Data classification device, data classification method and data classification program

Abstract

A data classification device according to the present invention comprises: a data acquisition unit which acquires data to be classified; a classification unit which classifies the data into one among a plurality of classes by using a trained model trained using a neural network; a similarity calculation unit which calculates the similarity between the data and the trained data used when generating the trained model; a determination unit which determines whether or not the data belongs to the plurality of classes on the basis of the similarity, and determines that the data belongs to an unknown class when the data does not belong to any of the plurality of classes.

Description

[Technical field]

[0001] The present invention relates to a data classification device, a data classification method, and a data classification program. [Background technology]

[0002] There is known a technique for classifying data using a classifier constructed by learning using a convolutional neural network (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2019-96313 A Summary of the Invention [Problem to be solved by the invention]

[0004] The above-mentioned classifier classifies data of an unknown class that cannot be classified into one of the known classes that have been learned. Therefore, when an unknown class may occur in the data, there is a risk that the accuracy of classification may decrease. Therefore, there is a need for a technology that can suppress the classification of data of an unknown class into a known class.

[0005] The present invention has been made in consideration of the above, and aims to provide a data classification device, a data classification method, and a data classification program that can suppress classification of data of an unknown class into a known class. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems and achieve the object, the data classification device of the present invention comprises a data acquisition unit that acquires data to be classified, a classification unit that classifies the data into one of a plurality of classes using a trained model trained using a neural network, a similarity calculation unit that calculates a similarity between the data and trained data used when generating the trained model, and a determination unit that determines whether the data belongs to one of the plurality of classes based on the similarity, and determines that the data belongs to an unknown class if the data does not belong to any of the plurality of classes.

[0007] In a data classification device according to a preferred aspect of the present invention, in the above invention, the trained model has the data as input parameters and the classification result and features of the data as output parameters.

[0008] In a preferred aspect of the present invention, in the data classification device of the above invention, the similarity calculation unit calculates the similarity using features of the data and the learned data, and the determination unit determines the class to which the data belongs in accordance with the classification result of the classification unit if the similarity satisfies a criterion, and determines that the data is of an unknown class if the similarity does not satisfy the criterion.

[0009] A preferred aspect of the present invention is the data classification device according to the above invention, further comprising a control unit that displays, on a display device, at least data that is most similar to the data, together with the determination result by the determination unit.

[0010] A data classification device according to a preferred aspect of the present invention, in the above invention, includes an evaluation value calculation unit that calculates a judgment evaluation value, which is the proportion of judgments made the reference number of times that determine that the data belongs to any of the plurality of classes, when the number of judgments made by the judgment unit reaches a reference number, and a control unit that outputs at least one of an evaluation result and information that prompts reconstruction of the trained model when the judgment evaluation value is smaller than a threshold value.

[0011] In a data classification device according to a preferred aspect of the present invention, in the above invention, the similarity calculation unit calculates the minimum Euclidean distance in a feature space between the features of the data and the features of the trained data for each class as the similarity with each class.

[0012] In a data classification device according to a preferred aspect of the present invention, in the above invention, the similarity calculation unit calculates a predetermined distance in a feature space between the features of the data and the learned data representing each of the multiple classes as the similarity with each class.

[0013] The data classification method of the present invention includes a data acquisition step of acquiring data to be classified, a classification step of reading out a trained model trained using a neural network from a memory unit and classifying the data into one of a plurality of classes using the trained model, a similarity calculation step of calculating a similarity between the data and trained data used when generating the trained model, and a determination step of determining whether the data belongs to one of the plurality of classes based on the similarity, and determining that the data belongs to an unknown class if the data does not belong to any of the plurality of classes.

[0014] The data classification program of the present invention causes a computer to execute the following steps: a data acquisition step of acquiring data to be classified; a classification step of classifying the data into one of a plurality of classes using a trained model trained using a neural network; a similarity calculation step of calculating a similarity between the data and trained data used when generating the trained model; and a determination step of determining whether the data belongs to one of the plurality of classes based on the similarity, and determining that the data belongs to an unknown class if the data does not belong to any of the plurality of classes. Effect of the Invention

[0015] According to the present invention, it is possible to suppress classification of data of an unknown class into a known class. [Brief description of the drawings]

