JP2021182372A - Program providing method, program providing device, and computer program - Google Patents

Abstract

To provide a program providing method for supporting a user to smoothly introduce a defect detection program including a learning model.SOLUTION: A defect detection program, which includes an unlearned learning model and detects a defect of an object to be detected by making the learning model learn, is distributed to a user's information processor, the information processor verifies detection accuracy of a learning model learned using machine learning by a user who temporarily introduces the distributed defect detection program, the information processor transmits verification result data of detection accuracy, the verification result data transmitted from the information processor is acquired, and information on an actual introduction plan of the defect detection program is transmitted to the information processor on the basis of the acquired verification result data.SELECTED DRAWING: Figure 7

Description

The present invention relates to a program providing method, a program providing device, and a computer program.

In recent years, work using artificial intelligence and machine learning can be expected to be applied in a wide range of fields such as image recognition, voice recognition, personal authentication, automatic driving and behavior prediction, so many companies (users) are considering introducing it. Is going.

However, there are many cases where a certain amount of technical know-how is required in order to properly learn the learning model and ensure the required accuracy. Depending on the purpose and accuracy of the business, it may be relatively easy to introduce the program, but it may be difficult to introduce it. In addition, a certain amount of technical know-how is required to determine whether or not the learning model is properly learned. When a user who is not familiar with machine learning introduces a program using machine learning, it is difficult to accurately judge whether or not it is better to receive the support of an expert.

Japanese Unexamined Patent Publication No. 2018-195119

An object of the present invention is to provide a program providing method, a program providing device, and a computer program that enable a user to make a proposal to a user for smoothly introducing a program including a learning model.

The program providing method according to the present invention distributes a program including an unlearned learning model to a user's information processing apparatus, and the accuracy of the output data of the learning model machine-learned by the user who temporarily introduced the distributed program. The data related to the above is acquired, and the information regarding the present introduction plan of the program is output based on the acquired data.

The program providing device according to the present invention determines the accuracy of the output data of the learning model machine-learned by the distribution unit that distributes the program including the unlearned learning model to the user and the user who temporarily introduced the distributed program. An acquisition unit for acquiring the relevant data and an output unit for outputting information regarding the main introduction plan of the program based on the acquired data are provided.

The computer program according to the present invention distributes a program including an unlearned learning model to a user, and acquires data related to the accuracy of the output data of the learning model machine-learned by the user who temporarily introduced the distributed program. Then, based on the acquired data, the computer is made to execute a process of outputting information regarding the main introduction plan of the program.

According to the present invention, it is possible to support the smooth introduction by the user of the program including the learning model.

It is a schematic diagram explaining the configuration example of the program providing system which concerns on Embodiment 1. FIG. It is a block diagram which shows the configuration example of the server which concerns on Embodiment 1. FIG. It is a block diagram which shows the defect detection apparatus. It is a conceptual diagram which shows the defect detection method using the image data of the inspection object. It is a conceptual diagram which shows the defect detection method using the time-series data obtained by operating an inspection object. It is a conceptual diagram which shows the record layout of a user DB. It is a flowchart which shows the processing procedure of the program provision method which concerns on Embodiment 1. It is a frequency distribution showing the verification result of the detection accuracy of defective products using a learning model. This is an example of a display screen of the present introduction plan according to the first embodiment. It is a flowchart which shows the processing procedure of the program provision method which concerns on Embodiment 2. It is an example of the display screen of this introduction plan which concerns on Embodiment 2. It is a flowchart which shows the processing procedure of the program provision method which concerns on Embodiment 3. It is a flowchart which shows the processing procedure of the program provision method which concerns on Embodiment 4. It is a conceptual diagram which shows an example of the plan learning model which concerns on Embodiment 5.

Specific examples of the program providing method, the program providing apparatus, and the computer program according to the embodiment of the present invention will be described below with reference to the drawings. It should be noted that the present invention is not limited to these examples, and is indicated by the scope of claims, and is intended to include all modifications within the meaning and scope equivalent to the scope of claims. In addition, at least a part of the embodiments described below may be arbitrarily combined.
Hereinafter, the present invention will be specifically described with reference to the drawings showing the embodiments thereof.

(Embodiment 1)
FIG. 1 is a schematic diagram illustrating a configuration example of the program providing system according to the first embodiment. The program providing system according to the first embodiment includes a server (program providing device) 1, an information processing device 2 of a user, and a terminal device 3 of a person in charge of supporting the introduction of the program. Each device is connected to the server 1 via a communication network N such as the Internet. In the following, as an example of providing a program in this system, a case where a defect detection program is provided will be described.

The server 1 can communicate with the information processing device 2 and the terminal device 3, and supports the user to introduce the defect detection program 5. Specifically, the server 1 distributes the defect detection program 5 to the user free of charge. The defect detection program 5 is installed in the user's information processing device 2. The defect detection program 5 includes an unlearned learning model 51 (see FIG. 2). For example, the user gives image data of a non-defective product as learning data to the learning model 51 to learn the characteristics of the non-defective product. Then, the user's information processing device 2 verifies the detection accuracy of the defective product using the learned learning model 51, and determines the accuracy of the verification result data (the accuracy of the output data of the learning model machine-learned by the provisionally introduced user). An example of such data) is transmitted to the server 1. The verification result data includes, for example, information regarding the discrepancy between the data obtained by using the learning model 51 machine-learned by the user and the target data. The server 1 receives the verification result data, and based on the verification result, transmits information about the present introduction plan of the defect detection program 5 according to the learning state of the learning model 51 to the information processing apparatus 2. The information processing apparatus 2 receives information about the introduction plan and displays the information. The user can examine the present introduction plan displayed on the information processing apparatus 2 and examine the introduction method of the defect detection program 5. The content of this introduction plan is generated according to the learning state of the learning model 51, and proposes to the user for the user to smoothly introduce the defect detection program 5 by using the information processing device 2. It can be carried out.

