Learning device, image classification device, learning method, and computer program

Abstract

To provide a learning model reduced in specific erroneous discrimination of discriminating a first object as a second object.SOLUTION: A learning device includes: a teacher data storage unit for storing teacher data including one or more sheets of image data capturing one or more discrimination objects of at least any one or more types among two or more discrimination objects each having a different type; an estimation unit for estimating the types of the discrimination objects included in the image data of the teacher data storage unit by using the learning model; a loss function evaluation unit for respectively calculating a degree of erroneous discrimination of discriminating a certain type of a discrimination object as another type of a discrimination object by using an estimation result of the estimation unit; and a learning unit for updating the learning model so as to reduce erroneous discrimination of a specific pattern of discriminating a first object belonging to a first type as a second object belonging to a second type with respect to the degree of erroneous discrimination calculated by the loss function evaluation unit.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a learning device, an image classification device, a learning method, and a computer program. [Background technology]

[0002] There is a technique for calculating a loss value using a loss function for classification results using a learning model, and updating the learning model using the loss value (see, for example, Patent Document 1). In the machine learning device described in Patent Document 1, the smaller the difference value between the likelihood of the true value and the estimated value for each class, the larger the weight to be multiplied by the loss value is set, and the learning model is updated. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2020-204909 Summary of the Invention [Problem to be solved by the invention]

[0004] Although the technology described in Patent Document 1 can improve the learning efficiency of a learning model, it only compares a true class, which is an estimated class that is correct, with a false class, which is an estimated class that is incorrect. Specifically, Patent Document 1 considers cases where class A is mistaken for class B and cases where class B is mistaken for class A in parallel. However, there are cases where it is desired to distinguish between certain error patterns, such as when class A is mistaken for class B but when class B is mistaken for class A, which is not acceptable. In this case, even if the technology described in Patent Document 1 is applied, there is a risk that misclassification of a specific pattern, such as mistaking class B for class A, cannot be suppressed.

[0005] The present invention has been made to solve the above-mentioned problems, and aims to provide a learning model that reduces the misclassification of specific patterns that discriminate a first object as a second object. [Means for solving the problem]

[0006] The present invention has been made to solve the above-mentioned problems, and can be realized in the following forms. A learning device comprising: a training data storage unit for storing training data including one or more pieces of image data in which at least one or more of two or more different types of discrimination objects are captured; an estimation unit for estimating the type of discrimination object captured in the image data in the training data storage unit using a learning model; and a calculation unit for calculating, using an estimation result from the estimation unit, a degree of misclassification in which a first object of one type is mistakenly classified as an object of another type, and for calculating, based on the calculated degree of misclassification, a first object belonging to a first type as a second object belonging to a second type. an evaluation unit that evaluates the degree of each misclassification using a loss evaluation function in which a weight preset for misclassification of a specific pattern discriminated as a second object is greater than a weight preset for misclassification of other patterns; and a learning unit that updates the learning model so that misclassification of a specific pattern in which a first object belonging to a first class is discriminated as a second object belonging to a second class is reduced relative to the degree of misclassification evaluated by the evaluation unit, and the estimation unit calculates a prediction probability P ii and the predicted probability P ij and the evaluation unit predicts the type of each object to be discriminated using the predicted probability P ii 、P ij and for the i-th object, only the i-th component corresponding to the i-th kind of the correct answer is 1, and the j-th component other than the i-th component is zero. i and a weight w multiplied by the degree of misclassification of the i-th object being erroneously predicted to be the j-th type. ij and a learning device that calculates and evaluates a loss evaluation function L1 shown below using the above. The present invention can also be realized in the following forms.

[0007] (1) According to one aspect of the present invention, there is provided a learning device comprising: a training data storage unit that stores training data including one or more image data each containing one or more images of at least one type of discrimination target object among two or more different types of discrimination target objects, an estimation unit that uses a learning model to estimate the type of the discrimination target object included in the image data in the training data storage unit, a loss function evaluation unit that uses the estimation result of the estimation unit to calculate the degree of misclassification of one type of discrimination target object as another type of discrimination target object, and a learning unit that updates the learning model so that the degree of misclassification of a specific pattern in which a first object belonging to a first type is discriminated as a second object belonging to a second type is reduced relative to the degree of misclassification calculated by the loss function evaluation unit.