[0016] [Figure 1] FIG. 1 is a block diagram showing a functional configuration of a data classification device according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a diagram showing an example of defect image data stored in the learned data storage unit 92. As shown in FIG. [Diagram 3] FIG. 3 is a flowchart showing an outline of the process performed by the data classification device according to an embodiment of the present invention. [Figure 4] FIG. 4 is a diagram showing a display example (part 1) of the determination result. [Diagram 5] FIG. 5 is a diagram showing a display example (part 2) of the determination result. [Figure 6] FIG. 6 is a diagram illustrating the effect of an embodiment of the present invention. [Figure 7] FIG. 7 is a diagram showing the variation in the detection rate of unknown classes and the false positive rate of known classes when the threshold for discriminating between known classes and unknown classes is changed in the example shown in FIG. [Figure 8] FIG. 8 is a diagram illustrating another method of determining a threshold value that provides a boundary for discriminating between known classes and unknown classes in one embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] Hereinafter, a mode for carrying out the present invention (hereinafter, referred to as an "embodiment") will be described with reference to the accompanying drawings. Note that the present invention is not limited to the embodiment described below.

[0018] Fig. 1 is a block diagram showing the functional configuration of a data classification device according to an embodiment of the present invention. The data classification device 100 shown in the figure is a device that classifies defect image data obtained by capturing images of defects that occur in a manufacturing process of a film using a polymer into one of a number of different classes according to the type of defect. The defects to be classified include defects that occur when raw materials melted during the film manufacturing process are retained inside the manufacturing equipment and degenerated due to excessive temperature, heat, time, etc., foreign matter that is mixed in during the manufacturing process, scratches caused by the manufacturing equipment, etc.

[0019] The following describes the functional configuration of data classification device 100. Data classification device 100 includes an input unit 1, a data acquisition unit 2, an image processing unit 3, a classification unit 4, a similarity calculation unit 5, a judgment unit 6, an evaluation value calculation unit 7, a control unit 8, and a storage unit 9.

[0020] The input unit 1 receives input of various information and is configured using a user interface such as a keyboard, a mouse, and a microphone.

[0021] The data acquisition unit 2 acquires defect image data sent from a defect inspection machine that images the surface of the film, extracts defects, and generates defect image data including the defects, and outputs the data to the memory unit 9 for storage.

[0022] The image processing unit 3 performs known image processing on the data of the defect image, and also performs image processing in which the non-defective parts of the data of the defect image are used as a common background, and the defect parts are superimposed on the background. The background is, for example, a single color of gray.

[0023] The classification unit 4 inputs the defect image data processed by the image processing unit 3 to a trained model trained using a neural network, and classifies the defect image data into one of a plurality of classes. The trained model is a neural network consisting of an input layer, an intermediate layer, and an output layer, each layer having one or a plurality of nodes. The trained model is generated by performing training to optimize the network parameters of the neural network. The network parameters include information on the weights and biases between layers of the neural network. The input parameters of the trained model are the defect image data, and the output parameters are the probability of belonging to each of a plurality of classes and a predetermined feature value in the defect image data. This neural network is a convolutional neural network (CNN) that introduces a Softmax loss function, more preferably an L2 Softmax loss function that adds an L2 norm constraint to the Softmax loss function (for details on the L2 Softmax loss function, see, for example, R. Ranjan et al., “L2-contsrained Softmax Loss for Discriminative Face Verification”, preprint, arXiv:1703.09507 [cs.CV] (https: / / arxiv.org / pdf / 1703.09507.pdf)). A feature is generally defined as a multidimensional vector consisting of multiple components that give the features of image data.

[0024] The similarity calculation unit 5 calculates the similarity between the defect image data and the learned defect image data (hereinafter, also referred to as learned data) used when generating the learned model. The similarity calculation unit 5 calculates the similarity using the feature amount of the defect image data and the feature amount of the learned data. As the similarity, for example, any of Euclidean distance, Mahalanobis distance, cosine similarity, etc. can be applied.

[0025] The determination unit 6 determines whether the defect image data belongs to a plurality of classes based on the similarity, and if the data does not belong to any of the plurality of classes, it determines that the data belongs to an unknown class. If the similarity satisfies a predetermined criterion, the determination unit 6 determines the class to which the defect image data belongs according to the classification result of the classification unit 4, whereas if the similarity does not satisfy the criterion, it determines that the defect image data is of an unknown class.

