be, for example, information indicating the type (e.g., name) of an object appearing in each of the plurality of image data DT0.

[0020] The class classification unit 113 of the information processing device 1 classifies, for example, each of the multiple image data DT0 to which the data generation unit 114 has added the same first label into multiple classes (hereinafter also referred to as the first number of classes) for each image data DT0 to which the data generation unit 114 has added the same first label.

[0021] Specifically, for example, the classifying unit 113 classifies the image data DT0 corresponding to each type of wild animal (for example, wild animals that the operator has determined need to be detected) into a class in which only the image data DT0 corresponding to that type is stored. Also, for example, the classifying unit 113 classifies the image data DT other than the image data DT0 corresponding to wild animals into a class in which all the image data DT other than the image data DT0 corresponding to wild animals are stored together. That is, the classifying unit 113 classifies the image data DT0 so that the number of classes in the image data DT0 used to generate the learning model MD1 that is generated initially is as small as possible.

[0022] The data generation unit 114 of the information processing device 1 generates a plurality of training data DT11 (hereinafter also referred to as first training data DT11) by, for example, associating a label (hereinafter also referred to as second label) corresponding to each of the plurality of image data DT0 classified into each class for each of the first number of classes classified by the classifying unit 113. That is, the data generation unit 114 generates a plurality of first training data DT11 each including, for example, each image data DT0 and a second label corresponding to the class into which each image data DT0 is classified.

[0023] As shown in FIG. 3, the model generation unit 115 of the information processing device 1 generates a learning model MD1 (hereinafter also referred to as the first learning model MD1) by performing machine learning using, for example, a plurality of first training data DT11 generated by the data generation unit 114. Note that the model generation unit 115 may determine the progress of the learning of the first learning model MD1 (the determination accuracy of the first learning model MD) by using, for example, a plurality of validation data (not shown) during the learning of the learning model MD. The plurality of validation data may be, for example, a portion of the plurality of first training data DT11 generated by the data generation unit 114. Specifically, the plurality of validation data may be, for example, a plurality of first training data DT11 that was not used in generating the first learning model MD1.

[0024] The model evaluation unit 116 of the information processing device 1 determines second labels corresponding to the plurality of evaluation data DT2, for example, by using the first learning model MD1 generated by the model generation unit 115. The plurality of evaluation data DT2 may be, for example, a portion of the plurality of first training data DT11 generated by the data generation unit 114. Specifically, the plurality of evaluation data DT2 may be, for example, the plurality of first training data DT11 that was not used to generate the first learning model MD1 and that was not used as the plurality of validation data.

[0025] Then, as shown in FIG. 3, the model evaluation unit 116 calculates, for example, a cost C1 (hereinafter also referred to as a first cost C1) indicating the determination result of the second label corresponding to the plurality of evaluation data DT2.

[0026] The class addition unit 117 of the information processing device 1, for example, identifies evaluation data DT2 erroneously determined by the first learning model MD1 (hereinafter also referred to as erroneously determined data DT2) from among a plurality of evaluation data DT2. That is, the class addition unit 117 identifies, for example, erroneously determined data DT2 erroneously determined by the first learning model MD1 to correspond to a label (hereinafter also referred to as another label) other than the second label corresponding to each evaluation data DT2 from among the plurality of evaluation data DT2. Then, the class addition unit 117 identifies, for example, a second label (hereinafter also referred to as an erroneously determined label) corresponding to the identified erroneously determined data DT2. Furthermore, the class addition unit 117 identifies, for example, a label (hereinafter also referred to as a similar label) similar to the erroneously determined label from among the first labels.

[0027] Then, for example, for each piece of image data DT0 to which the data generation unit 114 has added the same first label, the class classification unit 113 classifies each of the multiple pieces of image data DT0 to which the data generation unit 114 has added the first label into multiple classes (hereinafter also referred to as the second number of classes) including the first number of classes and the similar labels identified by the class addition unit 117.

[0028] Next, the data generation unit 114 generates a plurality of training data DT12 (hereinafter also referred to as second training data DT12) by, for example, associating a second label corresponding to each class with each of the plurality of image data DT0 classified into each class, for each of the second number of classes classified by the classification unit 113. Note that hereinafter, the first training data DT11 and the second training data DT12, etc., will also be collectively referred to simply as training data DT1.

[0029] Next, the model generation unit 115 generates a learning model MD2 (hereinafter also referred to as the second learning model MD2) by performing machine learning using, for example, multiple second training data DT12 generated by the data generation unit 114, as shown in Figure 3.

[0030] Furthermore, the model evaluation unit 116 determines the second labels corresponding to the multiple pieces of evaluation data DT2, for example, by using the second learning model MD2 generated by the model generation unit 115. Then, the model evaluation unit 116 calculates, for example, a cost C2 (hereinafter also referred to as the second cost C2) indicating the determination result of the second labels corresponding to the multiple pieces of evaluation data DT2, as shown in FIG.

[0031] As shown in FIG. 3, the model determination unit 118 of the information processing device 1 determines the magnitude relationship between the first cost C1 calculated by the model evaluation unit 116 and the second cost C2 calculated by the model evaluation unit 116, for example.

[0032] For example, when the model determination unit 118 determines that the first cost C1 is greater than the second cost C2, that is, when it determines that the second cost C2 does not exceed the first cost C1, the information output unit 119 of the information processing device 1 outputs information indicating that the first learning model MD1 has higher determination accuracy than the second learning model MD2. Then, the information processing device 1 ends the learning model generation process.

[0033] Similarly, the information processing device 1 also terminates the learning model generation process when, for example, the class adding unit 117 determines that no similar label exists.

[0034] On the other hand, when the model determination unit 118 determines that the second cost CL2 is greater than the first cost CL1, the information processing device 1 continues the execution of the learning model generation process and starts generating a new learning model MD.

[0035] Then, the information processing device 1 repeats the learning model generation process until the cost of the newly generated learning model MD becomes larger than the cost of the previously generated learning model MD.

[0036] In other words, since the types of wild animals that appear in a specific, restricted area tend to be limited, it can be concluded that the learning model MD that detects wild animals that appear in a specific area is more likely to be able to improve its judgment accuracy by classifying the training data DT1 into fewer classes than the learning model MD that detects wild animals that appear in other, wider areas.

