Classification systems, classification methods, and program products

Abstract

The present disclosure provides a classification system, a classification method, and a program. Even when data that may belong to a new category that does not exist in the original category is input, it is easy to appropriately classify the data. The classification system (100) includes an input acceptance unit (11), a classification unit (F1), a calculation unit (F2), a determination unit (F3), and a prompt unit (12). The input acceptance unit (11) accepts input of object data. The classification unit (F1) classifies the object data into any category among a plurality of categories. The calculation unit (F2) calculates a feature value of the object data. The determination unit (F3) determines the possibility that the object data is classified into the new category based on the classification result in the classification unit (F1) and the feature value of the object data calculated by the calculation unit (F2). When the determination unit (F3) determines that there is a possibility that the object data is classified into the new category, the prompt unit (12) prompts the determination result of the determination unit (F3).

Description

Technical Field

[0001] The present disclosure generally relates to a classification system, a classification method, and a program. More specifically, the present disclosure relates to a classification system, a classification method, and a program for classifying input data. Background Art

[0002] Patent Document 1 discloses a classification method that uses a neural network to transform an input pattern, the target of classification, into an output pattern, the classification result. This method compares the neural network's output pattern for the input pattern with the correct output pattern subsequently output for that input pattern, and accumulates the comparison results as execution history records. Furthermore, the accumulated results are monitored to detect any anomalies. Another classification method, for example, is disclosed in Patent Document 2.

[0003] Prior art literature

[0004] Patent Literature

[0005] Patent Document 1: Japanese Patent Application Laid-Open No. 5-35707

[0006] Patent Document 2: U.S. Patent Application Publication No. 2017 / 0039469 Summary of the Invention

[0007] The classification method described in Patent Document 1 has a problem in that, when object data that may belong to a new category that does not exist in the original output pattern (category) is input, the object data cannot be appropriately classified.

[0008] The present disclosure is made in view of the above-mentioned problems, and its purpose is to provide a classification system, classification method and program that can easily and appropriately classify object data even when object data that may belong to a new category that does not exist in the original category is input.

[0009] One embodiment of the present disclosure relates to a classification system comprising an input receiving unit, a classification unit, a calculation unit, and a determination unit. The input receiving unit receives input of object data. The classification unit classifies the object data into any one of a plurality of categories. The calculation unit calculates a feature value of the object data. The determination unit determines the likelihood that the object data is classified into a new category different from the plurality of categories based on the classification result of the classification unit and the feature value of the object data calculated by the calculation unit.

[0010] A classification method according to one embodiment of the present disclosure includes an input acceptance step, a classification step, a calculation step, and a determination step. The input acceptance step is a step of accepting input of object data. The classification step is a step of classifying the object data into any one of a plurality of categories. The calculation step is a step of calculating a feature value of the object data. The determination step is a step of determining the likelihood that the object data is classified into a new category different from the plurality of categories based on the classification result in the classification step and the feature value of the object data calculated in the calculation step.

[0011] A program according to one embodiment of the present disclosure causes one or more processors to execute the above-described classification method.

[0012] The present disclosure has an advantage that even when object data that may belong to a new category that does not exist in the existing categories is input, it is easy to appropriately classify the object data. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a block diagram showing the structure of a classification system according to one embodiment of the present disclosure.

[0014] Figure 2 This is an explanatory diagram showing an example of object data input to the above-mentioned classification system.

[0015] Figure 3 This is an explanatory diagram showing the structure of the classification unit in the above-mentioned classification system.

[0016] Figure 4 This is an explanatory diagram showing an example of the first determination operation in the determination unit in the above-mentioned classification system.

[0017] Figure 5 This is an explanatory diagram showing an example of the second determination operation in the determination unit in the above-mentioned classification system.

[0018] Figure 6 This is an explanatory diagram showing another example of the second determination operation in the determination unit in the above-mentioned classification system.

[0019] Figure 7 1 is an explanatory diagram showing a feature quantity space of candidates for a new class in the above-described classification system.

[0020] Figure 8 This is a flowchart showing an example of the operation of the above-mentioned classification system. DETAILED DESCRIPTION

[0021] (1) Summary

[0022] The classification system 100 according to this embodiment is a system for classifying object data into any of a plurality of categories. In this embodiment, the object data is an object image A1 (see Figure 2 ) data. The target image A1 is an image of the determination target object 2. As an example, the determination target object 2 is a can. Figure 2 The object image A1 shown shows the bottom surface of a can in a normal state. Here, the object 2 can be in a normal state or at least one of multiple abnormal states. For example, abnormal states may include the object 2 having dotted dents, linear scratches, circular abrasions, attached dirt, or attached foreign matter (such as liquid). In other words, the classification system 100 classifies the object 2 into any of multiple categories (here, normal state and multiple abnormal states). Therefore, the classification system 100 can be used to inspect the object 2.

[0023] like Figure 1 As shown, the classification system 100 includes an input accepting unit 11 , a classification unit F1 , a calculation unit F2 , a determination unit F3 , and a presentation unit (output unit) 12 .