[0026] When the number of judgments made by the judgment unit 6 reaches a predetermined standard number, the evaluation value calculation unit 7 calculates a judgment evaluation value, which is the proportion of judgments made for the defect image data that belong to one of multiple classes out of the standard number of judgments.

[0027] The control unit 8 controls the overall operation of the data classification device 100. The control unit 8 has a display control unit 81 that displays the judgment result of the judgment unit 6 on a display device. The display control unit 81 outputs data of a defect image most similar to the data of the defect image to the display device and displays it together with the judgment result by the judgment unit 6. Furthermore, when the judgment evaluation value is smaller than a threshold value, the control unit 8 outputs at least one of the evaluation result and information that prompts reconstruction of the trained model. The control unit 8 may output the information to the display device to display it, or may output the information to an audio output device to output it as audio.

[0028] The storage unit 9 includes a trained model storage unit 91 and a trained data storage unit 92. The trained model storage unit 91 stores the trained model described above. The trained data storage unit 92 stores defect image data (hereinafter referred to as trained data) used when training the trained model, and the feature values ​​and class of the trained data. FIG. 2 is a diagram showing an example of trained data stored in the trained data storage unit 92. FIG. 2 illustrates an example in which there are four classes of defect types (defect type 1 to defect type 4).

[0029] The memory unit 9 stores the number of judgments made by the judgment unit 6. When the number of judgments reaches a predetermined reference number, the number is reset under the control of the control unit 8. The memory unit 9 also stores various programs for operating the data classifying device 100, and data including various parameters necessary for the operation of the data classifying device 100. The various programs include the data classification program according to this embodiment. The memory unit 9 is configured using a ROM (Read Only Memory) in which the various programs are installed in advance, and a RAM (Random Access Memory) for storing calculation parameters and data for each process, etc.

[0030] The various programs may be recorded on a computer-readable recording medium such as a hard disk drive (HDD), a solid state drive (SSD), a flash memory, a CD-ROM, a DVD-ROM, or a Blu-ray (registered trademark), or may be downloadable via a communication network. The communication network referred to here is configured using at least one of an existing public line network, a local area network (LAN), and a wide area network (WAN), and may be either wired or wireless.

[0031] Data classification device 100 having the above functional configuration is a computer configured using one or more hardware components such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an ASIC (Application Specific Integrated Circuit), and an FPGA (Field Programmable Gate Array).

[0032] 3 is a flowchart showing an outline of the process performed by the data classification device 100. First, the data acquisition unit 2 acquires defect image data from a defect inspection machine (step S1).

[0033] Next, the image processing unit 3 performs image processing (step S2). At this time, the image processing unit 3 performs rule-based image processing. The purpose here is to remove unnecessary areas other than the defect part in the image, and to eliminate images in which it is clear that no defect is captured. In the classification process from step S3 onwards, all images may be used, or the images to be classified may be selected in the image processing of step S2. Here, a process is performed to cut out a specific size with the defect part at the centre, and a process is performed to make the background part other than the defect a monochrome background such as grey, and the defect part is superimposed on this background.

[0034] Thereafter, the classification unit 4 performs classification using the trained model stored in the trained model storage unit 91 (step S3). Specifically, the classification unit 4 inputs the defect image data processed by the image processing unit 3 to the trained model, and acquires the probability of belonging to each of a plurality of classes and the feature amount of the input defect image data as output parameters.

[0035] Next, the similarity calculation unit 5 calculates the similarity between the defect image data acquired in step S1 and the learned data by referring to the learned data storage unit 92 (step S4). For example, the similarity calculation unit 5 calculates the Euclidean distance between the feature amount of the defect image data and the feature amount of the learned data, and calculates the minimum value of the Euclidean distance for each class as the similarity.

[0036] After that, the judgment unit 6 first judges whether the similarity satisfies a predetermined criterion (step S5). If the similarity satisfies the criterion in step S5 (step S5: Yes), the data classification device 100 proceeds to step S6. The criterion here is defined as the range of values ​​that the minimum value takes, for example, when the similarity is the minimum value of the Euclidean distance between the feature amount of the defect image data and the feature amount of the learned data of each class.

[0037] In step S6, the determination unit 6 makes a determination according to the classification result by the classification unit 4 (step S6). Specifically, the determination unit 6 determines that the class with the maximum probability obtained in step S3 is the class to which the data of the defect image belongs.

[0038] In step S5, if the similarity does not satisfy the criterion (step S5: No), the determining unit 6 determines that the data of the defect image is of an unknown class (step S7).

