JP7161948B2 - Machine learning device, machine learning method, machine learning program, and inspection device - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Law ▲1 ▼ Date of publication October 29, 2018 ▲2 ▼ Publication “Textbook for Deep Learning Utilization” Chapter 3, pp.126-130 ▲3 ▼ Publisher Nikkei BP

The present invention relates to a machine learning device, a machine learning method, and a machine learning program that cause a classifier to perform machine learning. The present invention also relates to an inspection apparatus for inspecting articles.

A technique for classifying the state of an article by referring to an image containing the article as a subject is widely used. When implementing such a technique using a classifier constructed by machine learning, it is difficult to maintain classification accuracy over a long period of time. This is because the generalization ability of the classifier constructed by machine learning is demonstrated only for items included in the images used as training data, or items similar to those items, which are the targets of classification. This is because if the environment surrounding the article itself to be classified or the environment surrounding the article to be classified changes, the article to be classified will exceed the range in which the generalization ability of the classifier can be demonstrated.

As a technique that may contribute to solving such problems, for example, a sound source type identification device described in Patent Document 1 can be cited. This sound source type determination device is a device that uses a neural network (corresponding to a classifier in this specification) to determine a sound source type. re-learning means for performing re-learning (corresponding to additional learning in this specification) of the neural network based on the re-learning means.

Japanese Patent Application Laid-Open No. 2001-33304 (published on February 9, 2001)

As in the sound source type determination means described in Patent Document 1, by performing additional learning of the classifier, even if the article itself to be classified or the environment surrounding the article to be classified changes, the classifier classification accuracy can be maintained for a long period of time. However, when additional learning is performed by such a method, the additional learning of the classifier cannot be performed before the classification accuracy of the classifier used for classifying the articles decreases. In other words, a temporary drop in classification accuracy of the classifier used to classify the articles is unavoidable.

The present invention has been made in view of the above problems, and is a machine learning apparatus and a machine learning method capable of performing additional learning of a classifier used for classifying articles before the classification accuracy of the classifier decreases. , and to realize a machine learning program. Another object of the present invention is to realize an inspection apparatus equipped with such a machine learning device.

To solve the above problem, a machine learning device according to an aspect of the present invention includes a plurality of classifiers that classify states of articles belonging to each of a plurality of groups with reference to an image including the article as a subject. A machine learning device for performing machine learning on a classifier group consisting of: an evaluation unit for evaluating the classification accuracy of each classifier included in the classifier group for each group; If the classification accuracy of one classifier does not reach a predetermined standard, the at least one of the classifiers included in the classifier group before performing the classification process for the articles belonging to the next group. and a learning control unit that causes a classifier other than the classifier to perform additional learning.

To solve the above problems, a machine learning method according to an aspect of the present invention includes a plurality of classifiers that classify states of articles belonging to each of a plurality of groups with reference to an image including the article as a subject. A machine learning method for performing machine learning on a classifier group consisting of: an evaluation step of evaluating the classification accuracy of each classifier included in the classifier group for each group; If the classification accuracy of one classifier does not reach a predetermined standard, the at least one of the classifiers included in the classifier group before performing the classification process for the articles belonging to the next group. and a learning control step of causing a classifier other than the classifier to perform additional learning.

In order to solve the above problems, a machine learning program according to one aspect of the present invention is a machine learning program that causes a computer to operate as the machine learning device described above, and causes the computer to function as each part of the machine learning device. .

In order to solve the above problems, an inspection device according to an aspect of the present invention includes the machine learning device described above and the classifier group, and uses the classifier group to inspect a plurality of articles divided into lots. conduct an inspection. The number of articles included in each lot may or may not be uniform.

According to these configurations, it is possible to perform additional learning of classifiers other than the classifier, triggered by a decrease in the classification accuracy of a certain classifier. In other words, it is possible to perform additional learning of the classifier used to classify the goods (classifier other than the classifier) before the classification accuracy of the classifier used to classify the goods (the classifier other than the classifier) decreases. can.