[0024] The input accepting unit 11 accepts input of target data. In this embodiment, as an example, image data obtained by imaging the determination target object 2 by an imaging device is input to the input accepting unit 11 as target data.

[0025] Classification unit F1 classifies the object data into any of a plurality of categories. In this embodiment, classification unit F1 classifies the object data using a pre-learned model that has been machine-learned to determine the input-output relationship between the object data and the classification results. In other words, the plurality of categories correspond to pre-defined categories defined by machine learning.

[0026] The calculation unit F2 calculates the feature value of the target data. In this embodiment, as an example, the feature value is calculated based on the difference between the feature vector of the image data as the target data and the feature vector of the reference image data.

[0027] The determination unit F3 determines the likelihood that the object data has been classified into a new category different from the multiple categories based on the classification result of the classification unit F1 and the feature value of the object data calculated by the calculation unit F2. In other words, the determination unit F3 determines the likelihood that the object data is more likely to be classified into the new category than into any of the multiple categories into which the object data has been originally classified.

[0028] When the determination unit F3 determines that the object data may be classified into a new category, the presentation unit 12 presents the determination result of the determination unit F3. For example, the presentation unit 12 presents the determination result to the user of the classification system 100 via a display device such as a liquid crystal display.

[0029] As described above, in this embodiment, when the classification system 100 determines that there is a possibility that object data may be classified into a new category, the classification system 100 does not automatically cause the classification system F1 to relearn and classify the object data into a new category. Instead, the system prompts the user to indicate that a new category is required. Therefore, this embodiment has the advantage of being able to easily and appropriately classify the object data even when object data is input that may belong to a new category that does not exist in the existing categories.

[0030] (2) Details

[0031] Below, refer to Figure 1 The classification system 100 according to this embodiment will be described. The classification system 100 includes an input receiving unit 11 , an output unit 12 , a storage unit 13 , and a processing unit 14 .

[0032] The input accepting unit 11 accepts input of target data. Specifically, the input accepting unit 11 is the main body that executes the input acceptance step S1 described below. In this embodiment, the input accepting unit 11 is a communication interface that receives signals transmitted from other devices via wired or wireless communication. It receives image data of the determination target object 2 captured by the imaging device as target data. The target data (image data of the determination target object 2) can be transmitted directly from the imaging device to the input accepting unit 11 or indirectly via an intermediate medium. The target data accepted by the input accepting unit 11 is provided to the processing unit 14.

[0033] The processing unit 14 is configured to control the entire classification system 100, specifically, the input receiving unit 11, the output unit 12, and the storage unit 13. The processing unit 14 primarily comprises a computer system comprising one or more processors and memory. Thus, the one or more processors execute programs stored in memory, thereby functioning as the processing unit 14. The programs may be pre-stored in memory, provided via telecommunication lines such as the Internet, or stored on non-volatile storage media such as memory cards.

[0034] The processing unit 14 includes a classification unit F1, a calculation unit F2, and a determination unit F3. Figure 1 In FIG. 1 , the classification unit F1 , the calculation unit F2 , and the determination unit F3 are not shown as having actual structures, but are functions implemented by the processing unit 14 .

[0035] The classification unit F1 is configured to classify the object data into any of a plurality of categories. In other words, the classification unit F1 is the execution subject of the classification step S3 described later. Figure 3 As shown, the classification unit F1 includes a plurality of classifiers F11 to F1n ("n" is an integer of 2 or more) and a NOR circuit F10.

[0036] Each classifier F11 to F1n may include, for example, a classifier using a neural network, or a classifier generated by deep learning (deep learning) using a multi-layer neural network, in addition to a linear classifier such as an SVM (Support Vector Machine). In this embodiment, each classifier F11 to F1n is a classifier using a learned neural network. The learned neural network may include, for example, a CNN (Convolutional Neural Network) or a BNN (Bayesian Neural Network). Each classifier F11 to F1n may be implemented by installing the learned neural network on an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field-Programmable Gate Array).

[0037] In this embodiment, the multiple categories include multiple corresponding categories (the first category, the second category, ..., the nth category) and other categories. The multiple corresponding categories are categories corresponding to the multiple classifiers F11 to F1n included in the classification unit F1. That is, in this embodiment, object data is input to the multiple classifiers F11 to F1n respectively. Then, the multiple classifiers F11 to F1n determine whether the object data respectively input is classified into the corresponding category corresponding to themselves. For example, whether the object data is classified into the nth category is classified by the nth classifier F1n. That is, when the object data is input to the nth classifier F1n, if the output of the nth classifier F1n is "High", it is determined that the object data is classified into the nth category, and if it is "Low", it is determined that the object data is not classified into the nth category.

[0038] In this embodiment, each classifier F11 to F1n outputs "High" if the reliability exceeds a threshold, and outputs "Low" if the reliability is below the threshold. "Reliability" as used herein refers to the probability that the object data belongs to the category classified by classifier F1. For example, if the reliability of the first classifier F11 is "1," the probability that classifier F1 determines that the object data belongs to category 1 is 100%. Alternatively, if the reliability of the first classifier F11 is "0.8," the probability that classifier F1 determines that the object data belongs to category 1 is 80%, and the probability that the object data belongs to another category is 20%.

