WO2020090290A1 - Image classifying device, image inspecting device, and image classifying method - Google Patents

Abstract

The objective of the present invention is to suppress over-classification and under-classification of images into an intermediate class. This image classifying device is provided with: a first classifier for classifying an inspection image into a first class or into another class, where images classified into the first class include those which possibly may be in the first class; a second classifier for classifying the inspection image into a second class or into another class, where images classified into the second class include those which possibly may be in the second class; and a determining unit for determining, on the basis of the result of the classification by the first classifier and the result of the classification by the second classifier, whether the inspection image is in an intermediate class between the first class and the second class.

Description

Image classification device, image inspection device, and image classification method

The present invention relates to an image classification device, an image inspection device, and an image classification method.

Conventionally, as an inspection / classification device of an image classification device, a defect detection unit that performs a defect inspection of image data of an inspection target area and generates defect image data when a defect is detected, and a defect image using a classifier. There is known a device including a classification control unit that classifies into any one of a plurality of defect classes (see Patent Document 1). The classifier of this inspection / classification apparatus includes M (M is an integer of 2 or more) core classifiers and one voting unit, and classifies into N (N is an integer of 2 or more) defect classes. is doing.

JP, 2016-40650, A

Here, for example, an image may be neither defective nor defective, that is, an intermediate class (gray class) may exist between a defective class and a non-defective class.

In the situation where this intermediate class exists, for example, when classifying a target image into a first class with a defect, a second class without a defect, and an intermediate class into a third class, It becomes necessary to set two thresholds, a threshold between the third class and a threshold between the first class and the third class.

However, since the boundary of the intermediate class is not clear, it is difficult to adjust these two threshold values. Depending on the two threshold values, too many images are classified into the third class (excessive classification), Alternatively, the number of images classified into the third class may be too small (underclassified).

Therefore, an object of the present invention is to provide an image classification device and an image classification method that can suppress excessive classification and underclassification of images into intermediate classes.

An image classification device according to an aspect of the present invention is a first classifier that classifies an image into a first class or another class, and an image classified into the first class may have the first class possibility. A second classifier that classifies an image into a second class or another class, including a certain class of The image is an intermediate class between the first class and the second class based on the second classifier, which includes the And a determination unit that determines whether or not there is.

According to this aspect, it is determined whether or not the image is an intermediate class between the first class and the second class based on the classification result of the first classifier and the classification result of the second classifier. .. Here, when the first classifier classifies an image into the first class or another class, if the image classified into the first class by the first classifier includes a possibility of the first class. The threshold value of can be easily set in comparison with the threshold value when the image is classified into the intermediate class between the first class and the second class. Similarly, if the images classified by the second classifier into the second class include the possible second class, the second classifier may classify the image into the second class or another class. The threshold for classifying can be easily set in comparison with the threshold for classifying an image into an intermediate class between the first class and the second class. Then, the inventor of the present invention has found that an image can be classified into an intermediate class based on the classification result of the first classifier and the classification result of the second classifier. Therefore, it is determined whether the image is an intermediate class between the first class and the second class based on the classification result of the first classifier and the classification result of the second classifier. It is possible to suppress over-classification and under-classification into intermediate classes.

In the above-described aspect, the determining unit determines that the image is an intermediate class when the image is classified into the first class by the first classifier and classified into the second class by the second classifier. Good.

According to this aspect, when the image is classified into the first class by the first classifier and the second class by the second classifier, the image is determined to be the intermediate class. Here, the first classifier classifies the class as a first class including a possibility of the first class, and the second classifier classifies as a second class including a possibility of the second class. The inventor of the present invention has found that the classified images have a high probability of being an intermediate class between the first class and the second class. Therefore, when the image is classified into the first class by the first classifier and is classified into the second class by the second classifier, it is determined that the image is the intermediate class, and thus the image is classified into the intermediate class. The accuracy of classification can be improved.

In the above-described aspect, the determining unit determines that the image is the first class when the image is classified into the first class by the first classifier and is classified into the other class by the second classifier. You may.

According to this aspect, when the image is classified into the first class by the first classifier and is classified into the other class by the second classifier, the image is determined to be the first class. Here, an image classified by the first classifier as a first class including a possibility of the first class and classified by the second classifier as other than the second class is a first class. The inventor of the present invention has found that there is a high probability. Therefore, when the image is classified into the first class by the first classifier and is classified into the class other than the second class by the second classifier, the image is classified into the first class, and thus the image is classified into the first class. The accuracy of classifying into one class can be improved.

In the above-described aspect, the determining unit determines that the image is the second class when the image is classified into the other class by the first classifier and the second class by the second classifier. You may.

According to this aspect, when the image is classified into the other class by the first classifier and the second class by the second classifier, the image is determined to be the second class. Here, an image classified by the first classifier as a class other than the first class and classified by the second classifier as a second class including a possibility of the second class is a second class. The inventor of the present invention has found that there is a high probability. Therefore, when the image is classified into a class other than the first class by the first classifier and is classified into the second class by the second classifier, the image is classified into the second class, and thus the image is classified into the second class. The accuracy of classifying into two classes can be improved.

In the above-mentioned aspect, each of the first classifier and the second classifier may be a two-class classifier.

According to this aspect, each of the first classifier and the second classifier is a two-class classifier. Accordingly, it is possible to easily and easily generate the first classifier and the second classifier with high classification accuracy as compared with a multi-class classifier having three or more classes.

In the above-described aspect, a third classifier that classifies an image into a third class or another class, and the images classified into the third class include those that are likely to be the third class. The determination unit further includes a third classifier, and the determination unit determines that the image is a first class, a second class, based on the classification result of the first classifier, the classification result of the second classifier, and the classification result of the third classifier. It may be determined whether or not it is an intermediate class between at least two of the third and third classes.

According to this aspect, the image is classified into the first class, the second class, and the third class based on the classification result of the first classifier, the classification result of the second classifier, and the classification result of the third classifier. Is an intermediate class between at least two of the above. Here, when the image classified into the third class by the third classifier includes a possibility of the third class, when the third classifier classifies the image into the third class or another class. Can be easily set in comparison with the threshold when classifying an image into an intermediate class between at least two of the first class, the second class and the third class. it can. Then, based on the classification result of the third classifier and the classification result of the first classifier and the classification result of the second classifier, it is possible to classify an image into an intermediate class. The person found. Therefore, based on the classification result of the first classifier, the classification result of the second classifier, and the classification result of the third classifier, the image has at least two of the first class, the second class, and the third class. By determining whether or not the image is an intermediate class, it is possible to suppress excessive classification and under-classification of an image into intermediate classes.