In the machine learning device according to an aspect of the present invention, the evaluation unit determines the classification accuracy of each classifier included in the classifier group, and the classification result obtained by the classifier classifying the state of each article. and a classification result obtained by a person classifying the state of each article.

According to the above configuration, the classification accuracy of each classifier is evaluated based on the result of classification of the state of the article by a person. Therefore, the timing for performing additional learning can be appropriately specified without depending on the generalization ability of the main classifier.

In the machine learning device according to an aspect of the present invention, among the plurality of classifiers, the classifier having the highest classification accuracy achieved in the initial learning phase is set as the main classifier, and the evaluation unit selects the classifier group as the main classifier. The classification accuracy of each classifier included is the classification result obtained by classifying the state of each article by the classifier and the classification result obtained by classifying the state of each article by the main classifier. is preferably evaluated by comparing .

According to the above configuration, the classification accuracy of each classifier is evaluated based on the result of classifying the state of the article by the main classifier. Therefore, it is possible to appropriately identify the timing at which additional learning should be performed without the need for a person to classify the state of the article.

In the machine learning device according to an aspect of the present invention, among the plurality of classifiers, the classifier with the highest classification accuracy achieved in the initial learning phase is used as the main classifier, and in the first additional learning phase, the evaluation The department determines the classification accuracy of each classifier included in the classifier group by combining the classification result obtained by classifying the state of each article by the classifier and the classification result obtained by classifying the state of each article by a person. and a second additional learning phase following the first additional learning phase, the evaluation unit evaluates the classification accuracy of each classifier included in the classifier group. is evaluated by comparing the classification result obtained by classifying the state of each article by the classifier and the classification result obtained by classifying the state of each article by the main classifier. , is preferred.

According to the above configuration, in the first additional learning phase, it is possible to appropriately identify the timing at which additional learning should be performed without depending on the generalization ability of the main classifier, and the second additional learning phase , it is possible to appropriately identify the timing at which additional learning should be performed without the need for a person to classify the state of an article.

In the machine learning device according to an aspect of the present invention, the learning control unit, when the classification accuracy of classifiers other than the main classifier included in the classifier does not reach a predetermined standard, determines that the classifier group It is preferable to cause the main classifier to perform additional learning before performing the classification process for items belonging to the next group.

According to the above configuration, it is possible to perform additional learning of the main classifier, triggered by a decrease in the classification accuracy of the classifiers other than the main classifier. In other words, additional learning of the main classifier can be performed before the classification accuracy of the main classifier used for classifying the articles deteriorates.

In the machine learning device according to an aspect of the present invention, the learning control unit determines that, when the classification accuracy of at least one classifier included in the classifier group does not reach a predetermined standard, the classifier group It is preferable to cause all the classifiers included in the classifier group to perform additional learning before executing the classification process for the articles belonging to the group.

According to the above configuration, it is possible to maintain the detection accuracy of classifiers other than the classifier that detects a sign of deterioration in the classification accuracy of the classifier used for classifying articles.

According to one aspect of the present invention, a machine learning device, a machine learning method, and a machine learning program capable of performing additional learning of a classifier used for classifying articles before the classification accuracy of the classifier decreases are realized. can do. Further, according to one aspect of the present invention, it is possible to realize an inspection device including such a machine learning device.

1 is a block diagram showing the physical configuration of a machine learning device according to a first embodiment of the present invention; FIG. 2 is a block diagram showing a functional configuration of the machine learning device shown in FIG. 1; FIG. (a) is a flow chart showing the flow of a machine learning method S1 executed by the machine learning apparatus shown in FIG. 1 in a first additional learning phase. (b) is a flow chart showing the flow of the machine learning method S2 executed in the second additional learning phase. It is a block diagram showing a functional configuration of an inspection apparatus according to a second embodiment of the present invention. 5 is a schematic diagram showing the flow of a semiconductor laser manufacturing process in the embodiment of the inspection apparatus shown in FIG. 4; FIG. 5 is a box-and-whisker diagram showing the classification accuracy of each classifier included in the classifier group for each group in the embodiment of the inspection apparatus shown in FIG. 4. FIG.