[0008] According to this configuration, in order to suppress misclassification of a specific pattern in which a first object is discriminated as a second object among multiple types of misclassification, the generated learning model is updated to reduce misclassification of the specific pattern. As a result, the learning device of this configuration can provide a learning model that suppresses misclassification of the specific pattern among multiple types of misclassification more than other misclassifications.

[0009] (2) In the learning device of the above aspect, the loss function evaluation unit may evaluate the degree of each misclassification using a loss evaluation function in which a weight preset for misclassification of the specific pattern is greater than a weight preset for misclassification of other patterns. According to this configuration, by appropriately setting the weighting value set for the misclassification of a specific pattern that is desired to be suppressed, it is possible to provide a learning model that easily suppresses the misclassification of a specific pattern.

[0010] (3) In the learning device of the above aspect, the estimation unit predicts that the i-th object belonging to the i-th type (i=1, 2, . . .) is the correct i-th type by using a prediction probability P i and the predicted probability P j and the loss function evaluation unit predicts the type of each discrimination object using the predicted probability P i , P j and for the i-th object, only the i-th component corresponding to the i-th kind of the correct answer is 1, and the j-th component other than the i-th component is zero. i and a weight w that is multiplied by the degree of suppression of the misclassification of the i-th object being erroneously predicted as the j-th type. ij and may be used to calculate and evaluate the loss evaluation function L1 shown in the following formula (1).

number

[0011] (4) In the learning device of the above aspect, the learning unit may further generate a learning model that determines, from any image data, the type of object to be determined contained in the image data. According to this configuration, since the learning model is generated from scratch, it is possible to provide a learning model that suppresses misclassification of specific patterns even if there is no learning model transmitted from another device, etc.

[0012] (5) According to another aspect of the present invention, there is provided an image discrimination device, comprising: the learning device of the above aspect; an imaging unit that captures an image of a discrimination target; and a discrimination unit that uses the learning model to discriminate the type of discrimination target included in image data captured by the imaging unit. According to this configuration, it is possible to suppress erroneous discrimination of a specific pattern in which a first object is discriminated as a second object, and to discriminate a plurality of discrimination objects.

[0013] The present invention can be realized in various forms, for example, a learning device, an image discrimination device, an identification device, a sorting device, and systems equipped with these devices, a learning system, a learning method, and an image discrimination method, a computer program for executing these devices and methods, a server device for distributing this computer program, a non-transitory storage medium storing a computer program, etc. [Brief explanation of the drawings]

[0014] [Figure 1] 1 is a schematic diagram of an image discrimination device according to an embodiment of the present invention. [Figure 2] 1 is a flowchart of an image discrimination method. [Figure 3] FIG. 1 is an explanatory diagram of misidentification in Examples 1 and 2 and Comparative Example 1. [Figure 4] FIG. 10 is an explanatory diagram of misidentification in Examples 3 and 4 and Comparative Example 1. DETAILED DESCRIPTION OF THE INVENTION

[0015] <Embodiment> FIG. 1 is a schematic diagram of an image discrimination device (learning device) 100 according to an embodiment of the present invention. The image discrimination device 100 discriminates waste materials (OB) conveyed by a belt conveyor BC according to their material type (material quality) using a learning model MD generated by a learning unit 13. The waste materials in this embodiment include cast aluminum materials containing many additive elements, wrought aluminum materials containing few additive elements, and printed circuit boards (PCBs) that inevitably become mixed in with scrap. The image discrimination device 100 discriminates the material quality of each waste material using an image of the waste material captured by a photographing unit 60 and the learning model MD. Wrought aluminum materials containing few additive elements and cast aluminum materials are used in aluminum recycling, and are regenerated into alloys with different additive elements, while PCBs are not recycled. Therefore, while it is acceptable to erroneously discriminate wrought aluminum materials as cast aluminum materials, it is desirable to prevent erroneous discrimination of cast aluminum materials as wrought aluminum materials. In the learning model MD of this embodiment, a predetermined weight is set for each parameter during learning, so that erroneous discrimination of cast material as wrought material is suppressed.