[0037] Therefore, the information processing device 1 in this embodiment generates a learning model MD with as few classes as possible, and further generates one or more learning models MD while gradually increasing the number of classes. Then, for example, when the cost of the newly generated learning model MD is no longer larger than the cost of the previously generated learning model MD, the information processing device 1 determines that the determination accuracy of the newly generated learning model MD does not exceed the determination accuracy of the previously generated learning model MD, and ends the generation of the new learning model MD.

[0038] This allows the information processing device 1 in this embodiment to generate a learning model MD suitable for detecting wild animals that appear in a specific area, thereby improving the accuracy of detecting wild animals that appear in a specific area.

[0039] Furthermore, the information processing device 1 in this embodiment generates multiple learning models MD while gradually increasing the number of classes, thereby making it possible to reduce the size (number of parameters) of the learning model MD generated at the end of the learning model generation process. Therefore, the information processing device 1 can generate a learning model MD that can operate even in an environment with limited resource amounts, for example.

[0040] Hereinafter, a case where a learning model MD that recognizes the type of object (hereinafter also referred to as an object recognition model MD) is generated will be described, but the learning model generation process in this embodiment may also be a case where a learning model MD that detects the object itself (hereinafter also referred to as an object detection model MD) is generated. In this case, the information processing device 1 generates the learning model MD (object detection model MD) by using training data DT1 that includes position information of the object in the image data DT0 (for example, position information specified by a bounding box) in addition to the image data DT0 and the second label.

[0041] [Details of the first embodiment] Next, details of the learning model generation process in the first embodiment will be described. Figures 4 to 9 are flowcharts explaining details of the learning model generation process in the first embodiment. Figures 10 to 24 are diagrams explaining details of the learning model generation process in the first embodiment.

[0042] [Image data storage processing] First, the process of storing image data used to generate a learning model in the storage medium 104 (hereinafter also referred to as image data storage process) will be described, out of the learning model generation process.

[0043] 4, the information management unit 111 waits until it receives input of image data DT0 used to generate the learning model MD (NO in S1). Specifically, the information management unit 111 waits until the worker inputs the image data DT0 via the worker terminal 2, for example.

[0044] Then, when the input of the image data DT0 is accepted (YES in S1), the information management unit 111 stores the accepted input of the image data DT0 in the storage medium 104 (S2).

[0045] [Label addition process] Next, the process of adding a first label to image data DT0 (hereinafter also referred to as label adding process) in the learning model generation process will be described.

[0046] 5, the label adding unit 112 waits until it receives input of a first label corresponding to an object (e.g., a wild animal) shown in image data DT0 stored in the storage medium 104 (NO in S11). Specifically, the label adding unit 112 waits until the worker inputs the first label via the worker terminal 2, for example.

[0047] Then, when the input of the first label is accepted (YES in S11), the label adding unit 112 adds the accepted input first label to the image data DT0 corresponding to the accepted input first label (S12). A specific example of the image data DT0 will be described below.

[0048] [Example of image data DT0] Fig. 10 is a diagram illustrating a specific example of image data DT0. Specifically, Fig. 10 is a diagram illustrating a specific example of image data DT0 after a first label has been added. Note that the following description will be given assuming that the character string written above each image data DT0 is the first label.

[0049] The image data DT0 shown in Figure 10 includes, for example, daytime image data DT0 in which the object is a bear (image data DT0 with the first label "bear"), daytime image data DT0 in which the object is a boar (image data DT0 with the first label "boar"), and daytime image data DT0 in which the object is a car (image data DT0 with the first label "car").

[0050] Furthermore, the image data DT0 shown in Figure 10 includes, for example, nighttime image data DT0 in which the object is a bear (image data DT0 with the first label "bearnight"), daytime image data DT0 in which the object is a boar (image data DT0 with the first label "boarnight"), and daytime image data DT0 in which the object is a car (image data DT0 with the first label "carnight").

[0051] Furthermore, the image data DT0 shown in Figure 10 includes, for example, daytime image data DT0 in which no objects are shown (image data DT0 with the first label ``background (bg)''), nighttime image data DT0 in which no objects are shown (image data DT0 with the first label ``bg night''), and image data DT0 in which it is difficult to identify the objects (image data DT0 with the first label ``abnormal (dark)'').

[0052] That is, the image data DT0 shown in Fig. 10 includes image data DT0 corresponding to 18 types of first labels including "bear", "boar", etc. Description of the other image data DT0 included in Fig. 10 will be omitted.

[0053] [Similar List Memory Processing] Next, the process of storing the similarity list LS (hereinafter also referred to as similarity information LS) in the storage medium 104 (hereinafter also referred to as similarity list storage process) in the learning model generation process will be described.

[0054] 6, the information management unit 111 waits, for example, until it receives input of a similarity list LS (NO in S21). The similarity list LS is a list that is generated in advance by an operator based on information such as the color and shape of the object, and indicates the similarity relationships of each of the first labels. Specifically, the information management unit 111 waits, for example, until the operator inputs the similarity list LS via the operator terminal 2.

[0055] Then, when the input of the similarity list LS is accepted (YES in S21), the information management unit 111 stores the accepted input of the similarity list LS in the storage medium 104 (S22). A specific example of the similarity list LS will be described below.

[0056] [Example of similar list LS] FIG. 11 is a diagram illustrating a specific example of the similarity list LS.

[0057] The information in the first row of the similarity list LS shown in Fig. 11 indicates, for example, that image data DT0 whose first label is "bear" is similar to image data DT0 whose first label is "craw." The information in the second row of the similarity list LS shown in Fig. 11 indicates, for example, that image data DT0 whose first label is "boar" is similar to image data DT0 whose first label is "monkey," and further indicates that image data DT0 whose first label is "boar" is similar to image data DT0 whose first label is "racoon." A description of the other information included in Fig. 11 will be omitted.

[0058] [Main process of learning model generation] Next, the main processing of the learning model generation processing will be described.

[0059] 7, the classifying unit 113 waits until it is time to generate a model (NO in S31), for example. The model generation time is the time when the worker inputs information to start generating a learning model via the worker terminal 2, for example.

[0060] Then, when it is time to generate a model (YES in S31), the classification unit 113 classifies each of the multiple pieces of image data DT0 stored in the storage medium 104 into a first number of classes, for example, for each piece of image data DT0 to which the same first label is added (S32). Note that the first number of classes may be, for example, designated in advance by the worker via the worker terminal 2. A specific example of the processing of S32 will be described below.