[First embodiment]
(Physical configuration of machine learning device)
A physical configuration of the machine learning device 1 according to the first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing the physical configuration of the machine learning device 1. As shown in FIG.

The machine learning device 1 is a computer including a bus 10, a main memory 11, a processor 12, an auxiliary memory 13, and an input/output interface 14, as shown in FIG. Main memory 11 , processor 12 , auxiliary memory 13 and input/output interface 14 are interconnected via bus 10 . As the main memory 11, for example, single or multiple semiconductor RAMs (random access memories) are used. Processor 12 may be, for example, a single or multiple microprocessors, a single or multiple digital signal processors, a single or multiple microcontrollers, or a combination thereof. As the auxiliary memory 13, for example, a single or multiple HDDs (Hard Disk Drives), single or multiple SSDs (Solid State Drives), or a combination thereof is used. Also, part or all of the auxiliary memory 13 may be storage on a network connected via a communication interface (not shown). As the input/output interface 14, for example, a USB (Universal Serial Bus) interface, a short-range communication interface such as infrared rays or Bluetooth (registered trademark), or a combination thereof is used.

For example, an input device 20 and an output device 30 are connected to the input/output interface 14 . As the input device 20, for example, a keyboard, mouse, touch pad, microphone, or a combination thereof is used. A display, a printer, a speaker, or a combination thereof is used as the output device 30, for example. Note that the machine learning device 1 may incorporate a keyboard and touchpad functioning as the input device 20 and a display functioning as the output device 30, like a notebook computer. Further, the machine learning device 1 may incorporate a touch panel functioning as the input device 20 and the output device 30 like a smart phone or a tablet computer.

The auxiliary memory 13 stores a program P for causing the processor 12 to execute a machine learning method S1, which will be described later. The processor 12 expands the program P stored in the auxiliary memory 13 onto the main memory 11, and executes each instruction included in the program P expanded onto the main memory 11, thereby performing a machine learning method S1 described later. Execute each included step. In addition, the auxiliary memory 13 stores various data referred to by the processor 12 in order to execute the machine learning method S1, which will be described later.

Here, a mode in which the processor 12 executes the machine learning method S1, which will be described later, according to the program P stored in the auxiliary memory 13, which is an internal storage medium, has been described, but the present invention is not limited to this. That is, it is also possible to employ a form in which the processor 12 executes the machine learning method S1, which will be described later, according to the program P stored in the external recording medium. In this case, as the external recording medium, a computer-readable "non-transitory tangible medium" such as a tape, disk, card, semiconductor memory, or programmable logic circuit can be used. Alternatively, a mode may be employed in which the processor 12 implements a machine learning method S1 described later according to a program P obtained from a network connected via a communication interface (not shown). In this case, the network may be, for example, the Internet, a wired LAN (Local Area Network), a wireless LAN, or a combination of at least some of them.

Also, here, a mode of realizing the machine learning device 1 using a single computer has been described, but the present invention is not limited to this. That is, a form in which the machine learning device 1 is realized using a plurality of computers configured to be able to communicate with each other may be adopted. In this case, each step constituting the machine learning method S1, which will be described later, can be executed in parallel by these computers.

(Functional configuration of machine learning device)
A functional configuration of the machine learning device 1 will be described with reference to FIG. FIG. 2 is a block diagram showing the functional configuration of the machine learning device 1. As shown in FIG.

The machine learning device 1 includes a classifier group 101, an evaluation unit 102, and a learning control unit 103, as shown in FIG. These blocks are functional blocks implemented by executing the instructions of the program P described above by the processor 12 described above.