[0016] As shown in FIG. 1, the image discrimination device 100 includes a PC (Personal Computer) 50 and an imaging unit 60 that captures images of waste as discrimination objects OB transported by a belt conveyor BC. The PC 50 includes a CPU (Central Processing Unit) 10 and a memory unit 20. The memory unit 20 is configured with a hard disk drive (HDD) or the like. The memory unit 20 includes a teacher database (teacher DB, teacher data storage unit) 21 that stores teacher data necessary for generating and updating a learning model MD, and a model database (model DB) 22 that saves the generated and updated learning model MD.

[0017] The training data stored in the training DB 21 of this embodiment includes one or more image data in which at least one of three different types of discrimination objects OB is captured. That is, one image data contains at least one of the wrought material, the cast material, and the PCB.

[0018] The CPU 10 included in the PC 50 executes the functions of various programs by expanding the programs stored in a read-only memory (ROM) (not shown) into a random access memory (RAM). The CPU 10 functions as a learning unit 13 that generates a learning model MD, an estimation unit 11 that uses the learning model MD to estimate the type of object to be discriminated OB appearing in each image data stored in the training data DB 21, an evaluation unit (loss function evaluation unit) 12 that evaluates the estimation results of the estimation unit 11, and a discrimination unit 14 that discriminates the type of object to be discriminated OB.

[0019] The learning unit 13 generates a learning model MD that distinguishes the type of discrimination object OB contained in any image data from the image data. The learning unit 13 of this embodiment generates a learning model MD that distinguishes between cast material, wrought material, and PCB using image data stored in the teacher data DB 21. The learning unit 13 generates the learning model MD by, for example, deep learning using a convolutional neural network. The learning unit 13 also updates the learning model MD using evaluation by the evaluation unit 12, which will be described later. The learning unit 13 saves the generated learning model MD in the model DB 22.

[0020] The estimation unit 11 of this embodiment uses the learning model MD to output, as estimation results, the probabilities that the type of each discrimination object OB is cast material, wrought material, and PCB for all discrimination objects OB captured in each image data stored in the training data DB 21. For example, when the estimation unit 11 estimates that the probability that one discrimination object OB captured in certain image data is cast material is 50%, the probability that it is wrought material is 40%, and the probability that it is PCB is 10%, the estimation unit 11 outputs predicted probabilities P=(0.5, 0.4, 0.1) as estimation results.

[0021] In addition, the estimation unit 11 of this embodiment calculates the predicted probability P i (i=1,2,...) and the predicted probability P j (j=1, 2, . . .) to predict the type of each discrimination object OB. The prediction probability P i is the probability of predicting that the i-th type object, whose type is the i-th type of discrimination object OB, is the correct i-th type object. On the other hand, the prediction probability P j is the probability of predicting that the i-th type object is the j-th type object, which is not the correct answer. In the above example, if the type of the discrimination object OB captured in the image data is cast material (i=1), P1=0.5, P2=0.4 (wrought material: j=2), and P3=0.1 (j=3). Note that, like the learning unit 13, the estimation unit 11 of this embodiment predicts the type of discrimination object OB captured in the image data by, for example, deep learning using a convolutional neural network.

[0022] The evaluation unit 12 calculates a loss evaluation function L1 expressed by the following formula (2) as the degree of misclassification of a certain type of object OB being erroneously classified as another type of object, using the estimation result of the estimation unit 11. As expressed by the following formula (2), the evaluation unit 12 of this embodiment calculates a loss evaluation function L1 by using a weight w 12 The loss evaluation function L1 is calculated using

[0023]

number

[0024] As shown in the above formula (2), the loss evaluation function L1 is calculated based on the predicted probability P j ,P j and the variable t i And the weight w ij and are calculated using the variable t i For the i-th object, only the i-th component corresponding to the i-th type of correct answer is set to 1, and the j-th component other than the i-th component is set to zero. ij is a weight multiplied by the degree of suppression of misclassification in which the ith object of the ith kind is mistakenly predicted to be the jth kind.