In the above-mentioned aspect, each further comprises a third classifier to an m-th classifier (m is an integer of 4 or more), the third classifier classifies the image into a third class or another class, and The images classified into the class include those having the possibility of the third class, and the m-th classifier classifies the image into the m-th class or another class, and the image classified into the m-th class is Are included in the m-th class, and the determination unit determines that the image is classified from the first class to the m-th class based on the classification result of the first classifier to the m-th classifier. May be determined to be an intermediate class between at least two of the above.

According to this aspect, whether or not the image is an intermediate class between at least two of the first class to the m-th class based on the classification result of the first classifier to the classification result of the m-th classifier. Is determined. Here, when the image classified into the third class by the third classifier includes a possibility of the third class, when the third classifier classifies the image into the third class or another class. The threshold value of can be easily set in comparison with the threshold value when the image is classified into the intermediate class between at least two of the first class to the m-th class. Similarly, when the m-th classifier classifies an image into the m-th class or any other class, if the m-th classifier includes an m-th class possible image. The threshold value of can be easily set in comparison with the threshold value when the image is classified into the intermediate class between at least two of the first class to the m-th class. Then, the image can be classified into the intermediate class based on the classification result of the third classifier to the classification result of the m-th classifier, and the classification result of the first classifier and the classification result of the second classifier. The inventor of the present invention has found that Therefore, it is determined whether or not the image is an intermediate class between at least two of the first class to the m-th class based on the classification result of the first classifier to the classification result of the m-th classifier. With this, it is possible to suppress excessive classification and underclassification of the image into the intermediate class.

An image inspection apparatus according to another aspect of the present invention includes the above-described image classification device, and inspects the image of the inspection object using the image classification device.

According to this aspect, the image of the inspection object is inspected using the image classification device described above. Accordingly, it is possible to easily realize an image inspection apparatus that suppresses over-classification and under-classification of images into intermediate classes.

An image classification method according to another aspect of the present invention is a first classification step in which a first classifier classifies an image into a first class or another class, and the image is classified into the first class. Is a first classification step that includes the possible first classes and a second classification step in which the second classifier classifies the image into a second class or another class, the second class The images classified into are included in the images that have the possibility of the second class. The second classification step and the determination unit are based on the classification result of the first classification step and the classification result of the second classification step. , A determination step of determining whether or not the image is an intermediate class between the first class and the second class.

According to this aspect, it is determined whether the image is an intermediate class between the first class and the second class based on the classification result of the first classification step and the classification result of the second classification step. .. Here, when the first classifier classifies an image into the first class or another class, if the image classified into the first class by the first classifier includes a possibility of the first class. The threshold value of can be easily set in comparison with the threshold value when the image is classified into the intermediate class between the first class and the second class. Similarly, if the images classified by the second classifier into the second class include the possible second class, the second classifier may classify the image into the second class or another class. The threshold for classifying can be easily set in comparison with the threshold for classifying an image into an intermediate class between the first class and the second class. Then, the inventor of the present invention has found that an image can be classified into an intermediate class based on the classification result of the first classifier and the classification result of the second classifier. Therefore, it is determined whether the image is an intermediate class between the first class and the second class based on the classification result of the first classification step and the classification result of the second classification step. It is possible to suppress over-classification and under-classification into intermediate classes.

In the above-described aspect, the determining step determines that the image is the intermediate class when the image is classified into the first class in the first classifying step and the second class in the second classifying step. The determination may be included.

According to this aspect, when the image is classified into the first class in the first classification step and classified into the second class in the second classification step, the image is determined to be the intermediate class. Here, the first classifier classifies the class as a first class including a possibility of the first class, and the second classifier classifies as a second class including a possibility of the second class. The inventor of the present invention has found that the classified images have a high probability of being an intermediate class between the first class and the second class. Therefore, when the image is classified into the first class in the first classifying step and is classified into the second class in the second classifying step, the image is classified into the intermediate class by determining that the image is the intermediate class. The accuracy of classification can be improved.

In the above-described aspect, in the determining step, when the image is classified into the first class in the first classifying step and into the other class in the second classifying step, the determining unit classifies the image into the first class. It may include determining that there is.

According to this aspect, when the image is classified into the first class in the first classification step and classified into other classes in the second classifier, the image is determined to be the first class. Here, an image classified by the first classifier as a first class including a possibility of the first class and classified by the second classifier as other than the second class is a first class. The inventor of the present invention has found that there is a high probability. Therefore, when the image is classified into the first class in the first classifying step and is classified into a class other than the second class in the second classifying step, the image is classified into the first class, and thus the image is classified into the first class. The accuracy of classifying into one class can be improved.

In the above-described aspect, in the determination step, when the image is classified into the other class in the first classification step and the second class in the second classification step, the determination unit classifies the image into the second class. It may include determining that there is.

According to this aspect, when the image is classified into the other class in the first classification step and the second class in the second classification step, the image is determined to be the second class. Here, an image classified by the first classifier as a class other than the first class and classified by the second classifier as a second class including a possibility of the second class is a second class. The inventor of the present invention has found that there is a high probability. Therefore, when the image is classified into a class other than the first class in the first classification step and classified into the second class in the second classification step, the image is determined to be the second class, and thus the image is classified into the second class. The accuracy of classifying into two classes can be improved.

In the above-mentioned aspect, the third classifying step is a third classifying step of classifying an image into a third class or another class, and the image classified into the third class may have the third class. The determination unit further includes a third classification step that includes an image, and the determination unit includes an image based on the classification result of the first classification step, the classification result of the second classification step, and the classification result of the third classification step. May include determining if it is an intermediate class between at least two of the first class, the second class, and the third class.

According to this aspect, the image is classified into the first class, the second class, and the third class based on the classification result of the first classification step, the classification result of the second classification step, and the classification result of the third classification step. Is an intermediate class between at least two of the above. Here, when the image classified into the third class by the third classifier includes a possibility of the third class, when the third classifier classifies the image into the third class or another class. Can be easily set in comparison with the threshold when classifying an image into an intermediate class between at least two of the first class, the second class and the third class. it can. Then, based on the classification result CR3 of the third classifier and the classification result of the first classifier and the classification result of the second classifier, it is possible to classify the image into the intermediate class. The inventor found out. Therefore, based on the classification result of the first classification step, the classification result of the second classification step, and the classification result of the third classification step, the image has at least two of the first class, the second class, and the third class. By determining whether or not the image is an intermediate class, it is possible to suppress excessive classification and under-classification of an image into intermediate classes.