A classifier group 101 is a set of a plurality of classifiers C1, C2, . . . , Cn. Here, n is an arbitrary natural number representing the number of classifiers C1, C2, . . . , Cn. Each classifier Ci (i=1, 2, . . . , n) refers to an image containing an article as a subject and performs classification processing for classifying the state of the article using an algorithm determined by machine learning (hereinafter referred to as "model ”). A model that can be used for the classification process is, for example, a neural network that receives an image including an article as an object and outputs the state of the article. In this embodiment, the first convolution layer (eg conv32 3*3), the first polling layer (eg MaxPooling(2,2)), the second convolution layer (eg conv32 3*3), the 2 polling layers (e.g. MaxPooling(2,2)), ..., 6th polling layers (e.g. MaxPooling(2,2)), 6th convolution layers (e.g. conv32 3*3), 7th polling layers ( For example, a CNN (Convolutional Neural Network) in which MaxPooling(2,2)), a seventh convolution layer (eg GlobalAveragePooling), and a softmax layer (eg softmax) are connected in this order is used for classification processing.

The model used by each classifier Ci is constructed in the initial learning phase before the classifier group 101 starts operating, and is optimized in the additional learning phase after the classifier group 101 starts operating. be. In the initial learning phase, (1) model selection, (2) model tuning (network structure and hyperparameter tuning), and (3) model machine learning (hereinafter referred to as “initial learning”) are performed. done. In the additional learning phase, (1) model machine learning (hereinafter referred to as “additional learning”) is performed. In this embodiment, since CNN is used as the model, machine learning of the model in the initial learning phase and the additional learning phase is realized by deep learning.

In this embodiment, the structure of the model used by each classifier Ci is common. In addition, in this embodiment, the training data of the model used by each classifier Ci is common. Even in this case, the models used by each classifier Ci are mutually independent models with different weighting factors. This is because random numbers are used for learning the model used by each classifier Ci. Therefore, the generalization ability of each classifier Ci may differ from each other. Hereinafter, among the classifiers C1, C2, . do. Further, among the classifiers C1, C2, . Here, for convenience of explanation, it is assumed that classifier C1 is the primary classifier and classifiers C2, C3, . . . , Cn are secondary classifiers. During operation of the classifier group 101, the classification result by the main classifier C1 is referred to as the classification result by the classifier group 101. FIG. The classification results by the classifiers C2, C3, .

Note that articles to be classified are divided into a plurality of groups in advance. The classifier group 101 performs a classification process for classifying the states of articles for each group. That is, for example, the classifier group 101 finishes classifying each item belonging to the first group, starts classifying each item belonging to the second group, and classifies each item belonging to the second group. After finishing the processing, the classification processing of the articles belonging to the third group is started.

The evaluation unit 102 is a block that executes evaluation processing for evaluating the classification accuracy of each classifier Ci included in the classifier group 101 for each group. In the present embodiment, the evaluation unit 102 executes evaluation processing according to the first evaluation method in the first half of the additional learning phase (hereinafter referred to as “first additional learning phase”), and performs the evaluation process in the second half of the additional learning phase. In (hereinafter referred to as "second additional learning phase"), evaluation processing is executed according to the second evaluation method. The end timing of the first additional learning phase (the start timing of the second additional learning phase) and the end timing of the second additional learning phase are arbitrary. The second additional learning phase may continue until the operation of the classifier group 101 ends.

The first evaluation method is to evaluate the classification accuracy of each classifier Ci, for each item selected from the group, the result of classifying the state of the item by the classifier Ci and the result of classifying the state of the item by a person. It is a method of comparing with Here, the items selected from the group may be all or part of the items belonging to the group. In this case, for example, among the items selected from the group, the number of items whose state classification result by the classifier Ci matches the state classification result by the person is divided by the number of items selected from the group. The quotient (correct answer rate of the classifier Ci when the result of classification of the state by a person is regarded as correct) is the classification accuracy of the classifier Ci.

The second evaluation method is based on the classification accuracy of each classifier Ci, for each item selected from the group, the result of classifying the state of the item by the classifier Ci, and the state of the item by the main classifier C1. This is a method of comparing the results of classification. Here, the items selected from the group may be all or part of the items belonging to the group. In this case, for example, among the items selected from the group, the number of items for which the result of classifying the state by the classifier Ci and the result of classifying the state by the main classifier C1 are the same as the number of items selected from the group The quotient obtained by dividing by the number (correct answer rate of the classifier Ci when the result of classification by the main classifier C1 is regarded as correct) is the classification accuracy of the classifier Ci.