[0025] For example, if the correct answer for the object to be discriminated OB in the image data is a cast material, and the predicted probability P determined by the learning model MD is (0.5, 0.4, 0.1), the loss evaluation function L1 is calculated as shown in the following formula (3) using the above formula (2). Note that in this embodiment, the weight w 12 is set to 100, and the weight w 13 is set to 1.

[0026]

number

[0027] On the other hand, unlike the conditions under which the above formula (3) was calculated, if the correct answer for the object to be discriminated OB is similarly cast material and the prediction probability discriminated by the learning model MD is P = (0.5, 0.1, 0.4), the loss evaluation function L1 is calculated using the above formula (2) as shown in the following formula (4).

[0028]

number

[0029] The loss evaluation function L1 calculated by the above formula (4) is 11.7, which is smaller than the value of the loss evaluation function L1 calculated by the above formula (3), 51.8. This is because the predicted probability P of mistaking the cast material for the wrought material was 0.4 under the conditions of the above formula (3), but was 0.1 under the conditions of the above formula (4), and the weight w 12 And the weight w 13 The loss evaluation function L1 was different due to different numerical settings. In this way, even if the probability of correctly identifying cast material as cast material is the same, the value of the loss evaluation function L1 will change depending on the difference between the probability of identifying cast material as wrought material and the probability of identifying cast material as PCB.

[0030] The learning unit 13 updates the learning model MD so as to reduce the misclassification of a specific pattern in which the object to be discriminated OB, whose type is cast material, is mistakenly discriminated as wrought material, relative to the degree of misclassification calculated by the evaluation unit 12. Specifically, the learning unit 13 uses the evaluated loss evaluation function L1 to update the learning model MD so as to reduce the misclassification of a specific pattern different from other misclassifications (for example, misclassification of cast material as PCB). In this embodiment, the weight w 12 is the weight of other misclassifications w ij Since the value of the learning unit 13 is set to be larger than the value of the learning unit 13, the learning unit 13 can update the learning model MD so as to reduce the misclassification of the specific pattern that discriminates the cast material as the wrought material. The learning unit 13 stores the updated learning model MD in the model DB 22.

[0031] The discrimination unit 14 uses the learning model MD stored in the model DB 22 to discriminate waste as the discrimination object OB shown in the image data sent from the photographing unit 60. The discrimination unit 14 uses the learning model MD to discriminate waste flowing on the belt conveyor BC as either cast material, wrought material, or PCB. The discriminated waste is then sorted by workers, robots, etc. Note that the image data of the waste photographed by the photographing unit 60 is different data from the image data stored in the training data DB 21.

[0032] Fig. 2 is a flowchart of the image discrimination method. The image discrimination method shown in Fig. 2 includes a learning method for generating a learning model MD used for discrimination, and each step of discriminating waste using the learning model MD. In the image discrimination method, first, the learning unit 13 performs a training data acquisition step of acquiring image data from the training data DB 21 (step S1). The learning unit 13 then performs a learning step of generating a training model MD for discriminating the type of discrimination object OB using the acquired training data (step S2).

[0033] The estimation unit 11 performs an estimation step of estimating the type of the object to be discriminated OB appearing in each image data using the learning model MD generated by the learning unit 13 (step S3). The estimation unit 11 of this embodiment estimates the type of each object to be discriminated OB by calculating a predicted probability P i ,P j The evaluation unit 12 uses the estimation result of the estimation unit 11 to perform a loss function evaluation step (step S4) to evaluate the degree of misclassification of a specific pattern that erroneously classifies a cast material as a wrought material. The evaluation unit 12 of this embodiment calculates the weight w ij The estimation results are evaluated using a loss evaluation function L1 having the following formula:

[0034] The learning unit 13 performs an update process to update the generated learning model MD using the loss evaluation function L1 (step S5). The discrimination unit 14 uses the learning model MD updated by the learning unit 13 and image data captured by the imaging unit 60 to discriminate waste as discrimination objects OB flowing on the belt conveyor BC (step S6). The discrimination unit 14 discriminates whether there are still discrimination objects OB to be discriminated and whether there are still belt conveyors BC that are transported by the belt conveyor BC and that transport discrimination objects OB (step S7). If the discrimination unit 14 determines that there are still discrimination objects OB (step S7: YES), it continues the processing of step S6. If it determines that there are no more discrimination objects OB to be conveyed (step S7: NO), the flowchart of the image discrimination method ends.

[0035] Fig. 3 is an explanatory diagram of misclassification in Examples 1 and 2 and Comparative Example 1. Fig. 3 shows, in a table, the misclassification rate when the object to be classified OB is classified using a learning model MD generated using images of 33 pieces of cast aluminum, 30 pieces of wrought material (pipes), and 29 pieces of PCB as image data included in the training data.

[0036] In Example 1, the weight w 12 was set to 10, and the type of waste as the discrimination object OB photographed by the photographing unit 60 was discriminated by the learning model MD1 updated using the loss evaluation function L1 shown in equation (2) of the above embodiment. Example 2 was compared with Example 1 by using the weight w 12 was set to 100, and the type of the object OB to be discriminated was discriminated using the loss evaluation function L1 shown in the above formula (2). 12 Weight other than w ij is set to 1. In Comparative Example 1, the weight w ij and predicted probability P j The type of the object to be discriminated OB was discriminated by a learning model MD using a loss evaluation function L2 expressed by the following formula (5) that does not use and. Note that the learning unit 13 in Examples 1 and 2 used the stochastic gradient descent method when updating the learning model MD.

[0037]

number

[0038] When the loss evaluation function L2 of Comparative Example 1 is used, for example, if the probability estimated by the estimation unit 11 when the correct answer is the cast material is P=(0.5, 0.4, 0.1), (0.5, 0.1, 0.4), each loss evaluation function L2 is calculated as shown in the following equations (6) and (7).

[0039]

number

number

[0040] The loss evaluation function L2 calculated by the above formulas (6) and (7) is the same 1.31, even though the probability that the type of the discrimination object OB is estimated to be wrought material is different, 0.1 and 0.4. In other words, the loss evaluation function L2 of Comparative Example 1 evaluates whether or not there is a misclassification of the correct cast material, but does not take into account the type of misclassification. As a result, as shown in Figure 3, in Examples 1 and 2, the misclassification rate of the specific pattern in which the type of the discrimination object OB, which is cast material, is estimated to be wrought material is smaller than in Comparative Example 1.

[0041] Fig. 4 is an explanatory diagram of misclassification in Examples 3 and 4 and Comparative Example 1. Fig. 4 shows, in a table, the misclassification rate when the type of the object to be distinguished OB is distinguished using a learning model MD generated using images of 33 pieces of aluminum casting material, 30 pieces of pipe-shaped wrought material (hereinafter also simply referred to as "pipes"), and 30 pieces of fin-shaped wrought material (hereinafter also simply referred to as "fins") as image data included in the training data.

[0042] In Example 3, the weight w when a cast material is misidentified as a pipe is 12 is set to 100, and other weights W ij The type of the object to be discriminated OB was discriminated by the learning model MD1 updated using the loss evaluation function L1 shown in the formula (2) of the above embodiment, with the loss evaluation function L1 set to 1. In Example 4, in comparison with Example 3, the weight W 13 The loss evaluation function L1 shown in the above formula (2) was also set to 100, and the type of the discrimination object OB was discriminated using the loss evaluation function L1 shown in the above formula (2). In other words, Example 3 is a discrimination method that suppresses discrimination of a discrimination object OB whose type is a cast material as a pipe. Example 4 is a discrimination method that suppresses discrimination of a discrimination object OB whose type is a cast material as a pipe and discrimination of a discrimination object OB whose type is a cast material as a fin.