In the above-mentioned aspect, the method further includes each of a third classification step to an m-th (m is an integer of 4 or more) classification step, wherein the third classifier classifies the image into a third class or another class. , The images classified into the third class include those that are likely to be in the third class, and the m-th classifying step includes the m-th classifier classifying the image into the m-th class or other. The image classified into the m-th class includes an image having a possibility of being in the m-th class. It includes determining whether the image is an intermediate class between at least two of the first class to the m-th class based on the classification result of the m classification step.

According to this aspect, whether the image is an intermediate class between at least two of the first class to the m-th class based on the classification result of the first classification step to the classification result of the m-th classification step. Is determined. Here, when the image classified into the third class by the third classifier includes a possibility of the third class, when the third classifier classifies the image into the third class or another class. The threshold value of can be easily set in comparison with the threshold value when the image is classified into the intermediate class between at least two of the first class to the m-th class. Similarly, when the m-th classifier classifies an image into the m-th class or any other class, if the m-th classifier includes an m-th class possible image. The threshold value of can be easily set in comparison with the threshold value when the image is classified into the intermediate class between at least two of the first class to the m-th class. Then, the image can be classified into the intermediate class based on the classification result of the third classifier to the classification result of the m-th classifier, and the classification result of the first classifier and the classification result of the second classifier. The inventor of the present invention has found that Therefore, it is determined whether or not the image is an intermediate class between at least two of the first class to the m-th class based on the classification result of the first classification step to the classification result of the m-th classification step. With this, it is possible to suppress excessive classification and underclassification of the image into the intermediate class.

According to the present invention, it is possible to suppress over-classification and under-classification of images into intermediate classes.

FIG. 1 is a configuration diagram illustrating a schematic configuration of an image processing apparatus according to an embodiment. FIG. 2 is a block diagram illustrating a configuration of the classification unit and the determination unit illustrated in FIG. FIG. 3 is a Venn diagram illustrating an intermediate class between the first class and the second class. FIG. 4 is a table illustrating the results of the determination unit shown in FIG. FIG. 5 is a diagram illustrating the first classifier shown in FIG. 2 and the generation of the first classifier. FIG. 6 is a flow chart illustrating a schematic operation of the image inspection apparatus according to the first embodiment of the present invention. FIG. 7 is a diagram illustrating an inspection image and a classification result of the image inspection apparatus shown in FIG. FIG. 8 is a graph in which the inspection image of the image inspection device shown in FIG. 7 is mapped by the feature amount. FIG. 9 is a configuration diagram illustrating a schematic configuration of the image inspection apparatus according to the second embodiment of the present invention. FIG. 10 is a block diagram illustrating the configuration of the classification unit and the determination unit illustrated in FIG. FIG. 11 is a Venn diagram illustrating an intermediate class between at least two of the first class, the second class, and the third class. FIG. 12 is a table exemplifying the result of the determination unit shown in FIG. FIG. 13 is a block diagram illustrating a modification of the configurations of the classification unit and the determination unit illustrated in FIG.

A preferred embodiment of the present invention will be described with reference to the accompanying drawings. In addition, in each of the drawings, components denoted by the same reference numerals have the same or similar configurations.

<First Embodiment>
First, an example of the configurations of the image inspection apparatus and the image classification apparatus according to the first embodiment of the present invention will be described with reference to FIGS. 1 to 5. FIG. 1 is a configuration diagram illustrating a schematic configuration of an image inspection apparatus 100 according to the first embodiment of the present invention. FIG. 2 is a block diagram illustrating the configuration of the classification unit 31 and the determination unit 32 illustrated in FIG. FIG. 3 is a Venn diagram illustrating an intermediate class IC between the first class CL1 and the second class CL2. FIG. 4 is a table illustrating the result RE of the determination unit 32 shown in FIG. FIG. 5 is a diagram illustrating generation of the first classifier 31a and the second classifier 31b illustrated in FIG.

The image inspection device 100 is for inspecting whether or not the inspection object includes a defect from an image obtained by photographing the inspection object. The defect is not particularly limited, but for example, it can be visually recognized such as a scratch, a crack, a chip, a burr, an adhering substance, a foreign substance, a dent, unevenness in color, dirt, a faint print, a misalignment in print, etc. Including those.

As shown in FIG. 1, the image inspection device 100 includes an image classification device 50, an imaging device 60, an input device 70, an output device 80, and a line device 90. The image inspection device 100 captures an image of the inspection object on the line device 90 by the image capturing device 60, and inspects the image of the inspection object (hereinafter referred to as “inspection image”) using the image classification device 50. Is configured.

The image classification device 50 is, for example, an information processing device such as a computer. The image classification device 50 includes an input / output I / F (interface) 10, a storage unit 20, and a control unit 30. In addition, the image classification device 50 further includes a bus 40 configured to transmit signals and data between the respective units of the image classification device 50.

The input / output I / F 10 is an interface between the image classification device 50 and external devices. The input / output I / F 10 is configured to exchange data and signals with external devices. The input / output I / F 10 is also configured to control communication with an external device. The input / output I / F 10 is configured to include standardized interface connection terminals such as USB (Universal Serial Bus), HDMI (registered trademark) (High-Definition Mutimedia Interface), and IEEE 1394. The input / output I / F 10 is connected to the imaging device 60, the input device 70, and the output device 80.

The storage unit 20 is configured to store programs and data. The storage unit 20 includes, for example, a hard disk drive, a solid state drive, and the like. The storage unit 20 stores various programs executed by the control unit 30 and data necessary for executing the programs in advance. Further, the storage unit 20 stores the inspection image 21 input from the imaging device 60 via the input / output I / F 10.

The control unit 30 is configured to control the operation of each unit of the image classification device 50, such as the input / output I / F 10 and the storage unit 20. Further, the control unit 30 is configured to realize each function described below by executing a program stored in the storage unit 20 or the like. The control unit 30 includes a processor such as a CPU (Central Processing Unit), a memory such as a ROM (Read Only Memory), a RAM (Random Access Memory), and a buffer storage device such as a buffer. The control unit 30 includes, for example, a classification unit 31 and a determination unit 32 as its functional configuration.

As shown in FIG. 2, the classification unit 31 includes a first classifier 31a and a second classifier 31b.

The first classifier 31a is configured to classify the input image into the first class or another class. The images classified into the first class include those that are likely to be in the first class. As a result, an image that is not the first class is definitely classified into the other classes (classes other than the first class). The inspection image 21 stored in the storage unit 20 is input to the first classifier 31a, and the first classifier 31a outputs the classification result CR1.

The second classifier 31b is configured to classify the input image into the second class or another class. The images classified into the second class include those that are likely to be in the second class. As a result, an image that is not the second class is reliably classified into the other classes (classes other than the second class). The inspection image 21 stored in the storage unit 20 is input to the second classifier 31b, and the second classifier 31b outputs the classification result CR2.