[0039] The outputs of multiple classifiers F11 to F1n are input to the NOR circuit F10. Furthermore, when the outputs of all classifiers F11 to F1n are "Low," the NOR circuit F10 outputs "High." Here, "other categories" are categories that do not correspond to any of the multiple classifiers F11 to F1n. In other words, object data classified as "other categories" is data that is not classified into any of the multiple corresponding categories, i.e., the first to nth categories. Furthermore, in this embodiment, when the NOR circuit F10 outputs "High," that is, when the outputs of all classifiers F11 to F1n are "Low," the object data is classified as "other categories."

[0040] Thus, in this embodiment, the classification unit F1 classifies the object data received by the input reception unit 11 into any one of a plurality of categories using the plurality of classifiers F11 to F1n and the NOR circuit F10. Furthermore, while the classification unit F1 basically classifies the object data into any one of the plurality of categories, it may also classify the object data into two or more of the plurality of corresponding categories.

[0041] The calculation unit F2 is configured to calculate the feature quantity of the object data. In other words, the calculation unit F2 is the main body for executing the calculation step S2 described later. In the calculation of the feature quantity, reference data that serves as a reference for the object data is used. The reference data is an image of the determination object 2, just like the object data. In this embodiment, the reference data is data representing an image (reference image) of the determination object 2 in a normal state. Here, Figure 2 The target image A1 shown shows an image of the determination target object 2 in a normal state and can be used as the reference image A2. In this embodiment, when calculating the feature value, the target image A1 is normalized so that the position and size of the determination target object 2 match the position and size of the determination target object 2 in the reference image A2.

[0042] As an example, the calculation unit F2 calculates the feature quantity of the object data using feature vectors. For example, the calculation unit F2 calculates a feature vector for each block of equal size, consisting of one or more pixels, commonly defined in the target image A1 and the reference image A2. A feature vector uses the pixel values of the pixels contained in the block as its elements. The calculation unit F2 then calculates the difference vector between the feature vector for each block in the target image A1 and the feature vector for each block in the reference image A2 as the feature quantity of the object data. The calculated feature quantity is represented as a multidimensional vector based on the number of blocks and the number of pixels in each block.

[0043] In this embodiment, the calculation unit F2 further performs appropriate transformation processing such as t-SNE (t-distributed Stochastic Neighbor Embedding) on the feature amount of the target data calculated as described above. Figure 4 As shown, the calculation unit F2 calculates the feature quantity of the object data as a point D1 on the two-dimensional feature quantity space FS1. Of course, the feature quantity of the object data can also be expressed as a point on a multi-dimensional feature quantity space of three or more dimensions.

[0044] The determination unit F3 is configured to determine the possibility that the object data is classified into a new category different from the multiple categories based on the classification result of the classification unit F1 and the feature amount of the object data calculated by the calculation unit F2. In other words, the determination unit F3 is the main body for executing the determination steps S4 to S8 described below.

[0045] In this embodiment, determination unit F3 distinguishes whether the object data is classified as another category or a new category. That is, in this embodiment, if the object data is classified as another category by classification unit F1, determination unit F3 does not immediately classify the object data as a new category. Instead, for example, if a given condition is further satisfied, determination unit F3 considers the object data as a candidate for the new category. Furthermore, if the given condition is not satisfied, determination unit F3 classifies the object data as another category. Furthermore, in this embodiment, even if the object data is classified as two or more corresponding categories by classification unit F1, determination unit F3 determines whether the object data is a candidate for the new category.

[0046] For example, if the object data is classified into another category by the classification unit F1, and if the object data satisfies "Condition A," the determination unit F3 determines that the object data is a candidate for classification into a new category. On the other hand, if the object data does not satisfy "Condition A," the determination unit F3 determines that the object data is classified into another category.

[0047] "Condition A" is that the feature value of the object data is at least a given distance away from any of the plurality of corresponding categories in the feature value space FS1. Figure 4 An example of satisfying "Condition A" is explained below. Figure 4 In , "C1" to "C3" represent a set of feature quantities of object data that has been previously classified into the first to third categories in the feature quantity space FS1. Figure 4 In the example, different patterns are used to represent the feature quantity of the object data in "C1" to "C3". Figure 4 In the feature space FS1, "D1" represents the feature of the object data classified by the classification unit F1. Figure 4Although not shown, the feature space FS1 also includes a set of feature values of object data that has been previously classified into categories other than the first to third categories.

[0048] like Figure 4 As shown, "D1" is at least a given distance from any of "C1" to "C3." Here, "at least a given distance" means, as an example, that "D1" is outside the circle centered on the center point of each category in feature space FS1. Furthermore, "at least a given distance" means, as an example, that the feature closest to "D1" in each category of object data in feature space FS1 is at a distance greater than a given value from "D1."