[0039] After step S6 or S7, the control unit 8 increments the number of determinations stored in the storage unit 9 by 1 (step S8).

[0040] Next, the display control unit 81 outputs the judgment result to the display device and displays it (step S9). Figures 4 and 5 are diagrams showing examples of displaying the judgment result by the judgment unit 6 on the display device. Specifically, Figure 4 shows a case where the data of the defect image is of a known class, and Figure 5 shows a case where the data of the defect image is of an unknown class.

[0041] 4, a screen 201 is displayed showing information indicating that the defect image data (image name "Sample1.png") has been classified as defect type 1, along with four images and their Euclidean distances in order of shortest Euclidean distance. All four images are images of defect type 1.

[0042] 5, the screen 202 shows information indicating that the defect image data (image name "Sample3.png") is an unknown class, and four images are shown in order of the shortest Euclidean distance. These four images have different defect types.

[0043] According to these display examples, images similar to the image to be classified are displayed together with the judgment result, so that not only the judgment result but also information on the basis of the judgment is presented, and the judgment result can be visually convinced by the user. Note that at least one of the similar images and the Euclidean distance may be displayed. Furthermore, the number of images to be displayed is not limited to four, and may be set arbitrarily as long as it is one or more.

[0044] After step S9, the evaluation value calculation unit 7 refers to the storage unit 9 and compares the number of judgments made by the judgment unit 6 with a predetermined reference number (step S10). If the number of judgments made by the judgment unit 6 is less than the reference number (step S10: No), the process returns to the data acquisition process of step S1 and is repeated. If the number of judgments made by the judgment unit 6 reaches the reference number (step S10: Yes), the evaluation value calculation unit 7 calculates a judgment evaluation value, which is the proportion of judgments made that determine that the data of the defect image belongs to one of multiple classes among the judgments made the reference number of times (step S11).

[0045] If the calculated judgment evaluation value is smaller than a predetermined threshold (step S12: Yes), the display control unit 81 notifies the user of the judgment evaluation result by outputting and displaying information on the judgment evaluation result on the display device (step S13). Note that the display control unit 81 may display information prompting the reconstruction of the trained model on the display device instead of the judgment evaluation result, or may display the judgment evaluation result and information prompting the reconstruction of the trained model on the display device. This allows the user to recognize that the accuracy of the trained model is not appropriate.

[0046] Next, control unit 8 resets the number of judgments stored in storage unit 9 to 0 (step S14). After that, data classifying device 100 ends the series of processes.

[0047] In step S12, if the calculated judgment evaluation value is equal to or greater than a predetermined threshold value (step S12: No), data classifying device 100 proceeds to step S14, resets the judgment count, and ends the series of processes.

[0048] Fig. 6 is a diagram for explaining the effect of the embodiment described above. Specifically, Fig. 6 is a histogram showing (a) the result of calculating the minimum value of the Euclidean distance between a sample of a known class and each class of the learned data, and (b) the result of calculating the minimum value of the Euclidean distance between a sample of an unknown class and each class of the learned data. Here, the histogram is created assuming that the class to which the sample belongs has been determined in advance. Fig. 6 shows a case where two thresholds Th1 and Th2 are set as thresholds that provide a boundary for discriminating between a known class and an unknown class.

[0049] First, we will explain the case where the threshold is set to Th1, that is, where samples below the threshold Th1 are defined as a known class. In this case, according to (a), some of the samples of the known class are included in the unknown class, while according to (b), the samples of the unknown class are not included in the known class.

[0050] Next, we will explain the case where the threshold is set to Th2, that is, where samples below the threshold Th2 are defined as known classes. In this case, all samples of the known classes are included below the threshold according to (a), while most of the unknown classes are included in the known classes according to (b).

[0051] FIG. 7 is a diagram showing the variation of the detection rate of an unknown class and the false positive rate of a known class when the threshold is changed in the example shown in FIG. 6. A curve 301 shown in FIG. 7 is a curve called a receiver response curve (ROC: Receiver Operating Characteristic), and points P and Q on the curve correspond to thresholds Th1 and Th2, respectively. In the case of threshold Th1, the detection rate is 100%, while the false positive rate is less than 10%. In contrast, in the case of threshold Th2, the detection rate is about 15%, while the false positive rate is 0%. It was found that the AUC (Area Under ROC Curve), which is the area between this curve 301 and the horizontal axis, is 0.9899, ​​which is quite close to the case where the class is completely separable (AUC=1.00). That is, it was revealed that according to this embodiment, the known class and the unknown class can be separated with high accuracy.