Based on the classification result CR1 of the first classifier 31a and the classification result CR2 of the second classifier 31b, the determination unit 32 determines that the inspection image 21 input to the first classifier 31a and the second classifier 31b is It is configured to determine whether the class is an intermediate class between the first class and the second class. The classification result CR1 of the first classifier 31a and the classification result CR2 of the second classifier 31b are input to the determination unit 32, and the determination unit 32 outputs a result RE according to the determination result.

Here, when the inspection image 21 classified into the first class by the first classifier 31a includes a possibility of the first class, the first classifier 31a converts the inspection image 21 into the first class or other cases. Can be easily set in comparison with the threshold when the inspection image 21 is classified into the intermediate class between the first class and the second class. .. Similarly, when the inspection images 21 classified by the second classifier 31b into the second class include those having the possibility of the second class, the second classifier 31b converts the inspection image 21 into the second class. Alternatively, the threshold value when classifying into another class is set easily in comparison with the threshold value when classifying the inspection image 21 into an intermediate class between the first class and the second class. be able to. The inventor of the present invention has found that the inspection image 21 can be classified into the intermediate class based on the classification result CR1 of the first classifier 31a and the classification result CR2 of the second classifier 31b. .. Therefore, it is determined whether the inspection image 21 is an intermediate class between the first class and the second class based on the classification result CR1 of the first classifier 31a and the classification result CR2 of the second classifier 31b. By doing so, it is possible to suppress over-classification and under-classification of the inspection image 21 into intermediate classes.

For example, as shown in FIG. 3, when the first class CL1 is a class whose inspection image 21 is “OK” and the second class CL2 is a class whose inspection image 21 is “NG”, the first class CL1 There is an intermediate class IC between the second class CL2 and the second class CL2. The intermediate class IC includes an inspection image 21 that is “OK” in the first classifier 31a and “NG” in the second classifier 31b, and the inspection image 21 is both “OK” and “NG”. It is a gray image that cannot be classified.

In the following description, unless otherwise specified, as shown in FIG. 3, the inspection image 21 is “OK” for the first class CL1, the inspection image 21 is “NG” for the second class CL2, and the first class CL1. The intermediate class IC existing between the second class CL2 is "gray".

In the case of the example shown in FIG. 3, the determination unit 32 makes a determination according to the table shown in FIG. In the table of FIG. 4, the upper line represents negation, the upper line of “OK” means other than “OK”, and the upper line of “NG” means other than “NG”. In the following description, other than "OK" may be referred to as "? OK", and other than "NG" may be referred to as "? NG".

Specifically, when the inspection image 21 is classified as “OK” by the first classifier 31a and is classified as “NG” by the second classifier 31b, the determination unit 32 sets the inspection image 21 to the intermediate class. It is determined that the IC is IC, and “gray” indicating the intermediate class IC is output as the result RE. Here, the first classifier 31a classifies the class as a first class including a possibility of the first class, and the second classifier 31b classifies a second class including a possibility of the second class. The inventor of the present invention has found that the inspection image 21 classified as a certain one is highly likely to be an intermediate class between the first class and the second class. Therefore, when the inspection image 21 is classified as “OK” by the first classifier 31a and “NG” by the second classifier 31b, the inspection image 21 is determined to be the intermediate class IC. Thereby, the accuracy of classifying the inspection image 21 into “gray” can be improved.

Further, when the inspection image 21 is classified as “OK” by the first classifier 31a and is classified as other than “NG” by the second classifier 31b, the determination unit 32 sets the inspection image 21 to the first class. It is determined to be CL1, and “OK” indicating the first class CL1 is output as the result RE. Here, the inspection image 21 classified by the first classifier 31a as a first class including a possibility of the first class and classified by the second classifier 31b as other than the second class is: The inventor of the present invention has found that the probability of being in the first class is high. Therefore, when the inspection image 21 is classified as “OK” by the first classifier 31a and is classified as other than “NG” by the second classifier 31b, the inspection image 21 is determined to be the first class CL1. By doing so, the accuracy with which the inspection image 21 is classified into “OK” can be improved.

Further, when the inspection image 21 is classified by the first classifier 31a as other than “OK” and is classified as “NG” by the second classifier 31b, the determination unit 32 classifies the inspection image 21 into the second class. It is determined to be CL2, and “NG” indicating the second class CL2 is output as the result RE. Here, the inspection image 21 classified by the first classifier 31a as a class other than the first class and classified by the second classifier 31b as a second class including a possibility of the second class is: The inventor of the present invention has found that the probability of being in the second class is high. Therefore, when the inspection image 21 is classified as other than “OK” by the first classifier 31a and is classified as “NG” by the second classifier 31b, the inspection image 21 is determined to be the second class CL2. By doing so, the accuracy of classifying the inspection image 21 into “NG” can be improved.

When the inspection image 21 is classified by the first classifier 31a other than "OK" and by the second classifier 31b other than "NG", the inspection image 21 is in the example shown in FIG. It belongs to the area outside the first class CL1 and outside the second class CL2. That is, since such an inspection image 21 is an image that cannot be classified by the image classification device 50A, the determination unit 32 outputs "unknown" as the result RE.

The first classifier 31a and the second classifier 31b are two-class classifiers. As a result, it is possible to easily and easily generate the first classifier 31a and the second classifier 31b with high classification accuracy as compared with a multi-class classifier having three or more classes.

The first classifier 31a and the second classifier 31b are generated using an arbitrary machine learning model algorithm. For example, as shown in FIG. 5, the first classifier 31a is generated using the first neural network NW1 and the second classifier 31b is generated using the second neural network NW2. In the example shown in FIG. 5, each of the first neural network NW1 and the second neural network NW2 is a convolutional neural network (CNN: Convolutional Neural Network).

The first neural network NW1 and the second neural network NW2 are, for example, a convolutional filter layer / pooling layer which is a combination of a convolutional filter layer and a pooling layer, respectively, and a feature amount extracted by the convolutional filter layer / pooling layer, It includes a fully connected layer and an output.

A plurality of learning images LG are given as inputs to the first neural network NW1 and the second neural network NW2. An annotation, that is, related information is added to each learning image LG. That is, when each learning image LG is input to the first neural network NW1, the first label LA1 is given to each learning image LG, and each learning image LG is input when inputting each learning image LG to the second neural network NW2. Is attached with the second label LA2. The first label LA1 is “OK”, which is the first class, or “￢OK”, which is the other class. On the other hand, the second label LA2 is “NG” which is the second class or “−NG” which is the other class. The first label LA1 and the second label LA2 are set, for example, by the user viewing and classifying each learning image LG.