[0049] Furthermore, as an example, when the object data is classified by the classification unit F1 into two or more corresponding categories, if the object data satisfies "Condition B," the determination unit F3 determines that the object data is a candidate for classification into a new category. On the other hand, if the object data does not satisfy "Condition B," the determination unit F3 determines that the object data is classified into two or more corresponding categories.

[0050] “Condition B” includes a “first condition” and a “second condition.” In the present embodiment, when both the “first condition” and the “second condition” are satisfied, the determination unit F3 determines that “Condition B” is satisfied.

[0051] The "first condition" is that the feature value of the object data is between two or more corresponding categories in the feature value space FS1. Figure 5 An example of satisfying the "first condition" is explained below. Figure 5 In , "C1" and "C2" represent a set of feature quantities of object data that has been previously classified into the first category and the second category, respectively, in the feature quantity space FS1. Figure 5 In "C1" and "C2", the feature quantity of the object data is represented by different patterns. Figure 5 In the example, the feature quantity of the object data in the area surrounded by the dotted line belongs to both "C1" and "C2". Figure 5 In the feature space FS1, "D1" represents the feature of the object data classified by the classification unit F1. Figure 5 Although not shown in FIG. 1 , the feature space FS1 also includes a set of feature values of object data that has been previously classified into categories other than the first and second categories.

[0052] like Figure 5As shown in the figure, "D1" is located between "C1" and "C2". Here, "between two or more corresponding categories" means, as an example, that the object data is included inside a circle centered on the center point of an area defined by the center points of two or more corresponding categories (a line segment if there are two categories, or a polygon if there are three or more categories). Figure 5 In the example shown, “ D1 ” is included inside a circle centered at the center point of the line segment L1 defined by the center points of “ C1 ” and “ C2 ”.

[0053] The second condition is that any of the two or more classifiers F11 to F1n corresponding to the two or more corresponding categories focuses on the same part of the object data. In other words, if any of the two or more classifiers F11 to F1n corresponding to the two or more corresponding categories does not focus on the same part of the object data, the determination unit F3 excludes the object data from the candidate. Figure 6 An example satisfying the "second condition" will be described. Figure 6 The object image A1 is shown after appropriate visualization processing such as Grad-CAM (Gradient-weighted Class Activation Mapping) is performed. Figure 6 , a first range A11 focused on in the target data by the first classifier F11 among the two or more classifiers F11 to F1n, and a second range A12 focused on in the target data by the second classifier F12 among the two or more classifiers F11 to F1n are shown. Figure 6 In the example shown, the first range A11 and the second range A12 overlap, and two or more classifiers F11 and F12 focus on the same portion of the target data.

[0054] In this embodiment, if the number of object data considered as candidates for a new category exceeds a predetermined number as described above, determination unit F3 determines that the object data has been classified into a new category. Specifically, determination unit F3 does not determine that the object data has been classified into a new category at the stage of considering the object data as candidates for a new category. Only when the number of candidates exceeds a predetermined number (e.g., several dozen) does determination unit F3 determine that the candidate group for the object data has been classified into a new category.

[0055] The storage unit 13 includes one or more storage devices. Examples of the storage devices include RAM (Random Access Memory) and EEPROM (Electrically Erasable Programmable Read Only Memory). The storage unit 13 stores object data and the determination results of the determination unit F3 for the object data. In this embodiment, the storage unit 13 does not store all object data input to the input acceptance unit 11. Instead, it stores object data that the determination unit F3 determines has the potential to be classified into a new category (that is, candidates for the new category) and the determination results for that object data.

[0056] In this embodiment, the determination result of the determination unit F3 is associated with the group G0 corresponding to the similarity of the feature amount of the target data and stored in the storage unit 13. Figure 7 An example of classifying the feature amount of the object data into a plurality of groups G1, G2, and G3 will be described. Figure 7 In the example, groups G1-G3 each represent a set of feature values of object data that are candidates for a new class in feature value space FS1. Within each of groups G1-G3, the feature values of multiple object data are similar to each other. Furthermore, groups G1-G3 are located separately from each other in feature value space FS1, meaning they are not similar to each other. Furthermore, if the number of candidate object data in each group G1-G3 exceeds a predetermined number, the determination unit F3 can determine that the group G1-G3 represents a new class.

[0057] The output unit 12 is configured to output information processed by the processing unit 14. In this embodiment, the output unit 12 includes a port for outputting data and an image display device for displaying information. For example, the image display device may include a thin display device such as a liquid crystal display or an organic EL (electroluminescence) display.

[0058] In the present embodiment, the output unit 12 also serves as the prompt unit 12. The prompt unit 12 is configured to prompt the user of the classification system 100 with the determination result of the determination unit F3 when the determination unit F3 determines that there is a possibility that the object data is classified into a new category. That is, the prompt unit 12 is the execution body of the prompt step S9 described later. Here, the prompt unit 12 prompts the user with the determination result by displaying the determination result on the image display device. In the present embodiment, since the determination result is stored in the storage unit 13, the prompt unit 12 prompts the determination result read from the storage unit 13 by the processing unit 14. For example, the prompt unit 12 displays the object image A1 and a message indicating that the object image A1 may be classified into a new category on the image display device as the determination result.