When the classification accuracy of at least one classifier Ci included in the classifier group 101 does not reach a predetermined standard, the learning control unit 103 executes the classification process for the articles belonging to the next group of the classifier group 101. This block executes learning control processing for causing at least one classifier Cj (j≠i) other than the classifier Ci included in the classifier group 101 to perform additional learning. In this embodiment, when the classification accuracy of at least one classifier Ci included in the classifier group 101 does not reach a predetermined standard, the learning control unit 103 causes the classifier group 101 to Before executing the classification process, all the classifiers C1, C2, . Whether or not the classification accuracy of each classifier Ci has reached a predetermined standard is determined by, for example, comparing the classification accuracy of the classifier Ci evaluated by the evaluation unit 102 with a predetermined threshold value. Realized.

Here, making the classifier Ci perform additional learning means that the weight coefficients of the model used by the classifier Ci for classification processing are combined with teacher data (an image containing an article as a subject and the result of human classification of the state of the article). A set of pairs of ) to optimize using. Note that when the classifier Ci misclassifies an article that should be classified as state A as state B, the training data used in the additional learning consists of an image containing the article that should be classified as state A as a subject and the image of the article. At least one of a pair of a result of human classification of the state (state A) and a pair of an image including an article to be classified as state B as a subject and a result of human classification of the state of the article (state B). (preferably both). This is because it is possible to efficiently improve the generalization ability of the classifier Ci.

(Effect of machine learning device)
When implementing a technique for classifying the state of an article by referring to an image containing the article as a subject using a classifier constructed by machine learning, it is difficult to maintain classification accuracy over a long period of time. This is because the generalization ability of the classifier constructed by machine learning is demonstrated only for items included in the images used as training data, or items similar to those items, which are the targets of classification. This is because if the environment surrounding the article itself to be classified or the environment surrounding the article to be classified changes, the article to be classified will exceed the range in which the generalization ability of the classifier can be demonstrated. As an example, a classifier used at a manufacturing site, for example, a classifier that refers to an image containing a product as a subject and classifies the state of the product, may include manufacturing conditions (persons, machines, raw materials, manufacturing methods, etc.) Such problems can arise frequently, as changes may occur in

On the other hand, in the machine learning device 1, when the classification accuracy of at least one classifier Ci included in the classifier group 101 falls below the threshold, the classification accuracy of the main classifier C1 due to changes in manufacturing conditions Considering this as a "prediction" of a decrease in accuracy, additional learning is performed for all classifiers C1, C2, . . . , Cn including the main classifier C1. Therefore, the main classifier C1 can be made to perform additional learning before the classification accuracy of the main classifier C1 decreases, and the classification accuracy of the main classifier C1 can be maintained at a high level. In addition, when the main classifier C1 is made to perform additional learning, the sub-classifiers C2, C3, . . . This is for maintaining the detection accuracy with which Cn detects the above-described "prediction".

(Machine learning method in the first additional learning phase)
The flow of the machine learning method S1 performed in the first additional learning phase will be described with reference to FIG. 3(a). FIG. 3(a) is a flow chart showing the flow of the machine learning method S1 in the first additional learning phase.

The machine learning method S1 is a process performed before starting classification of articles belonging to the group next to the target group, with a group whose state of articles has been classified as a target group. As shown, it includes an evaluation step S101, a determination step S102, and an additional learning step S103.

Evaluation step S101 is a step in which the evaluation unit 102 executes evaluation processing for evaluating the classification accuracy of each classifier Ci included in the classifier group 101 according to the first evaluation method. Here, the first evaluation method is, as described above, the classification accuracy of each classifier Ci, for each article selected from the target group, the result of classifying the state of the article by the classifier Ci, This is a method of comparing the result of classification of the state of the article by a person. In the evaluation step S101, for example, among the items selected from the target group, the number of items whose state classification result by the classifier Ci matches the state classification result by the person is selected from the target group. The quotient obtained by dividing by the number of articles (correct answer rate of the classifier Ci when the result of classifying the state by a person is regarded as correct) is calculated as the classification accuracy of the classifier Ci.