In the example shown in FIG. 5, the first classifier 31a and the second classifier 31b are generated using the first neural network NW1 and the second neural network NW2, but the present invention is not limited to this. As the machine learning model algorithm for generating the first classifier 31a and the second classifier 31b, for example, a support vector machine, a logistic regression, a deep neural network, or the like may be used.

Returning to FIG. 1, each function of the control unit 30 can be realized by a program executed by a computer (microprocessor). Therefore, each function of the control unit 30 can be realized by hardware, software, or a combination of hardware and software, and is not limited to either case.

When each function of the control unit 30 is realized by software or a combination of hardware and software, the processing can be executed by multitasking, multithreading, or both multitasking and multithreading. It is not limited to such cases.

The imaging device 60 is configured to capture an image and record it as data. The imaging device 60 is a digital camera, and is configured to include, for example, optical system components such as a lens and electronic system components such as an image sensor (light receiving element). The optical system component may include a plurality of lenses. Further, the electronic component may include a light emitting device such as a flash. The imaging device 60 is arranged above the line device 90, photographs the inspection object on the line device 90, and outputs the photographed image to the image classification device 50 via the input / output I / F 10. The control unit 30 performs necessary processing on the image input from the imaging device 60 to generate a file of the inspection image 21, and causes the storage unit 20 to store the generated file of the inspection image 21.

The input device 70 is configured to be able to input information by a user (user) operation. The input device 70 includes, for example, a keyboard, a keypad, a mouse, a trackball, a touch panel, a microphone, and the like. It is possible to For example, when the user operates a keyboard, a keypad, a mouse, a trackball, a touch panel, a microphone, etc. (including a voice operation using a microphone), the input device 70 outputs data or a signal corresponding to the operation. It outputs to the image classification device 50 via the input / output I / F 10. The control unit 30 can input information to the image classification device 50 by generating data based on this data or signal.

The output device 80 is configured to output information. The output device 80 is configured to include a display device such as a liquid crystal display, an EL (Electro Luminescence) display, and a plasma display. For example, when the output device 80 displays the image data input from the image classification device 50 via the input / output I / F 10 on the display device, the information can be output.

The line device 90 is configured to convey the inspection object. The line device 90 is configured to include a transportation unit such as a belt conveyor. The line device 90 moves, stops, or excludes the inspection object on the line device 90, for example, based on a control signal input from the image classification device 50 via the input / output I / F 10. It will be possible.

Next, the operation of the image inspection apparatus 100 according to the first embodiment of the present invention will be described with reference to FIGS. 6 to 8. FIG. 6 is a flowchart illustrating a schematic operation of the image inspection apparatus 100 according to the first embodiment of the present invention. FIG. 7 is a diagram illustrating the inspection image 21 and the classification result of the image inspection apparatus 100 shown in FIG. FIG. 8 is a graph in which the inspection image 21 of the image inspection apparatus 100 shown in FIG. 7 is mapped by the feature amount.

The image inspection device 100 executes the image inspection process S200 shown in FIG. 6 when activated by, for example, a user's operation.

First, the first classifier 31a reads the inspection image 21 stored in the storage unit 20 and classifies the inspection image 21 into “OK” or other than “OK” (S201). Note that step S201 corresponds to the "first classification step" of the present invention.

Next, the second classifier 31b classifies the inspection image 21 read in step S201 into “NG” or other than “NG” (S202). Note that step S202 corresponds to the "second classification step" of the present invention.

Next, the determination unit 32 determines whether the inspection image 21 read in step S201 is the intermediate class “gray” based on the classification result in step S201 and the classification result in step S202 (S203). ). Note that step S203 corresponds to the "determination step" of the present invention.

If the inspection image 21 is “gray” as a result of the determination in step S203, it is considered necessary to re-inspect the inspection image 21. Therefore, for example, the control unit 30 displays the inspection image 21 on the output device 80 via the I / F 10 and re-inspects the inspection image 21 by the visual inspection of the user (S204). The user sees the inspection image 21 displayed on the output device 80 and operates the input device 70 to input “OK” or “NG” to the inspection image 21.

On the other hand, when the inspection image 21 is not “gray” as a result of the determination in step S203, the determination unit 32 determines whether the inspection image 21 is “OK” which is the first class (S205).

When the inspection image 21 is “OK” as a result of the determination in step S205, it is considered that the inspection target object corresponding to the inspection image 21 is ready to be shipped as a product, for example. Therefore, the control unit 30 outputs a control signal to the line device 90 via the I / F 10, and the line device 90 conveys and ships the inspection object corresponding to the inspection image 21 (S206).

On the other hand, when the inspection image 21 is not "OK" as a result of the determination in step S205, the determination unit 32 determines whether the inspection image 21 is the second class "NG" (S207).

If the inspection image 21 is “NG” as a result of the determination in step S207, it is considered that the inspection target corresponding to the inspection image 21 contains defects, for example, and cannot be shipped as a product. Therefore, the control unit 30 outputs a control signal to the line device 90 via the I / F 10, and the line device 90 excludes the inspection object corresponding to the inspection image 21 from the line (S208).

On the other hand, as a result of the determination in step S207, when the inspection image 21 is not “NG”, it is considered that the determination unit 32 cannot classify the image by the image classification device 50 like “unknown” shown in FIG. Therefore, the control unit 30 outputs a warning together with the inspection image 21 to the output device 80 via the I / F 10 (S209), for example. The user sees the inspection image and the warning output to the output device 80, and stops the line device 90 or operates the input device 70 to give “OK” or “NG” to the inspection image 21. Or type.

After step 204, step 206, step 208, or step 209, the control unit 30 repeats steps S201 to S209 until the image inspection apparatus 100 stops, for example.

Here, No. in FIG. 1 to No. When each inspection image 21 includes defects such as stains and stains like the inspection image 21 of No. 5, the first classifier 31a determines that the No. 1 to No. The inspection image 21 of No. 5 is classified into “OK” and No. 4 to No. The inspection image 21 of No. 5 is classified into “￢OK” (other than “OK”). On the other hand, the second classifier 31b is No. 1 to No. The inspection image 21 of No. 2 is classified as “NG” and No. 3 to No. The inspection image 21 of No. 5 is classified into “￢NG” (other than “NG”).

Also, in FIG. 6 to No. When each inspection image 21 includes a defect such as a crack or a scratch like the inspection image 21 of No. 10, the first classifier 31a determines that the No. 6 to No. The inspection image 21 of No. 7 is classified as “OK” and No. 8 to No. The inspection images 21 of 10 are classified into “￢OK” (other than “OK”). On the other hand, the second classifier 31b is No. The inspection image 21 of No. 6 is classified as “NG” and No. 7 to No. The inspection images 21 of 10 are classified into “￢NG” (other than “NG”).