[0059] In the present embodiment, the prompting unit 12 prompts the determination result in a manner corresponding to the number of object data belonging to the group G0. As an example, when "condition C" is not satisfied, the prompting unit 12 prompts the determination result of the determination unit F3. In this case, the prompting unit 12 may also display the number of candidates for object data belonging to one or more groups G0 on the image display device. Moreover, when "condition C" is satisfied, the prompting unit 12 displays an image or the like that prompts the user to re-learn the classification unit F1 together with the determination result of the determination unit F3 on the image display device. "Condition C" is that the number of object data selected as candidates for a new category by the determination unit F3 in any of the more than one groups G0 exceeds a given number.

[0060] In addition, in this embodiment, the prompting unit 12 displays the reliability and feature space FS1 of the object data in multiple categories. As an example, when the object data is classified into other categories by the classification unit F1 and satisfies "condition A", the prompting unit 12 displays the following Figure 4 The feature space FS1 and the reliability of each category are displayed on the image display device. In addition, as an example, when the object data is classified into two or more corresponding categories among a plurality of corresponding categories by the classification unit F1 and the "condition B" is satisfied, the presentation unit 12 displays the following information: Figure 5 The feature space FS1 shown and a list of the reliability in each category are displayed on the image display device.

[0061] (3) Action

[0062] Below, using Figure 8 An example of the operation of the classification system of this embodiment will be described. First, when the input accepting unit 11 accepts input of object data (S1), the calculation unit F2 calculates the feature value of the object data accepted by the input accepting unit 11 (S2). The classification unit F1 then uses the learned model to classify the object data accepted by the input accepting unit 11 (S3). The order of the calculation process for the feature value of the object data performed by the calculation unit F2 and the classification process for the object data performed by the classification unit F1 may be reversed.

[0063] Next, the determination unit F3 determines the likelihood that the object data has been classified into the new category based on the classification results of the classification unit F1 and the feature values of the object data calculated by the calculation unit F2. If the object data has been classified into another category by the classification unit F1 (S4: Yes), and if the object data satisfies "Condition A" (S5: Yes), the determination unit F3 considers the object data as a candidate for the new category (S6). On the other hand, if the object data has not been classified into another category by the classification unit F1 (S4: No) but has been classified into two or more of the corresponding categories (S7: Yes), and if "Condition B" (S8: Yes) has been satisfied, the determination unit F3 considers the object data as a candidate for the new category (S6).

[0064] If the determination unit F3 determines that the object data is a candidate for a new category, that is, if there is a possibility that the object data is classified into a new category, the presentation unit 12 presents the determination result of the determination unit F3 (S9). If "condition C" is satisfied (S10: Yes), the presentation unit 12 displays an image or the like prompting the user to re-learn the classifier F1, along with the determination result of the determination unit F3, on the image display device, thereby notifying the user that the classifier F1 should be re-learned (S11).

[0065] The user who has received the notification that re-learning is required performs a given operation to allow re-learning, for example, using an interface for accepting operations on the processing unit 14. Then, in the processing unit 14, a group of object data determined to be classified as a new category is used as teaching data to perform re-learning of the learned model of the classification unit F1. In this way, in the present embodiment, the classification unit F1 re-learns the object data determined by the determination unit F3 to be classified as a new category as teaching data. In the present embodiment, the re-learning of the classification unit F1 is equivalent to adding a new category. In addition, the so-called "adding a new category" is equivalent to adding a new classifier to the classification unit F1 to determine whether the input object data is classified as a new category.

[0066] (4) Advantages

[0067] As described above, in this embodiment, when the determination unit F3 determines that there is a possibility that object data may be classified into a new category, the classification system 100 does not automatically cause the classification unit F1 to relearn and classify the object data into a new category. Instead, the user is prompted to indicate that a new category is required. Therefore, this embodiment has the advantage of being able to easily and appropriately classify the object data even when object data is input that may belong to a new category that does not exist in the existing categories.

[0068] If the classifiers F11 to F1n of the classification unit F1 are not sufficiently trained, even if the object data is correctly classified into any of the multiple categories, the determination unit F3 may mistakenly determine that the object data is likely to be classified into a new category. Therefore, this embodiment has the advantage of displaying the determination result of the determination unit F3 by the display unit 12, thereby enabling the user to determine whether the object data is classified into a new category, thereby improving the reliability of the object data being classified into a new category.

[0069] (5) Modification

[0070] The above-mentioned embodiment is only one of various embodiments of the present disclosure. As long as the above-mentioned embodiment can achieve the purpose of the present disclosure, various changes can be made to the above-mentioned embodiment according to the design, etc. In addition, the same functions as the classification system 100 involved in the above-mentioned embodiment can also be achieved by a classification method, a computer program, or a non-volatile recording medium recording a computer program.