[0043] 4, the misclassification rate of the cast material as a pipe in Example 3 is 0.01, which is smaller than 0.10 in Comparative Example 1. Furthermore, the misclassification rate of the cast material as a pipe in Example 4 is 0.02, which is smaller than 0.10 in Comparative Example 1. Furthermore, the misclassification rate of the cast material as a fin in Example 4 is 0.01, which is smaller than 0.11 in Comparative Example 1. That is, in Examples 3 and 4, the misclassification rate of the cast material as a wrought material is smaller than in Comparative Example 1.

[0044] As described above, the image discrimination device 100 of this embodiment includes an estimation unit 11 that estimates discrimination objects OB captured in image data stored in the training data DB 21 using a learning model MD, and an evaluation unit 12 that evaluates the estimation results of the estimation unit 11. The evaluation unit 12 uses the estimation results of the estimation unit 11 to calculate the degree of misclassification, in which one type of discrimination object OB is mistakenly classified as another type of discrimination object. The learning unit 13 updates the learning model MD so as to reduce the misclassification of a specific pattern that erroneously classifies a discrimination object OB of a cast material type as wrought material, relative to the degree of misclassification evaluated by the evaluation unit 12. In this embodiment, for aluminum recycling, misclassification of other patterns that discriminate wrought material from cast material is tolerated, while the objective is to suppress misclassification of a specific pattern that discriminates cast material from wrought material. In order to suppress misclassification of a specific pattern among multiple types of misclassification, the learning unit 13 of this embodiment updates the generated learning model MD so as to reduce misclassification of a specific pattern, using the degree of misclassification evaluated by the evaluation unit 12. As a result, as shown in Figures 3 and 4, the image discrimination device 100 can provide a learning model MD that further suppresses misclassification of a specific pattern.

[0045] In addition, the evaluation unit 12 of this embodiment uses a weight w that is preset for the misidentification of a specific pattern that determines a cast material as a wrought material. ij By using the above, the image discrimination device 100 calculates a loss evaluation function L1 of an index that indicates the degree of misclassification. As a result, the image discrimination device 100 calculates a weight w ijBy appropriately setting the value of , it is possible to easily provide a learning model MD that suppresses misclassification of specific patterns.

[0046] Furthermore, the estimation unit 11 of this embodiment predicts that an i-th type object, which is the i-th type of discrimination object OB, is the correct i-th type object. i and the predicted probability P j The evaluation unit 12 predicts the type of each discrimination object OB using the predicted probability P i ,P j and the variable t i And the weight w ij The loss evaluation function L1 is calculated using the variable t i For the i-th object, only the i-th component corresponding to the i-th type of correct answer is set to 1, and the j-th component other than the i-th component is set to zero. ij is a weight to be multiplied by the degree of suppression of misclassification in which the i-th object is erroneously predicted to be the j-th type. That is, the evaluation unit 12 evaluates the prediction probability P i ,P j and the weight w corresponding to the misclassification of the specific pattern that we want to suppress. ij The learning unit 13 calculates a loss evaluation function L1 using the calculated loss evaluation function L1 and updates the learning model MD generated from the training data, thereby providing a learning model MD that suppresses misclassification of a specific pattern more than other misclassifications.

[0047] Furthermore, the learning unit 13 of this embodiment generates a learning model MD that discriminates the type of discrimination object OB from the image data stored in the training data DB 21. That is, in this embodiment, since the learning model MD is generated from scratch, it is possible to provide a learning model that suppresses erroneous discrimination of specific patterns even without a learning model MD transmitted from another device or the like.

[0048] The image discrimination device 100 of this embodiment also includes a photographing unit 60 that photographs the discrimination objects OB transported on the belt conveyor BC, and a discrimination unit 14 that uses a learning model MD to discriminate the type of discrimination object OB photographed by the photographing unit 60. Therefore, the image discrimination device 100 of this embodiment can discriminate the multiple conveyed discrimination objects OB by suppressing erroneous discrimination of a specific pattern that discriminates a discrimination object OB whose type is cast material as wrought material.