As described above, when each inspection image 21 includes a defect, there are various types of defects, and each inspection image 21 including a defect may have different characteristics.

No. shown in FIG. 1 to No. When the defects included in the ten inspection images 21 are mapped by, for example, the feature of "area" and the feature of "length", the result is as shown in FIG. The determination unit 32 determines No. based on the classification result of the first classifier 31a and the classification result of the second classifier 31b shown in FIG. 1, No. 2 and No. It is determined that the inspection image 21 of No. 6 is the first class CL1 surrounded by the broken line in FIG. Further, the determination unit 32 determines the No. based on the classification result of the first classifier 31a and the classification result of the second classifier 31b shown in FIG. 4, No. 5, and No. 8 to No. It is determined that the inspection image 21 of 10 is the second class CL2 surrounded by the one-dot chain line in FIG. Further, based on the classification result of the first classifier 31a and the classification result of the second classifier 31b shown in FIG. 3 and No. It is determined that the inspection image 21 of No. 7 is an intermediate class IC in which the first class CL1 and the second class CL2 overlap.

As described above, even when the plurality of inspection images 21 have different characteristics, the inspection images 21 are classified into the first class CL1 and the second class CL1 based on the classification result of the first classifier 31a and the classification result of the second classifier 31b. It can be determined whether or not it is an intermediate class IC with the class CL2.
it can.

<Second Embodiment>
Next, an image inspection apparatus and an image classification apparatus according to the second embodiment of the present invention will be described with reference to FIGS. 9 to 13. FIG. 9 is a configuration diagram illustrating a schematic configuration of an image inspection apparatus 100A according to the second embodiment of the present invention. FIG. 10 is a block diagram illustrating the configuration of the classification unit 31A and the determination unit 32A illustrated in FIG. FIG. 11 is a Venn diagram illustrating an intermediate class IC between at least two of the first class CL1, the second class CL2, and the third class CL3. FIG. 12 is a table exemplifying the result RE of the determination unit 32A shown in FIG. FIG. 13 is a block diagram illustrating a modification of the configurations of the classification unit 31A and the determination unit 32A illustrated in FIG. Note that the same or similar reference numerals are given to the same or similar configurations as those of the first embodiment. The points different from the first embodiment will be described below. In addition, similar operational effects due to the similar configuration will not be sequentially described.

As shown in FIG. 9, in the image inspection apparatus 100A and the image classification apparatus 50A according to the second embodiment of the present invention, the control unit 30 includes a classification unit 31A and a determination unit 32A. The image inspection apparatus 100 and the image classification apparatus 50 according to the present invention are different.

That is, the classification unit 31A further includes a third classifier 31c in addition to the first classifier 31a and the second classifier 31b. In addition to the classification result CR1 and the classification result CR2, the classification result CR3 of the third classifier 31c is input to the determination unit 32A.

The third classifier 31c is configured to classify the input image into the third class or another class. The images classified into the third class include those that may be in the third class. As a result, an image that is not the third class is definitely classified into the other classes (classes other than the third class). The inspection image 21 stored in the storage unit 20 is input to the third classifier 31c, and the third classifier 31c outputs the classification result CR3.

The determination unit 32A, based on the classification result CR1 of the first classifier 31a, the classification result CR2 of the second classifier 31b, and the classification result CR3 of the third classifier 31c, the first classifier 31a and the second classifier 31b. , And the inspection image 21 input to the third classifier 31c is configured to determine whether the inspection image 21 is an intermediate class between at least two of the first class, the second class, and the third class. ing. The determination unit 32A outputs the result RE according to the determination result.

Here, when the inspection image 21 classified into the third class by the third classifier 31c includes an image having a possibility of the third class, the third classifier 31c converts the inspection image 21 into the third class or other cases. The threshold value for classifying the inspection image 21 is compared with the threshold value for classifying the inspection image 21 into an intermediate class between at least two of the first class, the second class, and the third class. And can be easily set. Then, the inspection image 21 can be classified into the intermediate class based on the classification result CR3 of the third classifier 31c, the classification result CR1 of the first classifier 31a, and the classification result CR2 of the second classifier 31b. The inventor of the present invention has found that there is. Therefore, based on the classification result CR1 of the first classifier 31a, the classification result CR2 of the second classifier 31b, and the classification result CR3 of the third classifier 31c, the inspection image 21 includes the first class, the second class, and the By determining whether the inspection image 21 is an intermediate class between at least two of the three classes, it is possible to suppress excessive classification and underclassification of the inspection image 21 into the intermediate classes.

For example, as shown in FIG. 11, the first class CL1 is a class whose inspection image 21 is “A”, the second class CL2 is a class whose inspection image 21 is “B”, and the third class CL3 is When the inspection image 21 is a class of “C”, between the first class CL1 and the second class CL2, between the first class CL1 and the third class CL3, and between the second class CL2 and the third class CL3. Intermediate class ICs exist between the first class CL1, the second class CL2, and the third class CL3. The inspection image 21 included in the intermediate class IC is a gray image that cannot be classified into any of “A”, “B”, and “C”.

In the following description, unless otherwise specified, as shown in FIG. 11, the inspection image 21 is “A” for the first class CL1, the inspection image 21 is “B” for the second class CL2, and the third class CL3 is The intermediate class IC in which the inspection image 21 exists between at least two of “C”, the first class CL1, the second class CL2, and the third class CL3 is “gray”.

In the case of the example shown in FIG. 11, the determination unit 32A makes a determination according to the table shown in FIG. In the table shown in FIG. 12, like the table shown in FIG. 4, the upper line represents negation, the upper line of “A” means other than “A”, and the upper line of “B” is “B”. Means other than, and the upper line of "C" means other than "C". In the following description, other than "A" may be expressed as "￢A", other than "B" may be expressed as "￢B", and other than "C" may be expressed as "￢C".

Specifically, when the inspection image 21 is classified into “A” by the first classifier 31a, classified into “B” by the second classifier 31b, and classified into “C” by the third classifier 31c. The determination unit 32A determines that the inspection image 21 is the intermediate class IC, and outputs “gray” indicating the intermediate class IC as the result RE. Further, when the inspection image 21 is classified into “A” by the first classifier 31a, classified into “B” by the second classifier 31b, and classified into other than “C” by the third classifier 31c, The determination unit 32A determines that the inspection image 21 is the intermediate class IC, and outputs “gray” indicating the intermediate class IC as the result RE. Further, when the inspection image 21 is classified into “A” by the first classifier 31a, classified into other than “B” by the second classifier 31b, and classified into “C” by the third classifier 31c, The determination unit 32A determines that the inspection image 21 is the intermediate class IC, and outputs “gray” indicating the intermediate class IC as the result RE. In addition, when the inspection image 21 is classified by the first classifier 31a other than "A", by the second classifier 31b as "B", and by the third classifier 31c as "C", The determination unit 32A determines that the inspection image 21 is the intermediate class IC, and outputs “gray” indicating the intermediate class IC as the result RE.