[0071] A classification method according to one embodiment includes an input acceptance step S1, a classification step S3, a calculation step S2, judgment steps S4 to S8, and a prompting step S9. Input acceptance step S1 is a step of accepting input of object data. Classification step S3 is a step of classifying the object data into any one of a plurality of categories. Calculation step S2 is a step of calculating a feature value of the object data. Judgment steps S4 to S8 are steps of determining the possibility that the object data is classified into a new category different from the plurality of categories based on the classification result in classification step S3 and the feature value of the object data calculated in calculation step S2. Prompting step S9 is a step of prompting the judgment result of judgment steps S4 to S8 when judgment steps S4 to S8 determine that there is a possibility that the object data is classified into a new category. A (computer) program according to one embodiment causes one or more processors to execute the above-described classification method.

[0072] Modifications of the above-described embodiment are listed below. The modifications described below can be combined as appropriate and applied.

[0073] The classification system 100 in the present disclosure includes a computer system, for example, in the processing unit 14. The computer system mainly comprises a processor and a memory as hardware. The functions of the classification system 100 in the present disclosure are realized by the processor executing a program recorded in the computer system's memory. The program can be pre-recorded in the computer system's memory, provided via an electrical communication line, or recorded on a non-volatile recording medium such as a memory card, optical disk, or hard disk drive that can be read by the computer system. The processor of the computer system can be composed of one or more electronic circuits including semiconductor integrated circuits (ICs) or large-scale integrated circuits (LSIs). The integrated circuits such as ICs or LSIs referred to here are referred to differently depending on the degree of integration, including integrated circuits called system LSIs, VLSIs (Very Large Scale Integration), or ULSIs (Ultra Large Scale Integration). Furthermore, FPGAs (Field-Programmable Gate Arrays) that are programmable after LSI manufacturing, or logic devices that can reconfigure the internal connection relationships or circuit partitioning within the LSI, can also be used as processors. Multiple electronic circuits can be integrated into a single chip or distributed across multiple chips. Multiple chips can be integrated into a single device or distributed across multiple devices. The computer system referred to here includes a microcontroller having one or more processors and one or more memories. Therefore, a microcontroller can also be composed of one or more electronic circuits including semiconductor integrated circuits or large-scale integrated circuits.

[0074] Furthermore, it is not necessary for the classification system 100 to integrate multiple functions within a single housing. The components of the classification system 100 may be dispersed across multiple housings. Furthermore, at least a portion of the functions of the classification system 100 may be implemented via the cloud (cloud computing).

[0075] In the above embodiment, the presentation unit 12 may not present the user with the determination result each time the determination unit F3 obtains a determination result, but may present the determination result to the user only when the object data is determined to be classified into a new category. In other words, the presentation unit 12 may not present the determination result to the user only when the determination unit F3 determines that the object data is a candidate for classification into a new category.

[0076] In the above embodiment, when the classifier F1 relearns object data determined to be classified into a new category as teaching data, the correct label attached to the object data may be determined by the user or automatically determined by the processing unit 14 .

[0077] In the above-described embodiment, the user may cause the classifying unit F1 to re-learn after excluding object data considered not to correspond to the new category from the group of object data presented by the presentation unit 12 .

[0078] In the above-described embodiment, the target data is not limited to image data of the determination target object 2, but may also be sound data emitted by the determination target object 2. In this case, the "second condition" in the determination unit F3 is that any of the two or more classifiers F11 to F1n corresponding to the two or more corresponding categories focus on the same portion (time-series change) of the target data.

[0079] In the above-described embodiment, when determining whether there is a possibility that the object data is classified into a new category, the "second condition" of the "first condition" and "second condition" of "Condition B" need not be used. In other words, even if the object data is classified into two or more corresponding categories and the feature value of the object data is between the two or more corresponding categories in the feature value space FS1, the determination unit F3 may determine that the object data is a candidate for classification into a new category.

[0080] In the above embodiment, the presentation unit 12 may output the determination result of the determination unit F3 as sound through a speaker, etc. In this case, the presentation unit 12 may display the determination result of the determination unit F3 on the image display device together with the sound output.

[0081] (Summarize)

[0082] As described above, the classification system (100) according to the first embodiment includes an input receiving unit (11), a classification unit (F1), a calculation unit (F2), and a determination unit (F3). The input receiving unit (11) receives input of object data. The classification unit (F1) classifies the object data into any one of a plurality of categories. The calculation unit (F2) calculates a feature value of the object data. The determination unit (F3) determines the possibility that the object data is classified into a new category different from the plurality of categories based on the classification result of the classification unit (F1) and the feature value of the object data calculated by the calculation unit (F2).

[0083] This method has the advantage that even when object data that may belong to a new category that does not exist in the existing categories is input, it is easy to appropriately classify the object data.

[0084] The classification system (100) according to the second embodiment further comprises a presentation unit (12) in the first embodiment. The presentation unit (12) presents the determination result of the determination unit (F3) when the determination unit (F3) determines that there is a possibility that the object data is classified into a new category.