[0061] [Example of S32 processing] 12 to 14 are diagrams for explaining a specific example of the processing of S32. In the following, the explanation will be made assuming that the first number of classes is 9 classes, and that the minimum number of classes judged to be necessary for generating a learning model that enables the detection of bears and wild boars has been specified. That is, as shown in FIG. 12, the class into which image data DT0 whose first label is "bear" (hereinafter also referred to as the "bear class"), the class into which image data DT0 whose first label is "bg" (hereinafter also referred to as the "bg class"), the class into which image data DT0 whose first label is "boar" (hereinafter also referred to as the "boar class"), the class into which image data DT0 whose first label is "bearnight" (hereinafter also referred to as the "bearnight class"), and the class into which image data DT0 whose first label is "bgnight" (hereinafter also referred to as the "bgnight" class) are classified. The following explanation will be given assuming that the following classes have been specified in advance: a class into which image data DT0 with a first label of "boarnight" is classified (hereinafter also referred to as the boarnight class), a class into which image data DT0 with a first label of "dark" is classified (hereinafter also referred to as the dark class), a class into which daytime image data DT0 not classified into any class is classified (hereinafter also referred to as the other class), and a class into which nighttime image data DT0 not classified into any class is classified (hereinafter also referred to as the othernight class).

[0062] Specifically, the class classification unit 113, for example, classifies image data DT0 whose first label is "bear" into the bear class, image data DT0 whose first label is "bg" into the bg class, image data DT0 whose first label is "boar" into the boar class, image data DT0 whose first label is "bearnight" into the bearnight class, image data DT0 whose first label is "bgnight" into the bgnight class, image data DT0 whose first label is "boarnight" into the boarnight class, and image data DT0 whose first label is "dark" into the dark class.

[0063] In addition, as shown in Figure 13, the class classification unit 113 classifies, for example, image data DT0 whose first label is "car", image data DT0 whose first label is "cat", image data DT0 whose first label is "craw", image data DT0 whose first label is "dog", image data DT0 whose first label is "man", image data DT0 whose first label is "monkey", and image data DT0 whose first label is "racoon" into the "other" class.

[0064] Furthermore, as shown in Figure 14, the class classification unit 113 classifies, for example, image data DT0 whose first label is "racoonnight", image data DT0 whose first label is "carnight", image data DT0 whose first label is "mannight", and image data DT0 whose first label is "rabbitnight" into the othernight class.

[0065] Returning to FIG. 7, the data generation unit 114 generates a plurality of first training data DT11 (S33), for example, by associating a second label corresponding to each class with each of the plurality of image data DT0 classified into each class for each of the first number of classes classified in the processing of S32.

[0066] Specifically, for example, for each piece of image data DT0 classified into the bear class, the data generation unit 114 generates first training data DT11 having each piece of image data and information indicating "bear", which is a second label corresponding to each piece of image data. Furthermore, for example, for each piece of image data DT0 classified into the boar class, the data generation unit 114 generates first training data DT11 having each piece of image data and information indicating "boar", which is a second label corresponding to each piece of image data. Furthermore, for example, for each piece of image data DT0 classified into the other class, the data generation unit 114 generates first training data DT11 having each piece of image data and information indicating "other", which is a second label corresponding to each piece of image data. A description of the other pieces of first training data DT11 included in FIG. 12 will be omitted.

[0067] Returning to FIG. 7, the model generation unit 115 generates a first learning model MD1 by performing machine learning using, for example, the plurality of first training data DT11 generated in the processing of S33 (S34).

[0068] Next, the model evaluation unit 116 determines the second label corresponding to each of the multiple evaluation data DT2, for example, by using the first learning model MD1 generated in the process of S34 (S35).

[0069] Then, the model evaluation unit 116 calculates, for example, a first cost C1 indicating the determination result of the second label corresponding to each of the multiple evaluation data DT2 (S36). A specific example of the first cost C1 will be described below.

[0070] [Specific examples of costs for each learning model] 15, 18, 21, and 24 are diagrams illustrating specific examples of costs in each learning model MD. Specifically, FIG. 15 is a diagram illustrating a specific example of the first cost C1. More specifically, FIG. 15(A) is a diagram illustrating a specific example of a determination result corresponding to the bear class and the bearnight class. Also, FIG. 15(B) is a diagram illustrating a specific example of a determination result corresponding to the boar and the boarnight class. Note that the following description will be given assuming that the number of evaluation data DT2 for each first label is 25. Also, the following description will be given assuming that the number of types of first labels is 18, and the total number of evaluation data DT2 is 450 (18×25=450).

[0071] Specifically, as shown in the "bear" column in Figure 15(A), for example, among the evaluation data DT2 including image data DT0 with the first label "bear", the number of evaluation data DT2 for which the second label was correctly determined in the processing of S35 is 25, and as shown in the "bearnight" column in Figure 15(A), among the evaluation data DT2 including image data DT0 with the first label "bearnight", the number of evaluation data DT2 for which the second label was correctly determined in the processing of S35 is 22, the model evaluation unit 116 will, as shown in each column in Figure 15(A), identify 47 as the number of evaluation data DT2 that is True Positive (hereinafter also referred to as TP), identify 3 as the number of evaluation data DT2 that is False Negative (hereinafter also referred to as FN), identify 8 as the number of evaluation data DT2 that is False Positive (hereinafter also referred to as FP), and identify 392 as the number of evaluation data DT2 that is True Negative (hereinafter also referred to as TN).

[0072] Then, as shown in each column in Figure 15(A), the model evaluation unit 116 calculates, for example, 0.98 as Accuracy according to the following formula (1), 0.94 as Recall according to the following formula (2), 0.85 as Precision according to the following formula (3), and 0.90 as F-value (hereinafter also referred to as F value) according to the following formula (4).