On the other hand, the inspection image 21 is classified as "A" by the first classifier 31a, classified as other than "B" by the second classifier 31b, and classified as other than "C" by the third classifier 31c. At this time, the determination unit 32A determines that the inspection image 21 is the first class CL1 and outputs “A” indicating the first class CL1 as the result RE.

When the inspection image 21 is classified by the first classifier 31a other than "A", by the second classifier 31b as "B", and by the third classifier 31c other than "C". The determination unit 32A determines that the inspection image 21 is the second class CL2, and outputs “B” indicating the second class CL2 as the result RE.

Further, when the inspection image 21 is classified by the first classifier 31a other than "A", classified by the second classifier 31b other than "B", and classified by the third classifier 31c as "C". The determining unit 32A determines that the inspection image 21 is the third class CL3, and outputs “C” indicating the third class CL3 as the result RE.

The inspection image 21 was classified by the first classifier 31a as other than "A", by the second classifier 31b as other than "B", and by the third classifier 31c as other than "C". At this time, the inspection image 21 belongs to the area outside the first class CL1, outside the second class CL2, and outside the third class CL3 in the example shown in FIG. That is, since such an inspection image 21 is an image that cannot be classified by the image classification device 50A, the determination unit 32A outputs “unknown” as the result RE.

The third classifier 31c is a two-class classifier, like the first classifier 31a and the second classifier 31b. Further, the third classifier 31c may be generated by using a neural network similarly to the first classifier 31a and the second classifier 31b, or may be generated by a support vector machine, logistic regression, deep neural network, or the like. It may be generated using an algorithm of a machine learning model.

Further, the operation of the image inspection apparatus 100A according to the second embodiment of the present invention is, for example, the general operation of the image inspection apparatus 100 shown in FIG. Only the step of classifying into other classes (corresponding to the third classifying step of the present invention) and the step of determining whether or not the inspection image 21 is the third class are added. Therefore, illustration and description of the operation of the image inspection apparatus 100A are omitted.

In the present embodiment, an example in which the classification unit 31A further includes the third classifier 31c has been shown, but the present invention is not limited to this. For example, as shown in FIG. 13, the classification unit 31A may further include each of the third classifier 31c to the m-th (m is an integer of 4 or more) classifier 31x. In this case, the m-th classifier 31x is configured to classify the input image into the m-th class or another class. The images classified into the m-th class include those that may be in the m-th class. As a result, an image that is not the m-th class is definitely classified into the other classes (classes other than the m-th class). Further, the determination unit 32A determines that the inspection image 21 is between at least two of the first class to the m-th class based on the classification result CR1 of the first classifier 31a to the classification result CRm of the m-th classifier 31x. Determine if it is an intermediate class.

Here, when the images classified into the third class by the third classifier 31c include those having the possibility of the third class, the third classifier 31c converts the inspection image 21 into the third class or other classes. The threshold for classifying the inspection image 21 should be easily set in comparison with the threshold for classifying the inspection image 21 into an intermediate class between at least two of the first class to the m-th class. You can Similarly, when the image classified into the m-th class by the m-th classifier 31x includes a possible m-th class image, the m-th classifier 31x converts the inspection image 21 into the m-th class or another class. The threshold for classifying the inspection image 21 should be easily set in comparison with the threshold for classifying the inspection image 21 into an intermediate class between at least two of the first class to the m-th class. You can Then, based on the classification result CR3 of the third classifier 31c to the classification result CRm of the m-th classifier 31x, the classification result CR1 of the first classifier 31a, and the classification result CR2 of the second classifier 31b. The inventor of the present invention has found that images can be classified into intermediate classes. Therefore, based on the classification result CR1 of the first classifier 31a to the classification result CRm of the mth classifier 31x, is the inspection image 21 an intermediate class between at least two of the first class to the mth class? By determining whether or not the inspection image 21 is over-classified or under-classified into the intermediate class.

Specifically, when the inspection image 21 is classified as the n-th (n is an integer from 1 to m) class in two or more of the first classifier 31a to the m-th classifier 31x, The determination unit 32A determines that the inspection image 21 is the intermediate class.

On the other hand, when the inspection image 21 is classified into the nth class by the nth classifier and is classified into other classes in all the other classifiers, the determination unit 32A determines the inspection image 21 as the first class. It is determined to be n class.

The exemplary embodiments of the present invention have been described above. According to the image classification devices 50 and 50A according to the embodiment of the present invention, the inspection image 21 is classified into the first class based on the classification result CR1 of the first classifier 31a and the classification result CR2 of the second classifier 31b. It is determined whether the class is an intermediate class between the second class and the second class. Here, when the inspection image 21 classified into the first class by the first classifier 31a includes a possibility of the first class, the first classifier 31a converts the inspection image 21 into the first class or other cases. Can be easily set in comparison with the threshold when the inspection image 21 is classified into the intermediate class between the first class and the second class. .. Similarly, when the inspection images 21 classified by the second classifier 31b into the second class include those having the possibility of the second class, the second classifier 31b converts the inspection image 21 into the second class. Alternatively, the threshold value when classifying into another class is set easily in comparison with the threshold value when classifying the inspection image 21 into an intermediate class between the first class and the second class. be able to. The inventor of the present invention has found that the inspection image 21 can be classified into the intermediate class based on the classification result CR1 of the first classifier 31a and the classification result CR2 of the second classifier 31b. .. Therefore, it is determined whether the inspection image 21 is an intermediate class between the first class and the second class based on the classification result CR1 of the first classifier 31a and the classification result CR2 of the second classifier 31b. By doing so, it is possible to suppress over-classification and under-classification of the inspection image 21 into intermediate classes.

Further, according to the image classification method according to the embodiment of the present invention, the inspection image 21 is classified into the intermediate class between the first class and the second class based on the classification result of step S201 and the classification result of step S202. Is determined. Here, when the first classifier classifies an image into the first class or another class, if the image classified into the first class by the first classifier includes a possibility of the first class. The threshold value of can be easily set in comparison with the threshold value when the image is classified into the intermediate class between the first class and the second class. Similarly, if the images classified by the second classifier into the second class include the possible second class, the second classifier may classify the image into the second class or another class. The threshold for classifying can be easily set in comparison with the threshold for classifying an image into an intermediate class between the first class and the second class. Then, the inventor of the present invention has found that an image can be classified into an intermediate class based on the classification result of the first classifier and the classification result of the second classifier. Therefore, based on the classification result of step S201 and the classification result of step S202, it is determined whether the inspection image 21 is an intermediate class between the first class and the second class. It can be overclassified or underclassified into intermediate classes.