[0049] <Modifications of the above embodiment> The present invention is not limited to the above-described embodiment, and can be embodied in various forms without departing from the spirit of the invention. For example, the following modifications are also possible.

[0050] [Variation 1] The image discrimination device 100 of the above embodiment is merely an example, and various modifications are possible to the configuration of the image discrimination device 100 and the control performed by the image discrimination device 100. For example, the image discrimination device 100 may function as a learning device that generates a learning model MD without discriminating the type of waste flowing on the belt conveyor BC. The learning device may not include the imaging unit 60 and may not function as the discrimination unit 14. The learning device may discriminate the type of waste by exchanging various information with an imaging device configured as another device and a discrimination device functioning as the discrimination unit 14.

[0051] The learning unit in the modified example may not generate a learning model MD, but may acquire an existing learning model MD and update the acquired learning model MD using the results of the estimation unit 11 and the evaluation unit 12. The learning unit 13 and the estimation unit 11 may use a well-known learning method other than deep learning using a convolutional neural network. The learning unit 13 may generate the learning model MD using image data other than the image data included in the teacher data DB 21.

[0052] Although the evaluation unit 12 in the above embodiment evaluates the estimation result of the estimation unit 11 using the loss evaluation function L1 shown in the above formula (2), any known evaluation method may be used to evaluate the estimation result. The evaluation unit 12 may also calculate the degree of misclassification using a different loss evaluation function.

[0053] In the above embodiment, the objects to be distinguished OB are exemplified as cast material, wrought material, and PCB, but the objects to be distinguished OB are not limited to these. The type of the objects to be distinguished OB may be a parameter other than the material.

[0054] 2 shows an example of a flowchart of an image discrimination method, but discrimination using the learning model MD in step S6 does not have to be performed. For example, the learning method of the modified example may include a training data acquisition step (step S1), an estimation step using an existing learning model MD (step S3), a loss function evaluation step (step S4), and an update step (step S5).

[0055] In the above-described embodiments, a part of the configuration realized by hardware may be replaced by software, and conversely, a part of the configuration realized by software may be replaced by hardware. The present invention is not limited to the above-described embodiments, and can be embodied in various forms without departing from the spirit of the present invention.

[0056] This aspect has been described above based on embodiments and modifications. However, the above-described embodiments are intended to facilitate understanding of this aspect and are not intended to limit this aspect. This aspect may be modified or improved without departing from the spirit and scope of the claims, and equivalents thereof are included in this aspect. Furthermore, if a technical feature is not described as essential in this specification, it may be deleted as appropriate. [Explanation of symbols]

[0057] 10...CPU 11…Estimation part 12...Evaluation unit (loss function evaluation unit) 13...Study Department 14...Discrimination section 20...Storage section 60...Photography Department 100...Image discrimination device (learning device) BC: Belt conveyor 21...Teacher Database 22...Model Database L1, L2...Loss evaluation function MD, MD1...Learning model OB: Object to be distinguished P,P i ,P j …Predicted probability t i …variable

Claims

1. A learning device, comprising: a training data storage unit for storing training data including one or more pieces of image data in which at least one or more types of discrimination objects are captured among two or more types of discrimination objects that are different from each other; an estimation unit that estimates the type of the object to be discriminated that appears in the image data stored in the training data storage unit using a learning model; an evaluation unit that uses the estimation results of the estimation unit to calculate the degree of misclassification of a certain type of object to be discriminated as another type of object to be discriminated, and evaluates the degree of each misclassification using a loss evaluation function in which a weight previously set for misclassification of a specific pattern in which a first object belonging to a first type is discriminated as a second object belonging to a second type is made greater than a weight previously set for misclassification of other patterns; a learning unit that updates the learning model so that a misclassification of a specific pattern in which a first object belonging to a first type is discriminated as a second object belonging to a second type is reduced, relative to the degree of misclassification evaluated by the evaluation unit; and Equipped with the estimation unit predicts the type of each discrimination object using a prediction probability P ii that the i-th object belonging to the i-th type (i=1, 2, ...) is the correct i-th type and a prediction probability P ij that the i-th object is the incorrect j-th type (j=1, 2, ...), The evaluation unit the predicted probabilities P ii and P ij predicted by the estimation unit; a variable t i in which only the i-th component corresponding to the i-th type of the correct answer is 1 and the j-th component other than the i-th component is zero for the i-th object; and a weight wij by which the degree of misclassification of the i-th object being erroneously predicted to be the j-th type is multiplied; The learning device calculates and evaluates the loss evaluation function L1 shown in the following equation (8) using [Equation 8]