[0073] Accurary=(TP+TN) / (TP+FN+FP+TN) ···(Formula (1)) Recall=TP / (TP+FN)...(Formula (2)) Precision=TP / (TP+FP) (Formula (3)) F-value=(2×Recall×Precision) / (Recall+Precision) ···(Equation (4)) Similarly, as shown in the "boar" column in Figure 15(B), for example, among the evaluation data DT2 containing image data DT0 whose first label is "boar", the number of evaluation data DT2 for which the second label was correctly determined in the processing of S35 is 22, and as shown in the "boarnight" column in Figure 15(B), among the evaluation data DT2 containing image data DT0 whose first label is "boarnight", the number of evaluation data DT2 for which the second label was correctly determined in the processing of S35 is 19, the model evaluation unit 116 will, as shown in each column in Figure 15(B), identify 41 as the number of evaluation data DT2 that is TP, identify 9 as the number of evaluation data DT2 that is FN, identify 14 as the number of evaluation data DT2 that is FP, and identify 386 as the number of evaluation data DT2 that is TN.

[0074] Then, in this case, as shown in each column in FIG. 15(B), the model evaluation unit 116 calculates, for example, 0.95 as Accurate, 0.82 as Recall, 0.75 as Precision, and 0.78 as F-value.

[0075] Then, the model evaluation unit 116 calculates, for example, the first cost C1 of the first learning model MD1 as 0.78, which is the minimum of the F-values ​​0.90 calculated for the bear class and the bearnight class and 0.78 calculated for the boar class and the boarnight class.

[0076] Returning to FIG. 8, the class addition unit 117 determines whether or not it is possible to add (extend) a class for classifying a plurality of image data DT0 stored in the storage medium 104 (S41).

[0077] Specifically, the class adding unit 117, for example, identifies evaluation data DT2 (misjudged data DT2) in which the first learning model MD1 erroneously judges the second label corresponding to each evaluation data DT2 from among the multiple evaluation data DT2. Then, the class adding unit 117, for example, identifies a second label (misjudged label) corresponding to the identified misjudged data DT2. Furthermore, the class adding unit 117, for example, refers to the similarity list LS stored in the storage medium 104, and identifies, from the identified misjudged labels, a similar label that is similar to a label that has not yet been included in the first number of classes. Thereafter, the class adding unit 117 identifies a class corresponding to the identified similar label as a new class (hereinafter also referred to as an added class).

[0078] This allows the class addition unit 117 to identify, as an additional class, a class that is more likely to contribute to improving the determination accuracy of the next learning model MD (second learning model MD2) to be generated. A specific example of the process of S41 will be described below.

[0079] [Example of processing in S41] Fig. 16 is a diagram illustrating a specific example of the process of S41. Specifically, Fig. 16(A) is a diagram illustrating a specific example of the process of S41 corresponding to Fig. 15(A). Also, Fig. 16(B) is a diagram illustrating a specific example of the process of S41 corresponding to Fig. 15(B).

[0080] Specifically, the example shown in Fig. 16(A) shows that the three pieces of image data DT0 corresponding to FN described in Fig. 15(A) include two pieces of image data DT0 that were erroneously determined as "bear" and one piece of image data DT0 that was erroneously determined as "boarnight". Also, the example shown in Fig. 16(A) shows that the eight pieces of image data DT0 corresponding to FP described in Fig. 15(A) include two pieces of image data DT0 that should have been correctly determined as "boar", three pieces of image data DT0 that should have been determined as "boarnight", one piece of image data DT0 that should have been determined as "craw", and two pieces of image data DT0 that should have been determined as "racoonnight".

[0081] 16(A) shows that a misjudgment occurred between image data DT0 whose second labels are "bear," "boarnight," "boar," "craw," and "racoonnight," respectively, and image data DT0 whose second label is "bear" or "bearnight." Therefore, the class adding unit 117 specifies, as candidates for classes to be added, classes that are first labels corresponding to the image data DT0 in which a misjudgment occurred between the image data DT0 whose first label is "bear" or "bearnight," and that correspond to the first labels "craw" and "racoonnight" corresponding to the image data DT0 whose current second label is "other" or "othernight."

[0082] 11 indicates that image data D0 whose first labels are "craw" and "racoonnight" are similar to image data DT0 whose first label is either "bear" or "bearnight." Therefore, the class adding unit 117, for example, identifies, from among the identified candidates for added classes, the classes corresponding to "craw" and "racoonnight," respectively (the craw class and the racoonnight class) as added classes.

[0083] 16(B) shows that the 9 pieces of image data DT0 corresponding to FN described in FIG. 15(B) include 2 pieces of image data DT0 that were erroneously determined as "bear", 1 piece of image data DT0 that was erroneously determined as "other", 3 pieces of image data DT0 that were erroneously determined as "bearnight", and 3 pieces of image data DT0 that were erroneously determined as "boar". The example shown in FIG. 16(B) shows that the 14 pieces of image data DT0 corresponding to FP described in FIG. 15(B) include 1 piece of image data DT0 that should have been correctly determined as "bearnight", 4 pieces of image data DT0 that should have been determined as "monkey", 1 piece of image data DT0 that should have been determined as "rabbitnight", 7 pieces of image data DT0 that should have been determined as "racoon", and 1 piece of image data DT0 that should have been determined as "racoonnight".

[0084] 16(B) shows that a misjudgment occurred between image data DT0 whose second labels are "bear," "other," "bearnight," "boar," "monkey," "rabbitnight," "racoon," and "racoonnight," respectively, and image data DT0 whose second label is "boar" or "boarnight." Therefore, the class adding unit 117 specifies, as candidates for classes to be added, classes that are first labels corresponding to image data DT0 in which a misjudgment occurred between image data DT0 whose first label is "boar" or "boarnight," and that correspond to each of the first labels "monkey," "rabbitnight," "racoon," and "racoonnight" corresponding to image data DT0 whose current second label is "other" or "othernight."

[0085] 11 indicates that image data D0 whose first labels are "monkey," "racoon," and "racoonnight" are similar to image data DT0 whose first label is either "boar" or "boarnight." Therefore, the class adding unit 117, for example, identifies, from among the identified candidates for added classes, the classes corresponding to "monkey," "racoon," and "racoonnight," respectively (monkey class, racoon class, and racoonnight class) as added classes.

[0086] That is, as shown in FIG. 17, the class addition unit 117 identifies, as additional classes, the craw class, the racoonnight class, the monkey class, and the racoon class, which are the union of the craw class and the racoonnight class, which are additional classes identified from FIG. 16(A), and the monkey class, the racoon class, and the racoonnight class, which are additional classes identified from FIG. 16(B).