[0085] In a classification system (100) according to a third aspect, in the first aspect or the second aspect, the plurality of categories include a plurality of corresponding categories corresponding to the plurality of classifiers (F11 to F1n) included in the classification unit (F1) and other categories not corresponding to any of the plurality of classifiers (F11 to F1n). The determination unit (F3) distinguishes whether the determination target data is classified into the other categories or the new category.

[0086] According to this aspect, there is an advantage that, when the learning of the classifier ( F1 ) is insufficient, it is possible to reduce the likelihood of erroneous determination that the target data may be classified into a new category.

[0087] In a classification system (100) according to a fourth aspect, under the third aspect, the determination unit (F3) determines that the object data is a candidate for classification into a new category when the following conditions are satisfied: the object data is classified into another category, the reliability of any one of a plurality of corresponding categories is lower than a threshold, and the feature quantity of the object data is at least a given distance away from any one of the plurality of corresponding categories in the feature quantity space (FS1).

[0088] According to this aspect, there is an advantage that, when the learning of the classifier ( F1 ) is insufficient, it is possible to reduce the likelihood of erroneous determination that the target data may be classified into a new category.

[0089] In a classification system (100) according to a fifth aspect, under the third aspect, a determination unit (F3) determines that object data is a candidate for classification into a new category when the following conditions are satisfied: the object data is classified into two or more corresponding categories among a plurality of corresponding categories, and a feature value of the object data is between the two or more corresponding categories in a feature value space (FS1).

[0090] According to this aspect, there is an advantage that, when the learning of the classifier ( F1 ) is insufficient, it is less likely to erroneously determine that the target data is not classified into the new category.

[0091] In a classification system (100) according to a sixth aspect, under the fifth aspect, the determination unit (F3) excludes the target data from the candidate list when the following condition is satisfied: none of the two or more classifiers (F11 to F1n) corresponding to the two or more corresponding categories focuses on the same portion of the target data.

[0092] This method has the advantage of easily improving the accuracy of determining whether or not there is a possibility that the target data is classified into a new category.

[0093] In the classification system (100) according to the seventh aspect, in any one of the fourth to sixth aspects, if the number of candidate object data exceeds a predetermined number, the determination unit (F3) determines that the object data is classified into a new category.

[0094] This method has the advantage of easily improving the accuracy of determining whether or not object data is classified into a new category.

[0095] The classification system (100) according to the eighth aspect is further provided with a storage unit (13) in the second aspect. The storage unit (13) stores the target data and the determination result of the determination unit (F3) on the target data.

[0096] According to this embodiment, there is an advantage in that the presenting unit (12) can present the past determination results stored in the storage unit (13).

[0097] In the classification system (100) according to the ninth aspect, in the eighth aspect, the determination result is associated with the group (G0) corresponding to the similarity of the feature amount of the target data and stored in the storage unit (13).

[0098] This method has an advantage in that it is easier to classify the object data into an appropriate new category than when the object data is stored in association with one group (G0).

[0099] In the classification system (100) according to the tenth aspect, in the ninth aspect, the presentation unit (12) presents the determination result in a manner corresponding to the number of data belonging to the group (G0).

[0100] According to this method, there is an advantage that, by confirming the information presented by the presentation unit (12), the user can easily understand the degree of possibility that the object data is classified into the new category.

[0101] In the classification system (100) according to the eleventh aspect, in any one of the first to tenth aspects, the classifier (F1) relearns the object data determined by the determination unit (F3) to be classified into the new category as teaching data.

[0102] According to this embodiment, there is an advantage in that a new category can be added to the classification unit ( F1 ), and improvement in the classification accuracy of the object data classified by the classification unit ( F1 ) can be expected.

[0103] In the classification system (100) according to the twelfth aspect, in the second aspect or any one of the eighth to tenth aspects, the presentation unit (12) displays the reliability and feature value space (FS1) of the object data in a plurality of categories.

[0104] According to this embodiment, there is an advantage that the user can easily understand the determination result of the determination unit (F3) by observing the image displayed by the presentation unit (12).

[0105] The classification method according to the thirteenth embodiment includes an input acceptance step (S1), a classification step (S3), a calculation step (S2), and determination steps (S4 to S8). The input acceptance step (S1) is a step of accepting input of object data. The classification step (S3) is a step of classifying the object data into any one of a plurality of categories. The calculation step (S2) is a step of calculating a feature value of the object data. The determination steps (S4 to S8) are steps of determining the likelihood that the object data is classified into a new category different from the plurality of categories based on the classification result in the classification step (S3) and the feature value of the object data calculated in the calculation step (S2).

[0106] This method has the advantage that even when object data that may belong to a new category that does not exist in the existing categories is input, it is easy to appropriately classify the object data.

[0107] The classification method according to the fourteenth aspect is a method according to the thirteenth aspect, wherein the presenting step ( S9 ) is a step of presenting the determination results of the determining steps ( S4 to S8 ) when the determining steps ( S4 to S8 ) determine that there is a possibility that the object data is classified into a new category.