The determination step S102 is a step of determining whether or not there is a classifier Ci whose classification accuracy evaluated in the evaluation step S101 is below a predetermined threshold value. If there is a classifier Ci whose classification accuracy is below the threshold, an additional learning step S103, which will be described later, is executed. Conversely, if there is no classifier Ci whose classification accuracy is below the threshold, the additional learning step S103, which will be described later, is not performed.

Additional learning step S103 is a step in which the learning control unit 103 causes all the classifiers C1, C2, . . . , Cn included in the classifier group 101 to perform additional learning. In the additional learning step S103, the weighting coefficient of the model used by each classifier Ci for classification processing is optimized using teacher data.

As described above, in the first additional learning phase, the classification accuracy of each classifier Ci is evaluated based on the result of human classification of the state of the article. Therefore, the timing at which additional learning should be performed can be appropriately specified without depending on the generalization ability of the main classifier C1.

(Machine learning method in the second additional learning phase)
The flow of the machine learning method S2 performed in the second additional learning phase will be described with reference to FIG. 3(b). (b) of FIG. 3 is a flow chart showing the flow of the machine learning method S2 performed in the second additional learning phase.

The machine learning method S2 is a process performed before starting classification of articles belonging to the group next to the target group, with the group in which the state of the articles has been classified as the target group. As shown, it includes an evaluation step S201, a determination step S202, and an additional learning step S203.

Evaluation step S201 is a step in which the evaluation unit 102 executes evaluation processing for evaluating the classification accuracy of each classifier Ci included in the classifier group 101 according to the second evaluation method. Here, the second evaluation method is, as described above, the classification accuracy of each classifier Ci, for each article selected from the target group, the result of classifying the state of the article by the classifier Ci, This is a method of comparing the result of classifying the state of the article by the main classifier C1. In the evaluation step S201, for example, among the items selected from the target group, the number of items for which the result of classifying the state by the classifier Ci and the result of classifying the state by the main classifier C1 are the same is determined from the target group. The quotient obtained by dividing by the number of selected articles (correct answer rate of the classifier Ci when the result of classification by the main classifier C1 is regarded as correct) is calculated as the classification accuracy of the classifier Ci.

The determination step S202 is a step of determining whether or not there is a classifier Ci for which the classification accuracy evaluated in the evaluation step S201 is below a predetermined threshold value. If there is a classifier Ci whose classification accuracy is below the threshold, an additional learning step S203, which will be described later, is performed. Conversely, if there is no classifier Ci whose classification accuracy is below the threshold, the additional learning step S203, which will be described later, is not executed.

Additional learning step S203 is a step in which the learning control unit 103 causes all the classifiers C1, C2, . . . , Cn included in the classifier group 101 to perform additional learning. In the additional learning step S203, the weighting coefficient of the model used by each classifier Ci for classification processing is optimized using teacher data.

As described above, in the second additional learning phase, the classification accuracy of each classifier Ci is evaluated based on the result of classification of the article state by the main classifier C1. Therefore, it is possible to appropriately identify the timing at which additional learning should be performed without the need for a person to classify the state of the article.

(Modification)
In this embodiment, after starting the operation of the classifier group 101, the first additional learning phase (the classification accuracy of each classifier Ci, the result of classifying the state of the article by the classifier Ci, and the and a second additional learning phase (the classification accuracy of each classifier Ci is evaluated by comparing the result of classifying the state of the article by the classifier Ci with the result of classifying the state of the article by the main classifier C1). (Evaluate by comparing the result of classifying the state of the article), but the present invention is not limited to this. For example, a configuration may be adopted in which only the first additional learning phase is performed after starting operation of the classifier group 101, or only the second additional learning phase is performed after starting operation of the classifier group 101. You may adopt the structure which carries out.

(Modification 2)
In this embodiment, the classifier group 101 is included inside the machine learning device 1 (executed by the same computer as the machine learning device 1). Not limited. For example, a configuration in which the classifier group 101 is not included inside the machine learning device 1 (executed by a computer different from the machine learning device 1) may be adopted. That is, the classifier group 101 is not an essential component of the machine learning device 1 .