[0087] Returning to Figure 8, if it is determined that it is not possible to add (expand) a class to classify the multiple image data DT0 stored in the storage medium 104, that is, if it is determined that there is no additional class (NO in S41), the information processing device 1 terminates the learning model generation process.

[0088] On the other hand, if it is determined that it is possible to add (extend) classes to classify the multiple image data DT0 stored in the storage medium 104, that is, if it is determined that additional classes exist (YES in S41), the class classification unit 113 classifies each of the multiple image data DT0 stored in the storage medium 104, for example, for each image data DT0 to which the same first label is attached, into a second number of classes including the first number of classes and the additional classes determined to exist in the previous processing of S41 (S42).

[0089] Next, the data generation unit 114 generates a plurality of second training data DT12, for example, by associating a second label corresponding to each class with each of the plurality of image data DT0 classified into each class for each of the second number of classes classified in the processing of S42 (S43).

[0090] Then, the model generation unit 115 generates a second learning model MD2 by performing machine learning using, for example, the plurality of second training data DT12 generated in the processing of S43 (S44).

[0091] Furthermore, as shown in FIG. 9, the model evaluation unit 116 determines the second labels corresponding to the multiple pieces of evaluation data DT2, for example, by using the second learning model MD2 generated in the process of S44 (S51).

[0092] Then, the model evaluation unit 116 calculates, for example, a second cost C2 indicating the determination result of the second label corresponding to the plurality of evaluation data DT2 (S52).

[0093] Specifically, as shown in Figure 18, the model evaluation unit 116 calculates the second cost C2 of the second learning model MD2 as 0.82, which is the minimum of the F-value calculated for the bear class and the bearnight class (see Figure 18(A)) and the F-value calculated for the boar class and the boarnight class (see Figure 18(B)).

[0094] Thereafter, the model determining unit 118 determines whether or not the second cost C2 calculated in the process of S52 is greater than the first cost C1 calculated in the process of S36 (S53).

[0095] As a result, when it is determined that the second cost C2 is not greater than the first cost C1 (NO in S53), the information processing device 1 ends the learning model generation process.

[0096] In this case, the information output unit 119 may output information indicating that the first learning model MD1 has the best determination accuracy. Even if it is determined that the second cost C2 is greater than the first cost C1, the information processing device 1 may terminate the learning model generation process if classes corresponding to all labels included in the similarity list LS have already been identified as additional classes.

[0097] On the other hand, when it is determined that the second cost C2 is greater than the first cost C1 (YES in S53), the information processing device 1 performs the processes from S41 onwards again.

[0098] That is, as shown in FIG. 8, the class adding unit 117 determines again whether or not it is possible to add (extend) a class for classifying a plurality of image data DT0 stored in the storage medium 104 (S41).

[0099] Specifically, the class adding unit 117, for example, identifies evaluation data DT2 (misjudged data DT2) in which the second learning model MD2 has erroneously judged the second label corresponding to each evaluation data DT2 from among the multiple evaluation data DT2. Then, the class adding unit 117, for example, identifies a second label (misjudged label) corresponding to the identified misjudged data DT2. Furthermore, the class adding unit 117, for example, refers to the similarity list LS stored in the storage medium 104, and identifies, from the identified misjudged labels, a similar label that is similar to a label that has not yet been included in the second number of classes. Thereafter, the class adding unit 117 identifies the class corresponding to the identified similar label as an added class.

[0100] As a result, if it is determined that it is possible to add (extend) a class to classify the multiple image data DT0 stored in the storage medium 104, that is, if it is determined that an additional class exists (YES in S41), the class classification unit 113 classifies each of the multiple image data DT0 stored in the storage medium 104, for example, for each image data DT0 to which the same first label is attached, into multiple classes (hereinafter also referred to as the third number of classes) including the second number of classes and the additional class determined to exist in the processing of the immediately preceding S41 (S42).

[0101] Next, the data generation unit 114 generates a plurality of training data DT13 (hereinafter also referred to as third training data DT13) by, for example, associating a second label corresponding to each class with each of the plurality of image data DT0 classified into each class for each of the third number of classes classified in the processing of S42 (S43).

[0102] Then, the model generation unit 115 generates a learning model MD3 (hereinafter also referred to as the third learning model MD3) by performing machine learning using, for example, the multiple third training data DT13 generated by the processing of S43 (S44).

[0103] Furthermore, as shown in FIG. 9, the model evaluation unit 116 determines the second labels corresponding to the multiple pieces of evaluation data DT2 by using, for example, the third learning model MD3 generated in the process of S44 (S51).

[0104] Then, the model evaluation unit 116 calculates, for example, a cost C3 (hereinafter also referred to as a third cost C3) indicating the determination result of the second label corresponding to the plurality of evaluation data DT2 (S52).

[0105] Thereafter, as shown in FIG. 19, the model determination unit 118 determines, for example, whether the third cost C3 calculated in the previous (immediately preceding) processing of S52 is greater than the second cost C2 calculated in the processing of S52 before that (S53).

[0106] As a result, when it is determined that the third cost C3 is not greater than the second cost C2 (NO in S53), the information processing device 1 ends the learning model generation process.

[0107] On the other hand, if it is determined that the third cost C3 is greater than the second cost C2, the information processing device 1 performs the processes from S41 onwards again.

[0108] 17 and other examples, classes corresponding to all labels included in the similarity list LS described in FIG. 11 are identified as additional classes. Therefore, in this case, the information processing device 1, for example, terminates the learning model generation process (NO in S41). Then, in this case, the information output unit 119 outputs, for example, information indicating that the second learning model MD2 has the best determination accuracy.

[0109] In this way, the information processing device 1 in this embodiment generates, for example, a learning model (first learning model MD1) with as few classes as possible, and further generates one or more learning models (second learning model MD2 and third learning model MD3) while gradually increasing the number of classes. Then, for example, when the cost of the newly generated learning model is no longer larger than the cost of the previously generated learning model, the information processing device 1 determines that the determination accuracy of the newly generated learning model does not exceed the determination accuracy of the previously generated learning model, and terminates the generation of the new learning model.

[0110] This allows the information processing device 1 in this embodiment to generate a learning model suitable for detecting wild animals that appear in a specific area, thereby improving the accuracy of detecting wild animals that appear in a specific area.