2. The learning device according to claim 1 , The learning unit further generates a learning model for discriminating the type of discrimination target object contained in any image data from the image data.

3. An image discrimination device, The learning device according to claim 1 or 2; an imaging unit that images an object to be identified; a discrimination unit that uses the learning model to discriminate the type of discrimination object included in the image data captured by the imaging unit; An image discrimination device comprising:

4. A learning method, comprising: a training data acquisition step of acquiring training data including one or more pieces of image data in which at least one or more types of discrimination target objects are captured among two or more types of discrimination target objects that are different from each other; an estimation step of estimating the type of the object to be discriminated that appears in the image data of the acquired training data using a learning model; an evaluation step of calculating a degree of misclassification in which a certain type of object is erroneously classified as another type of object using the estimation result of the estimation step, and evaluating the degree of each misclassification in relation to the calculated degree of misclassification using a loss evaluation function in which a weight preset for misclassification of a specific pattern in which a first object belonging to a first type is classified as a second object belonging to a second type is made larger than a weight preset for misclassification of other patterns; an updating step of updating the learning model so that a misclassification of a specific pattern in which a first object belonging to a first type is discriminated as a second object belonging to a second type is reduced with respect to the evaluated degree of misclassification; Run The estimation step predicts the type of each discrimination object using a prediction probability P ii that an i-th object belonging to an i-th type (i=1, 2, ...) is the correct i-th type and a prediction probability P ij that it is the incorrect j-th type (j=1, 2, ...), The evaluation step includes: the predicted probabilities P ii and P ij predicted by the estimation step; a variable t i in which only the i-th component corresponding to the i-th type of the correct answer is 1 and the j-th component other than the i-th component is zero for the i-th object; and a weight wij by which the degree of misclassification of the i-th object being erroneously predicted to be the j-th type is multiplied; A learning method in which the loss evaluation function L1 shown in the following equation (8) is calculated and evaluated using [Equation 9]

5. A computer program comprising: A training data storage function for storing training data including one or more pieces of image data in which at least one or more types of discrimination objects are captured among two or more types of discrimination objects that are different from each other; an estimation function that estimates the type of the object to be discriminated that appears in the image data included in the stored training data using a learning model; an evaluation function that calculates the degree of misclassification of a certain type of object to be discriminated as another type of object to be discriminated using the estimation result of the estimation function, and evaluates the degree of each misclassification using a loss evaluation function in which a weight preset for misclassification of a specific pattern in which a first object belonging to a first type is discriminated as a second object belonging to a second type is made greater than a weight preset for misclassification of other patterns; a learning function that updates the learning model so that the degree of misclassification of a specific pattern in which a first object belonging to a first type is discriminated as a second object belonging to a second type is reduced, based on the evaluated degree of misclassification; and on the computer, The estimation function predicts the type of each discrimination object using a prediction probability P ii that an i-th object belonging to an i-th type (i=1, 2, ...) is the correct i-th type and a prediction probability P ij that it is the incorrect j-th type (j=1, 2, ...), The evaluation function is the predicted probabilities P ii and P ij predicted by the estimation function; a variable t i in which only the i-th component corresponding to the i-th type of the correct answer is 1 and the j-th component other than the i-th component is zero for the i-th object; and a weight wij by which the degree of misclassification of the i-th object being erroneously predicted to be the j-th type is multiplied; A computer program that calculates and evaluates a loss evaluation function L1 shown in the following formula (8) using [Equation 10]

Images