The embodiments described above are for facilitating the understanding of the present invention and are not for limiting the interpretation of the present invention. Each element provided in each embodiment and the arrangement, material, condition, shape, size, and the like thereof are not limited to the exemplified ones, but can be appropriately changed. Further, the configurations shown in different embodiments can be partially replaced or combined.

(Appendix)
1. A first classifier (31a) for classifying the inspection image (21) into a first class or a class other than the first class, and the images classified into the first class include those which may be the first class. A first classifier (31a),
A second classifier (31b) for classifying the inspection image (21) into a second class or a class other than the second class, and the images classified into the second class include those which may be the second class. A second classifier (31b),
Based on the classification result (CR1) of the first classifier (31a) and the classification result (CR2) of the second classifier (31b), the inspection image (21) is an intermediate image between the first class and the second class. A determination unit (32) for determining whether the class is a class,
Image classification device (50).
9. A first classifying step in which the first classifier (31a) classifies the inspection image (21) into a first class or another class, and the image classified into the first class has a possibility of being in the first class. A first classification step, which includes
A second classifying step in which the second classifier (31b) classifies the inspection image (21) into the second class or another class, and the image classified into the second class has a possibility of being in the second class. A second classification step, which includes
Based on the classification result (CR1) of the first classification step and the classification result (CR2) of the second classification step, the determination unit (32) determines that the inspection image (21) is between the first class and the second class. A determination step of determining whether or not it is an intermediate class,
Image classification method.

10 ... Input / output I / F, 20 ... Storage part, 21 ... Inspection image, 30 ... Control part, 31, 31A ... Classification part, 31a ... 1st classifier, 31b ... 2nd classifier, 31c ... 3rd classifier , 31x ... mth classifier, 32, 32A ... Judgment unit, 40 ... Bus, 50, 50A ... Image classification device, 60 ... Imaging device, 70 ... Input device, 80 ... Output device, 90 ... Line device, 100, 100A Image inspection apparatus, CL1 ... First class, CL2 ... Second class, CL3 ... Third class, CR1, CR2, CR3, CRm ... Classification result, IC ... Intermediate class, LA1 ... First label, LA2 ... Second label , LG ... Learning image, NW1 ... First neural network, NW2 ... Second neural network, RE ... Result, S200 ... Image inspection processing.

Claims (14)

A first classifier for classifying an image into a first class or another class, wherein the images classified into the first class include those that may be of the first class. When,
A second classifier for classifying the image into a second class or another class, wherein the images classified into the second class include those that are likely to be in the second class. A vessel,
A determination unit that determines whether the image is an intermediate class between the first class and the second class based on the classification result of the first classifier and the classification result of the second classifier. And,
Image classifier. The determination unit determines that the image is the intermediate class when the image is classified into the first class by the first classifier and classified into the second class by the second classifier. To do
The image classification device according to claim 1. The determining unit determines that the image is the first class when the image is classified into the first class by the first classifier and is classified into another class by the second classifier. judge,
The image classification device according to claim 1. The determining unit determines that the image is the second class when the image is classified into another class by the first classifier and is classified into the second class by the second classifier. judge,
The image classification device according to any one of claims 1 to 3. Each of the first classifier and the second classifier is a two-class classifier,
The image classification device according to claim 1. A third classifier for classifying the image into a third class or any other class, wherein the images classified into the third class include those having a possibility of the third class. Further equipped with
The determination unit determines that the image is the first class, the second class, and the second classifier based on the classification result of the first classifier, the classification result of the second classifier, and the classification result of the third classifier. Determining whether it is an intermediate class between at least two of the third classes,
The image classification device according to claim 1. Further comprising each of a third classifier to an m-th classifier (m is an integer of 4 or more),
The third classifier classifies the image into a third class or any other class, and the images classified into the third class include the possible ones of the third class,
The m-th classifier classifies the image into the m-th class or any other class, and the images classified into the m-th class include those that may be in the m-th class.
The determination unit determines that the image is an intermediate class between at least two of the first class to the m-th class based on the classification result of the first classifier to the classification result of the m-th classifier. Determine whether there is,
The image classification device according to claim 1. An image classification device according to any one of claims 1 to 7,
Inspecting the image of the inspection object using the image classification device,
Image inspection device. A first classifying step in which a first classifier classifies an image into a first class or another class, and the images classified into the first class include those that are likely to be in the first class. , The first classification step,
A second classifying step in which a second classifier classifies the image into a second class or another class, wherein the images classified into the second class include those that are likely to be in the second class. The second classification step,
The determination unit determines whether the image is an intermediate class between the first class and the second class based on the classification result of the first classification step and the classification result of the second classification step. A determination step of determining,
Image classification method. In the determining step, when the image is classified into the first class in the first classifying step and classified into the second class in the second classifying step, the judging section classifies the image into the intermediate class. Including determining that
The image classification method according to claim 9. In the determining step, when the image is classified into the first class in the first classifying step and is classified into another class in the second classifying step, the determining section determines the image to be the first class. Including determining that it is a class,
The image classification method according to claim 9. In the determining step, when the image is classified into the other class in the first classifying step and is classified into the second class in the second classifying step, the determining unit determines the image to be the second class. Including determining that it is a class,
The image classification method according to any one of claims 9 to 11. A third classifying step in which a third classifier classifies the image into a third class or another class, and the images classified into the third class include those that are likely to be in the third class. Further comprising a third classification step,
In the determining step, the determining unit determines, based on the classification result of the first classification step, the classification result of the second classification step, and the classification result of the third classification step, that the image is the first class, Determining whether the class is an intermediate class between at least two of the second class and the third class,
The image classification method according to claim 9. Further including each of the third classification step to the m-th (m is an integer of 4 or more) classification step,
The third classifying step includes a third classifier classifying the image into a third class or another class, and images classified into the third class may be in the third class. Stuff included,
The m-th classifying step includes the m-th classifier classifying the image into an m-th class or another class, and an image classified into the m-th class may be in the m-th class. Stuff included,
In the determination step, the determination unit determines that the image is at least two of the first class to the m-th class based on the classification result of the first classification step to the classification result of the m-th classification step. Including determining whether it is an intermediate class between
The image classification method according to claim 9.

Images