[0111] Furthermore, the information processing device 1 in this embodiment generates multiple learning models MD while gradually increasing the number of classes, thereby making it possible to reduce the size (number of parameters) of the learning model MD generated at the end of the learning model generation process. Therefore, the information processing device 1 can generate a learning model MD that can operate even in an environment with limited resource amounts, for example.

[0112] In addition, the information processing device 1 in this embodiment is capable of generating not only an object recognition model MD but also an object detection model MD by using training data DT1 including position information of an object in image data DT0 (for example, position information specified by a bounding box).

[0113] The information processing device 1 may be configured to generate a learning model MD capable of detecting wild animals at daytime with higher accuracy by training the model using only image data DT0 captured during the day. Similarly, the information processing device 1 may be configured to generate a learning model MD capable of detecting wild animals at night with higher accuracy by training the model using only image data DT0 captured during the night.

[0114] [Modification of the first embodiment] Next, a modified example of the first embodiment will be described, focusing on the differences from the learning model generation process described with reference to Figs. 4 to 9 (processing from S41 onwards).

[0115] In a modification of the first embodiment, in the process of S41, a class that the learning model MD has erroneously determined as a class corresponding to a wild animal (a wild animal to be detected) a predetermined number of times or more (for example, three times or more) is identified as an additional class. A specific example of the process of S41 will be described below.

[0116] [Example of processing in S41 (1)] The example shown in Fig. 16(A) shows that misjudgments occurred between image data DT0 whose second labels are "bear", "boarnight", "boar", "craw", and "racoonnight", respectively, and image data DT0 whose second labels are "bear" or "bearnight". The example shown in Fig. 16(A) also shows that the number of times misjudgments occurred between image data DT0 whose second label is "boarnight" and image data DT0 whose second label is "bear" or "bearnight" was three or more.

[0117] Therefore, the class addition unit 117 determines that there is no first label corresponding to image data DT0 for which erroneous judgments have occurred three or more times between the image data DT0 whose first label is "bear" or "bearnight" and the image data DT0 whose current second label is "other" or "othernight", and does not identify a candidate class to add.

[0118] On the other hand, the example shown in Figure 16(B) shows that misjudgments occurred between image data DT0 whose second labels are "bear", "other", "bearnight", "boar", "monkey", "rabbitnight", "racoon", and "racoonnight", respectively, and image data DT0 whose second label is "boar" or "boarnight". Also, the example shown in Figure 16(B) shows that the number of misjudgments that occurred between image data DT0 whose second labels are "bearnight", "monkey", and "racoon" and image data DT0 whose second label is "boar" or "boarnight" was three or more times.

[0119] Therefore, the class addition unit 117 identifies, as candidates for classes to be added, the classes corresponding to the first labels "bearnight," "monkey," and "racoon" that correspond to image data DT0 for which three or more misjudgments have occurred between the image data DT0 whose first label is "boar" or "boarnight" and the image data DT0 whose current second label is "other" or "othernight."

[0120] 11 indicates that image data DT0 whose first labels are "monkey" and "racoon," respectively, are similar to image data DT0 whose first label is either "bear" or "bearnight." Therefore, the class adding unit 117, for example, identifies, from among the identified candidates for added classes, the classes corresponding to "monkey" and "racoon," respectively (monkey class and racoon class) as added classes.

[0121] That is, as shown in FIG. 20, the class adding unit 117 identifies, for example, the monkey class and the racoon class, which are the classes to be added identified from FIG. 16(B), as the classes to be added.

[0122] Next, in the process of S42, the classification unit 113 classifies each of the multiple image data DT0 stored in the storage medium 104 into a second number of classes, for example, for each image data DT0 to which the same first label is added.

[0123] Then, in the processing of S43, the data generation unit 114 generates a plurality of second training data DT12, for example, by associating a second label corresponding to each class with each of the plurality of image data DT0 classified into each class, for each of the second number of classes.

[0124] Furthermore, in the processing of S44, the model generation unit 115 generates a second learning model MD2 by performing machine learning using, for example, a plurality of second training data DT12.

[0125] Then, as shown in Figure 21, the model evaluation unit 116 calculates the second cost C2 of the second learning model MD2 as 0.80, which is the minimum of the F-value calculated for the bear class and the bearnight class (see Figure 21(A)) and the F-value calculated for the boar class and the boarnight class (see Figure 21(B)).

[0126] 21, the first cost C1 of the first learning model is 0.78, and the second cost C2 of the second learning model MD2 is 0.80. Therefore, in this case, the information processing device 1 performs the processes from S41 onwards again. A specific example of the process of S41 (second time) will be described below.

[0127] [Example of processing in S41 (2)] Fig. 22 is a diagram illustrating a specific example of the process of S41. Specifically, Fig. 22(A) is a diagram illustrating a specific example of the process of S41 corresponding to Fig. 21(A). Also, Fig. 22(B) is a diagram illustrating a specific example of the process of S41 corresponding to Fig. 21(B).

[0128] Specifically, the example shown in Fig. 22(A) shows that the 13 pieces of image data DT0 corresponding to FN shown in Fig. 21(A) include 5 pieces of image data DT0 that were erroneously determined as "bear," 1 piece of image data DT0 that was erroneously determined as "boar," 1 piece of image data DT0 that was erroneously determined as "other," 2 pieces of image data DT0 that were erroneously determined as "bgnight," 2 pieces of image data DT0 that were erroneously determined as "boarnight," and 2 pieces of image data DT0 that were erroneously determined as "othernight." Also, the example shown in Fig. 22(A) shows that the 5 pieces of image data DT0 corresponding to FP shown in Fig. 21(A) include 3 pieces of image data DT0 that should have been correctly determined as "boar," 1 piece of image data DT0 that should have been determined as "boarnight," and 1 piece of image data DT0 that should have been determined as "racoonnight."

[0129] That is, the example shown in Fig. 22(A) shows that misjudgments occurred between image data DT0 whose second labels are "bear", "boar", "other", "bgnight", "boarnight", and "othernight", respectively, and image data DT0 whose second label is "bear" or "bearnight". Also, the example shown in Fig. 22(A) shows that the number of times misjudgments occurred between image data DT0 whose second labels are "bear" and "boar", respectively, and image data DT0 whose second label is "bear" or "bearnight" was three or more.