[0108] The program according to the fifteenth aspect causes one or more processors to execute the classification method according to the thirteenth aspect or the fourteenth aspect.

[0109] This method has the advantage that even when object data that may belong to a new category that does not exist in the existing categories is input, it is easy to appropriately classify the object data.

[0110] The configurations according to the second to twelfth aspects are not essential to the classification system (100) and can be omitted as appropriate.

[0111] Industrial applicability

[0112] The classification system, classification method, and program disclosed herein have the advantage of being able to easily and appropriately classify object data even when object data is input and may belong to a new category that does not already exist. Therefore, the classification system, classification method, and program disclosed herein can be utilized for, for example, object inspection and are therefore industrially useful.

[0113] Explanation of symbols

[0114] 11 Input acceptance department;

[0115] 12 output unit (prompt unit);

[0116] 13. Storage Department;

[0117] 100 classification system;

[0118] F1 Classification Department;

[0119] F11~F1n classifiers;

[0120] F2 Computing Department;

[0121] F3 Judgment Department;

[0122] FS1 feature space;

[0123] G0, G1, G2, G3 groups;

[0124] S1 input acceptance steps;

[0125] S2 calculation steps;

[0126] S3 classification step;

[0127] S4~S8 determination steps;

[0128] S9 Tips Steps.

Claims

1. A classification system comprising: Input acceptance department, accepting input of object data; a classification unit having a plurality of classifiers, and classifying the object data into any one of a plurality of categories; a calculation unit that calculates a feature value of the object data; and a determination unit that determines, based on the classification result of the classification unit and the feature amount of the object data calculated by the calculation unit, a possibility that the object data is classified into a new category different from the plurality of categories; The plurality of categories include a plurality of corresponding categories corresponding to the plurality of classifiers included in the classification unit and other categories not corresponding to any of the plurality of classifiers. The determination unit distinguishes and determines whether the target data is classified into the other category or the new category. When the object data is classified into two or more corresponding categories among the plurality of corresponding categories and the feature quantity of the object data is between the two or more corresponding categories in the feature quantity space, the determination unit considers the object data as a candidate for classification into the new category, The determination unit excludes the object data from the candidates when any of the two or more classifiers corresponding to the two or more corresponding categories does not focus on the same portion of the object data.

2. The classification system according to claim 1, wherein: Also features: The presenting unit presents a determination result of the determining unit when the determining unit determines that there is a possibility that the object data is classified into the new category.

3. The classification system according to claim 1 or 2, wherein: The classification unit includes the plurality of classifiers and a NOR circuit.

4. The classification system according to claim 1 or 2, wherein: When the object data is classified into the other category, the credibility in any of the multiple corresponding categories is lower than a threshold, and the feature value of the object data is more than a given distance away from any of the multiple corresponding categories in the feature value space, the judgment unit regards the object data as a candidate for classification into the new category.

5. The classification system according to claim 4, wherein: If the number of the object data serving as the candidates exceeds a given number, the determination unit determines that the object data is classified into the new category.

6. The classification system according to claim 2, wherein: Also features: The storage unit stores the target data and the determination result of the determination unit with respect to the target data.

7. The classification system according to claim 6, wherein: The determination result is associated with a subgroup corresponding to the similarity of the feature amount of the object data and stored in the storage unit.

8. The classification system according to claim 7, wherein: The presenting unit presents the determination result in a format corresponding to the number of the target data belonging to the group.

9. The classification system according to claim 1 or 2, wherein: The classifier relearns the object data determined by the determiner to be classified into the new category as teaching data.

10. The classification system according to any one of claims 2, 6 to 8, wherein: The presentation unit displays the reliability and feature value space of the object data in the plurality of categories.

11. A classification method for classifying object data in a plurality of classifiers, comprising: An input acceptance step of accepting input of the object data; a classification step of classifying the object data into any category among a plurality of categories; a calculation step of calculating a feature quantity of the object data; and a determination step of determining, based on the classification result in the classification step and the feature amount of the object data calculated in the calculation step, a possibility that the object data is classified into a new category different from the plurality of categories; The plurality of categories include a plurality of corresponding categories corresponding to the plurality of classifiers respectively and other categories not corresponding to any of the plurality of classifiers. In the determination step, it is determined whether the target data is classified into the other category or the new category. In the determination step, when the object data is classified into two or more corresponding categories among the plurality of corresponding categories and the feature quantity of the object data is between the two or more corresponding categories in the feature quantity space, the object data is considered as a candidate for classification into the new category. If any of the two or more classifiers corresponding to the two or more corresponding categories does not focus on the same portion of the object data, the object data is excluded from the candidates in the determination step.

12. The classification method according to claim 11, wherein: Also includes: The presenting step is a step of presenting a determination result of the determining step when the determining step determines that there is a possibility that the object data is classified into the new category.

13. A program product comprising a program for causing one or more processors to execute the classification method according to claim 11 or 12.

Images