[0052] FIG. 8 is a diagram for explaining another method of determining a threshold value that provides a boundary for discriminating between a known class and an unknown class in the above-described embodiment. FIG. 8(a) is a diagram showing a histogram showing the result of calculating the minimum value of the Euclidean distance between a sample of a known class and each class of learned data, and its normal distribution curve C. This normal distribution curve C is a line drawn on the assumption that the histogram is normally distributed. FIG. 8(b) is a histogram showing the result of calculating the minimum value of the Euclidean distance between a sample of an unknown class and each class of learned data. In FIG. 8, as in FIG. 6, the histogram is created assuming that the class to which the sample belongs has been determined in advance. FIG. 8 shows a case where a threshold value Th3 or Th4 is set as a threshold value that provides a boundary for discriminating between a known class and an unknown class.

[0053] In setting a threshold value that provides a boundary for discriminating between known classes and unknown classes, it may be difficult to prepare samples of the unknown class in advance. In such cases, the threshold value is set based on the minimum value of the Euclidean distance between the samples of the known class and each class of the trained data.

[0054] First, a normal distribution is assumed for a histogram showing the result of calculating the minimum value of the Euclidean distance between a sample of a known class and each class of the trained data. Next, a threshold is determined from a confidence interval calculated from the mean μ and standard deviation σ of the normal distribution. Note that the distribution to be sought is not limited to a normal distribution, and a half-normal distribution, a Poisson distribution, or a mixed distribution that takes multimodality into consideration may be assumed as a probability distribution.

[0055] Here, a case where the threshold is set to Th3 (66% confidence interval of normal distribution: mean μ+1σ), that is, a case where samples below the threshold Th3 are defined as a known class, will be described. In this case, as shown in (a) of Fig. 8, some samples of the known class are included in the unknown class, while as shown in (b), no samples of the unknown class are included in the known class.

[0056] In addition, a case where the threshold is set to Th4 (99.9997 confidence interval of normal distribution: mean μ+6σ), that is, a case where samples below the threshold Th4 are defined as known classes, will be described. In this case, as shown in (a) of Fig. 8, all samples of known classes are included below the threshold, while as shown in (b), many samples of unknown classes are included in known classes.

[0057] From the above, when it is difficult to prepare samples of an unknown class in advance, a threshold value according to the operational purpose may be set only from samples of known classes. For example, when it is desired to avoid erroneously determining that a sample of a known class is a sample of an unknown class, the threshold value may be set to Th4. Conversely, when it is desired to avoid erroneously determining that a sample of an unknown class is a sample of a known class, the threshold value may be set to Th3. The threshold value setting is not limited to the above, and it is preferable to set the threshold value according to the operational purpose.

[0058] According to one embodiment of the present invention described above, the similarity between the acquired defect image data and the learned data used when generating the learned model is calculated, and whether the defect image data belongs to multiple classes is determined based on the calculated similarity. If the data does not belong to any of the multiple classes, it is determined that the data belongs to an unknown class, thereby preventing classification of data of an unknown class into a known class.

[0059] Furthermore, according to this embodiment, at least the data of the defect image that is most similar to the data of the defect image is displayed on the display device along with the judgment result, so that the user can visually grasp not only whether the data of the defect image belongs to a known class or an unknown class, but also the basis for the judgment.

[0060] Further, according to this embodiment, when the number of judgments reaches a reference number, a judgment evaluation value is calculated, which is the ratio of the judgments made the reference number of times that the data of the defect image belongs to one of a plurality of classes, and when the calculation result is smaller than the threshold value, information that prompts the reconstruction of the evaluation result or the learned model is output, so that if the learned model being used is invalid, it is possible to promptly notify the fact. In this regard, in the conventional film manufacturing process, if the optical system of the defect inspection machine is changed or the manufacturing conditions of the product are changed, there is a risk that the learned model used up to that point will not be able to properly classify. In addition, it is difficult to evaluate the classification itself during operation. In contrast, in this embodiment, the classification itself is evaluated at a predetermined interval, so that if there is a change in the manufacturing process, the user can be accurately notified of the incompatibility of the learned model. In addition, although the image is pre-processed in step S2, it is possible to classify the image without the image processing of step S2 prepared in advance, considering the possibility that there may be an error here, or to perform another image processing method under conditions in which the threshold value, filter, etc. prepared in advance are changed, and to try classification from step S3 onwards again.