[0130] Therefore, the class addition unit 117 determines that there is no first label corresponding to image data DT0 for which erroneous judgments have occurred three or more times between the image data DT0 whose first label is "bear" or "bearnight" and the image data DT0 whose current second label is "other" or "othernight", and does not identify a candidate class to add.

[0131] On the other hand, the example shown in FIG. 22(B) shows that misjudgments occurred between image data DT0 whose second labels are "bear," "bearnight," "othernight," "dog," "monkey," and "racoon," respectively, and image data DT0 whose second label is "boar" or "boarnight." The similarity list LS shown in FIG. 11 also shows that, among the image data DT0 whose second label is "othernight," the first label of image data DT0 that is similar to image data DT0 whose second label is "boar" or "boarnight" is "racoonnight." Therefore, the example shown in FIG. 22(B) shows that the number of times misjudgments occurred between image data DT0 whose second label is "boar" or "boarnight" and image data DT0 whose second label is "bear" or "racoonnight" was three or more.

[0132] Therefore, the class addition unit 117 identifies, as a candidate class to be added, a class that is a first label corresponding to image data DT0 for which a misjudgment has occurred three or more times between the image data DT0 whose first label is "boar" or "boarnight" and the image data DT0 whose current second label is "other" or "othernight", and is a first label corresponding to "racoonnight".

[0133] 11 indicates that image data D0 whose first label is "racoonnight" is similar to image data DT0 whose first label is either "boar" or "boarnight." Therefore, the class adding unit 117 identifies, for example, a class corresponding to "racoonnight" (racoonnight class) as the class to be added.

[0134] That is, as shown in FIG. 23, the class adding unit 117 identifies, for example, the racoonnight class, which is the added class identified from FIG. 22(B), as the added class.

[0135] Next, in the process of S42, the classification unit 113 classifies each of the multiple image data DT0 stored in the storage medium 104 into a third number of classes, for example, for each image data DT0 to which the same first label is added.

[0136] Then, in the processing of S43, the data generation unit 114 generates a plurality of third training data DT13, for example, by associating a second label corresponding to each class with each of the plurality of image data DT0 classified into each class, for each of the third number of classes.

[0137] Furthermore, in the process of S44, the model generation unit 115 generates a third learning model MD3 by performing machine learning using, for example, the multiple pieces of third training data DT13 generated in the process of S43.

[0138] Then, as shown in Figure 24, the model evaluation unit 116 calculates, as the third cost C3 of the second learning model MD2, the minimum value of 0.83 (see Figure 24(B)), between the F-value of 0.83 calculated for the bear class and the bearnight class (see Figure 24(A)) and the F-value of 0.83 calculated for the boar class and the boarnight class.

[0139] 24, the second cost C2 of the second learning model is 0.80, and the third cost C3 of the third learning model MD3 is 0.83. Therefore, in this case, the information processing device 1 performs the processes from S41 onwards again. A specific example of the process of S41 (third time) will be described below.

[0140] [Example of processing in S41 (3)] Fig. 25 is a diagram illustrating a specific example of the process of S41. Specifically, Fig. 25(A) is a diagram illustrating a specific example of the process of S41 corresponding to Fig. 24(A). Also, Fig. 25(B) is a diagram illustrating a specific example of the process of S41 corresponding to Fig. 24(B).

[0141] Specifically, the example shown in Fig. 25(A) shows that the 8 (pieces) of image data DT0 corresponding to FN shown in Fig. 24(A) include 2 (pieces) of image data DT0 that were erroneously determined as "monkey", 2 (pieces) of image data DT0 that were erroneously determined as "bear", 1 (piece) of image data DT0 that was erroneously determined as "boar", 1 (piece) of image data DT0 that was erroneously determined as "boarnight", and 2 (pieces) of image data DT0 that were erroneously determined as "racoonnight". Also, the example shown in Fig. 25(A) shows that the 9 (pieces) of image data DT0 corresponding to FP shown in Fig. 24(A) include 5 (pieces) of image data DT0 that should have been correctly determined as "boar", 3 (pieces) of image data DT0 that should have been determined as "boarnight", and 1 (piece) of image data DT0 that should have been determined as "racoonnight".

[0142] That is, the example shown in Fig. 25(A) shows that misjudgments occurred between image data DT0 whose second labels are "monkey", "bear", "boar", "boarnight", and "racoonnight", respectively, and image data DT0 whose second label is "bear" or "bearnight". Also, the example shown in Fig. 22(A) shows that there is no second label for which the number of times misjudgments occurred between image data DT0 whose second label is "bear" or "bearnight" is three or more.

[0143] Therefore, the class addition unit 117 determines that there is no first label corresponding to image data DT0 for which erroneous judgments have occurred three or more times between the image data DT0 whose first label is "bear" or "bearnight" and the image data DT0 whose current second label is "other" or "othernight", and does not identify a candidate class to add.

[0144] On the other hand, the example shown in Fig. 25(B) shows that misjudgments occurred between image data DT0 whose second labels are "bear", "bearnight", "boar", "monkey", "rabbitnight", "racoon", and "racoonnight" and image data DT0 whose second label is "boar" or "boarnight". Also, the example shown in Fig. 22(B) shows that there is no second label in which the number of times misjudgments occurred between image data DT0 whose second label is "boar" or "boarnight" was three or more.

[0145] Therefore, the class addition unit 117 determines that there is no first label corresponding to image data DT0 for which three or more erroneous judgments have occurred between the image data DT0 whose first label is "boar" or "boarnight" and the image data DT0 whose current second label is "other" or "othernight", and does not identify a candidate class to add.