[Second embodiment]
(Functional configuration of inspection device 2)
A functional configuration of an inspection apparatus 2 according to a second embodiment of the present invention will be described with reference to FIG. FIG. 4 is a block diagram showing the functional configuration of the inspection device 2. As shown in FIG. The physical configuration of the inspection device 2 is the same as the physical configuration of the machine learning device 1 shown in FIG. Therefore, a detailed description of the physical configuration of the inspection device 2 will be omitted.

The inspection device 2 is a device for inspecting a plurality of articles divided into groups (for example, lots), and includes a machine learning device 1 as shown in FIG. That is, the inspection device 2 includes a classifier group 101 , an evaluation section 102 and a learning control section 103 . The plurality of articles to be inspected may be divided into groups in advance, or the articles inspected in each of a plurality of predetermined time periods may be grouped (for example, the first group). Items inspected in a time slot are in a first group, items inspected in a second time slot following the first time slot are in a second group, etc.).

The classifier group 101 consists of a plurality of classifiers C1, C2, . . The evaluation unit 102 executes evaluation processing for evaluating the classification accuracy of each classifier Ci included in the classifier group 101 for each group. If the classification accuracy of at least one classifier Ci included in the classifier group 101 does not reach a predetermined standard, the learning control unit 103 controls the classifier group 101 before inspecting articles belonging to the next group. Additional learning (or re-learning) is performed on the classifiers Cj other than the at least one classifier Ci included in . For example, if the classification accuracy of at least one classifier Ci included in the classifier group 101 does not reach a predetermined standard, all the items included in the classifier group 101 are additional learning (or re-learning) to classifiers C1, C2, . . . , Cn. The details of the functions of the classifier group 101, the evaluation unit 102, and the learning control unit 103 are as described in the first embodiment. The flow of the machine learning method in the inspection device 2 is the same as the flow of the machine learning methods S1 and S2 in the machine learning device 1 described with reference to FIG.

(Example 1)
An embodiment of the inspection device 2 will be described with reference to FIGS. 5 and 6. FIG.

FIG. 5 is a schematic diagram showing the flow of the manufacturing process of the semiconductor laser.

As shown in the upper part of FIG. 5, the manufacturing process of a semiconductor laser consists of a crystal growth process, an electrode process, and a device forming process. In a semiconductor laser manufacturing process, a wafer formed in a crystal growth process is visually inspected.

As shown in the lower part of FIG. 5, the appearance inspection of the wafer consists of (1) an imaging process for imaging a part of the wafer, and (2) a partial image corresponding to each LD chip from the image obtained in the imaging process. and (3) a classification step of classifying the state of the LD chip based on the partial image (image including the LD chip as an object) obtained in the cutting step.

The inspection apparatus 2 according to this embodiment performs this classification process using a classifier group 101 consisting of three classifiers C1, C2, and C3. Each classifier Ci classifies the state of the LD chip into one of the following 7 states based on the partial image including the LD chip as a subject. An LD chip whose state is classified into any of states 1-3 is regarded as a non-defective product, and an LD chip whose state is classified as any of states 4-7 is regarded as a defective product.

State 1: no defects,
Condition 2: There is a thin point defect inside the stripes formed on the chip.
State 3: there is a defect on the stripe,
State 4: There is a dark point defect inside the stripe,
State 5: there is a defect other than a point defect inside the stripe,
State 6: there is some defect outside the stripe,
State 7: Particulate foreign matter is present on the chip.

FIG. 6 is a box and whisker diagram showing the classification accuracy of each classifier Ci included in the classifier group 101 for each group. In this example, in the inspection for group 5, the classification accuracy of classifier C2 is below the threshold of 98%, and in the inspection for group 9, the classification accuracy of classifier C3 is below the threshold of 98%. For this reason, in this embodiment, additional training of the three classifiers C1, C2, C3 is performed before the examination for Groups 6 and 10 is started. In addition, in FIG. 6, the fact that the classification accuracy of at least one classifier Ci included in the classifier group 101 falls below the threshold is expressed as a "prediction". In addition, in FIG. 6, performing additional learning of the three classifiers C1, C2, and C3 included in the classifier group 101 is expressed as "movement".