[0061] Although the embodiment for carrying out the present invention has been described above, the present invention should not be limited to the above-mentioned embodiment. For example, the similarity calculation unit 5 may calculate the similarity between the feature amount of the defect image data and the feature amount of the representative image data of each class. The representative image may be an image having a feature amount closest to the center of gravity in the feature space of the feature amount of each class, or an image having a feature amount closest to the average of the feature amount of each class.

[0062] Furthermore, the present invention can be applied not only to defect detection during film production, but also to abnormality detection in time-series data (for example, arrhythmia detection in human heart rate data). [Industrial Applicability]

[0063] The data classification device, data classification method, and data classification program of the present invention can be preferably applied to suppress classification of data of an unknown class into a known class, but the scope of application is not limited to this. [Explanation of symbols]

[0064] 1 Input section 2 Data Acquisition Section 3. Image Processing Section 4 Classification section 5 Similarity determination section 6 Judgment section 7 Evaluation value calculation section 8. Control Unit 9 Storage section 81 Display control unit 91 Trained model memory 92 Learned data storage unit 100 Data Classifier 201, 202 screens 301 curve

Claims

1. A data acquisition unit that acquires data to be classified; A classification unit that classifies the data into one of a plurality of classes using a trained model trained using a neural network; A similarity calculation unit that calculates a similarity between the data and trained data used in generating the trained model; a determination unit that determines whether the data belongs to the plurality of classes based on the similarity, and determines that the data belongs to an unknown class when the data does not belong to any of the plurality of classes; Equipped with The trained model has the data as input parameters, and the classification result and the feature amount of the data as output parameters, The similarity calculation unit calculates a similarity using features of the data and the learned data. Data Classifier.

2. The determination unit: If the similarity satisfies a criterion, a class to which the data belongs is determined according to a classification result of the classification unit; 2. The data classification apparatus according to claim 1, wherein if the similarity does not satisfy the criterion, the data is determined to belong to an unknown class.

3. 3. The data classification device according to claim 1, further comprising a control unit that displays, on a display device, at least data that is most similar to the data, together with the determination result by the determination unit.

4. an evaluation value calculation unit that calculates, when the number of judgments made by the judgment unit reaches a reference number, a judgment evaluation value that is a ratio of judgments made the reference number of times that the data belongs to any one of the plurality of classes; A control unit that outputs at least one of an evaluation result and information that prompts reconstruction of the trained model when the judgment evaluation value is smaller than a threshold value; The data classification device according to any one of claims 1 to 3, comprising:

5. The similarity calculation unit is The data classification device according to any one of claims 1 to 4, wherein the minimum value of the Euclidean distance in a feature space between the features of the data and the features of the trained data in each class is calculated as the similarity with each class.

6. The similarity calculation unit is The data classification device according to any one of claims 1 to 4, wherein a predetermined distance in a feature space between the feature of the data and the learned data representing each of the plurality of classes is calculated as a similarity with each class.

7. A data acquisition step of acquiring data to be classified; A classification step of reading out a trained model trained using a neural network from a storage unit and classifying the data into one of a plurality of classes using the trained model; A similarity calculation step of calculating a similarity between the data and trained data used in generating the trained model; a determination step of determining whether the data belongs to the plurality of classes based on the similarity, and determining that the data belongs to an unknown class when the data does not belong to any of the plurality of classes; Including, The trained model has the data as input parameters, and the classification result and the feature amount of the data as output parameters, The similarity calculation step calculates a similarity using feature amounts of the data and the learned data. Data classification methods.

8. A data acquisition step of acquiring data to be classified; A classification step of classifying the data into one of a plurality of classes using a trained model trained using a neural network; A similarity calculation step of calculating a similarity between the data and trained data used in generating the trained model; a determination step of determining whether the data belongs to the plurality of classes based on the similarity, and determining that the data belongs to an unknown class when the data does not belong to any of the plurality of classes; Run the following on your computer: The trained model has the data as input parameters, and the classification result and the feature amount of the data as output parameters, The similarity calculation step calculates a similarity using feature amounts of the data and the learned data. Data classification program.

Images