[0146] That is, the class adding unit 117 determines that there is no added class in the process of S41 performed for the third time. Therefore, in this case, the information processing device 1 ends, for example, the learning model generation process. [Explanation of symbols]

[0147] 1: Information processing device 2: Worker terminal 101:CPU 102: Memory 103: Communication interface 104:Storage medium 105: Bus 111: Information Management Department 112: Label adding section 113: Classification section 114: Data generation unit 115: Model generation unit 116: Model evaluation unit 117: Class Addition Section 118: Model determination unit 119: Information output unit

Claims

1. adding a first label corresponding to an object appearing in each of the plurality of image data to each of the plurality of image data; classifying the plurality of image data into a first number of classes for each image data to which the same first label is added; generating a plurality of first training data sets by associating a second label corresponding to each class with each of the plurality of image data sets classified into each class for each of the first number of classes; generating a first learning model by performing machine learning using the plurality of first training data; determining the second label corresponding to a plurality of evaluation data by using the first learning model; Identifying, among the plurality of evaluation data, evaluation data that has been erroneously determined to correspond to a label other than the second label corresponding to each evaluation data as erroneously determined data, and identifying the second label corresponding to the identified erroneously determined data as an erroneously determined label; classifying the plurality of image data into a second number of classes including the first number of classes and a class corresponding to a similar label similar to the erroneous determination label, for each image data to which the same first label is added; generating a plurality of second training data sets by associating the second labels corresponding to each class with each of the plurality of image data sets classified into each class for each of the second number of classes; generating a second learning model by performing machine learning using the plurality of second training data; determining the second labels corresponding to the plurality of evaluation data by using the second learning model; If a second cost indicating the result of the determination when the second learning model is used is greater than a first cost indicating the result of the determination when the first learning model is used, output information indicating that the second learning model has higher determination accuracy than the first learning model. A learning model generation program that causes a computer to execute processing.

2. In claim 1, the first number of classes includes a class into which the first label corresponding to a predetermined detection target object is classified, and a class into which the first label corresponding to an object other than the detection target object is classified, A learning model generation program characterized by:

3. In claim 1, Each of the first learning model and the second learning model is an object recognition model or an object detection model. A learning model generation program characterized by:

4. In claim 1, In the process of identifying the erroneously determined label, a label for which a number of evaluation data items that have been erroneously recognized as corresponding to the same other label is equal to or exceeds a predetermined number is identified as the erroneously determined label, In the process of classifying into the second number of classes, the other identical labels are identified as the similar labels. A learning model generation program characterized by:

5. In claim 4, further comprising: storing similarity information indicating the similarity relationship of the first labels in a storage unit; Have the computer execute the process, In the process of classifying the labels into the second number of classes, the similarity information stored in the storage unit is referenced, and if it is determined that the similarity information includes information indicating that the erroneously determined label and the other label are in a similar relationship, the other label is identified as the similar label. A learning model generation program characterized by:

6. In claim 1, the first cost is an F-value corresponding to the result of the determination when the first learning model is used; The second cost is an F-value corresponding to the result of the determination when the second learning model is used. A learning model generation program characterized by:

7. In claim 1, In the outputting process, when the first cost is greater than the second cost, information indicating that the first learning model has a higher determination accuracy than the second learning model is output. A learning model generation program characterized by:

8. In claim 1, further comprising: Identifying, among the plurality of evaluation data, evaluation data that has been erroneously determined by the second learning model to correspond to a label other than the second label corresponding to each evaluation data as erroneously determined data, and identifying the second label corresponding to the identified erroneously determined data as a new erroneously determined label; classifying the plurality of image data into a third number of classes including the second number of classes and a class corresponding to a similar label similar to the new erroneous determination label, for each image data to which the same first label is added; generating a plurality of third training data sets by adding the second labels corresponding to each class to each of the plurality of image data sets classified into each class for each of the third number of classes; generating a third learning model by performing machine learning using the plurality of third training data; determining the second label corresponding to the plurality of evaluation data by using the third learning model; If a third cost indicating the result of the determination when the third learning model is used is greater than the second cost, output information indicating that the third learning model has higher determination accuracy than the second learning model. A learning model generation program that causes a computer to execute processing.

9. a label adding unit that adds a first label corresponding to an object appearing in each of a plurality of image data to each of the plurality of image data; a classification unit that classifies the plurality of image data into a first number of classes for each image data to which the same first label is added; a data generation unit that generates a plurality of first training data by associating a second label corresponding to each class with each of the plurality of image data classified into each class, for each of the first number of classes; a model generation unit that generates a first learning model by performing machine learning using the plurality of first training data; a model evaluation unit that uses the first learning model to determine the second label corresponding to a plurality of evaluation data; a class addition unit that identifies, among the plurality of evaluation data, evaluation data that has been erroneously determined to correspond to a label other than the second label corresponding to each evaluation data as erroneously determined data, and identifies the second label corresponding to the identified erroneously determined data as an erroneously determined label, the classifying unit classifies the plurality of image data into a second number of classes including the first number of classes and a class corresponding to a similar label similar to the erroneous determination label, for each image data to which the same first label is added; the data generation unit generates a plurality of second training data by associating the second label corresponding to each class with each of the plurality of image data classified into each class, for each of the second number of classes; the model generation unit generates a second learning model by performing machine learning using the plurality of second training data; The model evaluation unit determines the second label corresponding to the plurality of evaluation data by using the second learning model, and further an information output unit that outputs information indicating that the second learning model has higher determination accuracy than the first learning model when a second cost indicating the result of the determination when the second learning model is used is greater than a first cost indicating the result of the determination when the first learning model is used; A learning model generation device characterized by:

10. adding a first label corresponding to an object appearing in each of the plurality of image data to each of the plurality of image data; classifying the plurality of image data into a first number of classes for each image data to which the same first label is added; generating a plurality of first training data sets by associating a second label corresponding to each class with each of the plurality of image data sets classified into each class for each of the first number of classes; generating a first learning model by performing machine learning using the plurality of first training data; determining the second label corresponding to a plurality of evaluation data by using the first learning model; Identifying, among the plurality of evaluation data, evaluation data that has been erroneously determined to correspond to a label other than the second label corresponding to each evaluation data as erroneously determined data, and identifying the second label corresponding to the identified erroneously determined data as an erroneously determined label; classifying the plurality of image data into a second number of classes including the first number of classes and a class corresponding to a similar label similar to the erroneous determination label, for each image data to which the same first label is added; generating a plurality of second training data sets by associating the second labels corresponding to each class with each of the plurality of image data sets classified into each class for each of the second number of classes; generating a second learning model by performing machine learning using the plurality of second training data; determining the second labels corresponding to the plurality of evaluation data by using the second learning model; If a second cost indicating the result of the determination when the second learning model is used is greater than a first cost indicating the result of the determination when the first learning model is used, output information indicating that the second learning model has higher determination accuracy than the first learning model. A learning model generation method characterized in that the processing is executed by a computer.

Images