[Additional notes]
The present invention is not limited to the above-described embodiments, but can be modified in various ways within the scope of the claims, and can be obtained by appropriately combining technical means disclosed in different embodiments. is also included in the technical scope of the present invention.

1 machine learning device 2 inspection device 10 bus 11 main memory 12 processor 13 auxiliary memory 14 input/output interface 20 input device 30 output device 101 classifier group 102 evaluation unit 103 learning control unit C1, C2, C3, Ci, Cj classifier

Claims (9)

A machine learning device for performing machine learning on a classifier group consisting of a plurality of classifiers for classifying the state of an article belonging to each of a plurality of groups with reference to an image containing the article as a subject,
an evaluation unit that evaluates the classification accuracy of each classifier included in the classifier group for each group;
If the classification accuracy of at least one classifier included in the classifier group does not reach a predetermined standard, before the classifier group performs classification processing for articles belonging to the next group, the classifier group a learning control unit that causes a classifier other than the at least one classifier to perform additional learning,
A machine learning device characterized by: The evaluation unit determines the classification accuracy of each classifier included in the classifier group, the classification result obtained by classifying the state of each article by the classifier, and the classification accuracy of each classifier included in the classifier group. evaluated by comparing the classification results obtained by
2. The machine learning device according to claim 1, wherein: Among the plurality of classifiers, the classifier with the highest classification accuracy achieved in the initial learning phase is used as the main classifier,
The evaluation unit determines the classification accuracy of each classifier included in the classifier group, the classification result obtained by classifying the state of each article by the classifier, and the classification result obtained by classifying the state of each article by the main classifier. Evaluate by comparing the classification result obtained by classifying,
2. The machine learning device according to claim 1, wherein: Among the plurality of classifiers, the classifier with the highest classification accuracy achieved in the initial learning phase is used as the main classifier,
In the first additional learning phase,
The evaluation unit determines the classification accuracy of each classifier included in the classifier group, the classification result obtained by classifying the state of each article by the classifier, and the classification accuracy of each classifier included in the classifier group. evaluated by comparing the classification results obtained by
In a second additional learning phase following the first additional learning phase,
The evaluation unit determines the classification accuracy of each classifier included in the classifier group, the classification result obtained by classifying the state of each article by the classifier, and the classification result obtained by classifying the state of each article by the main classifier. Evaluate by comparing the classification result obtained by classifying,
2. The machine learning device according to claim 1, wherein: When the classification accuracy of classifiers other than the main classifier included in the classifier does not reach a predetermined standard, the learning control unit performs classification processing for articles belonging to the next group of the classifier group. before performing additional learning on the main classifier,
5. The machine learning device according to claim 3, wherein: When the classification accuracy of at least one classifier included in the classifier group does not reach a predetermined standard, the learning control unit performs classification processing on articles belonging to the next group of the classifier group. , causes all classifiers included in the classifier group to perform additional learning,
6. The machine learning device according to any one of claims 1 to 5, characterized by: A machine learning method for performing machine learning on a classifier group consisting of a plurality of classifiers for classifying the state of an article belonging to each of a plurality of groups with reference to an image containing the article as a subject, the method comprising:
an evaluation step of evaluating the classification accuracy of each classifier included in the classifier group for each group;
If the classification accuracy of at least one classifier included in the classifier group does not reach a predetermined standard, before the classifier group performs classification processing for articles belonging to the next group, the classifier group and a learning control step of causing a classifier other than the at least one classifier to perform additional learning,
A machine learning method characterized by: A machine learning program that causes a computer to operate as the machine learning device according to any one of claims 1 to 6, wherein the computer functions as each part of the machine learning device. Equipped with the machine learning device according to any one of claims 1 to 6 and the classifier group, and inspecting a plurality of items divided into lots using the classifier group,
An inspection device characterized by:

Images