FIG. 2 is a block diagram showing a configuration example of the server 1 according to the first embodiment. The server 1 is a computer including a processing unit 11, a storage unit 12, and a communication unit 13. The server 1 may be a multi-computer composed of a plurality of computers, or may be a virtual machine virtually constructed by software.

The processing unit 11 has one or a plurality of arithmetic processing units (Central Processing Units), MPUs (Micro-Processing Units), and the like, and reads and executes the computer program 4 stored in the storage unit 12. The process related to the distribution of the defect detection program 5, the provision of the main introduction plan of the defect detection program 5, the main introduction support, and the like is executed.

The storage unit 12 is a storage device such as a hard disk, EEPROM (Electrically Erasable Programmable ROM), and a flash memory. The storage unit 12 stores the computer program 4, the defect detection program 5, and the user DB 6 necessary for the processing unit 11 to execute the process related to the provision of the defect detection program 5. The defect detection program 5 is a program for causing the information processing apparatus 2 to execute the defect detection process. Some functions of the defect detection program 5 are restricted. The function restriction can be canceled by the release key. The release key is an example of information for enabling the defect detection program 5 whose function of deleting the parameter is canceled, in other words, information for enabling the program whose function is canceled. The user DB 6 is a database in which the defect detection program 5 is installed and the information of the user who has been temporarily installed or has been installed is stored. The storage unit 12 may be an external storage device connected to the server 1.

The computer program 4 is a program that causes the computer to function as the program providing device according to the first embodiment. The computer program 4 causes the computer to execute the distribution process of the defect detection program 5, the process related to the generation and provision of the main introduction plan of the defect detection program 5, the process related to the present introduction support, and the like.
The computer program 4 may be recorded on the recording medium 7 so that it can be read by a computer. The storage unit 12 stores the computer program 4 read from the recording medium 7 by a reading device (not shown). The recording medium 7 is a semiconductor memory such as a flash memory, an optical disk, a magnetic disk, a magnetic disk disk, or the like. Further, the computer program 4 according to the present embodiment may be downloaded from an external server (not shown) connected to the communication network N and stored in the storage unit 12.

The communication unit 13 includes a processing circuit, a communication circuit, and the like for performing processing related to communication, and transmits / receives information between the information processing device 2 of the user and the terminal device 3 of the person in charge.

The user shown in FIG. 1 is, for example, a manufacturer of mass-produced products. The mass-produced product is the inspection target in the first embodiment. The information processing device 2 is a computer including a calculation unit, a main storage unit, an auxiliary storage unit, a display device, an input device, a communication circuit, and an input terminal (not shown). The arithmetic unit is an arithmetic processing circuit such as one or a plurality of CPUs, MPUs, GPUs (Graphics Processing Units), GPGPUs (General-purpose computing on graphics processing units), and TPUs (Tensor Processing Units). The communication circuit transmits / receives data to / from the server 1. The information processing apparatus 2 receives the defect detection program 5 delivered from the server 1 by the communication circuit. The auxiliary storage unit is a non-volatile storage device such as a hard disk or a semiconductor memory, and stores the defect detection program 5 received by the communication circuit. The arithmetic unit reads and executes the defect detection program 5 stored in the auxiliary storage unit to perform information processing related to defect detection. The main storage unit is a temporary storage area for SRAM (Static Random Access Memory), DRAM (Dynamic Random Access Memory), flash memory, etc., and temporarily stores data necessary for the arithmetic unit to execute arithmetic processing. .. The display device is a display device such as a liquid crystal display or an organic EL (Electro Luminescence) display, and displays an image given by a calculation unit. The input device is a keyboard, a mouse, or the like, and accepts user operations and inputs.
A data acquisition device is connected to the input terminal of the information processing device 2. The data acquisition device is, for example, a camera that captures an image of an inspection object and a sensor that detects the operation of the inspection object. The sensor detects vibration, acceleration, current, voltage, etc. of the inspection object, and outputs time-series data indicating the operation of the inspection object. The image data and time series data output from the data acquisition device are input to the information processing device 2.

The person in charge is an expert, an engineer, or the like of the defect detection program 5, and supports the introduction of the defect detection program 5 by the user. The terminal device 3 is a computer, a smartphone, or the like similar to the information processing device 2, and has a function of communicating with the server 1. The terminal device 3 may also be a multi-computer including a plurality of computers, or may be a virtual machine virtually constructed by software.

FIG. 3 is a block diagram showing a defect detection device, and FIG. 4 is a conceptual diagram showing a defect detection method using image data of an inspection object. The information processing device 2 in which the defect detection program 5 is installed functions as a defect detection device. The information processing device 2 as a defect detection device includes a learning model 51, a defect degree calculation unit 52, a quality determination unit 53, and a learning processing unit 54.

The learning model 51 is, for example, an autoencoder. When the image data of the inspection object of a good product is input to the learning model 51, the characteristics of the inspection object included in the captured image are extracted, and the image data of the self-reproduced image in which the characteristics of the inspection object are reproduced is output.

The learning model 51 has an input layer 51a, an output layer 51c, and an intermediate layer 51b. The intermediate layer 51b has a convolution layer (CONV layer) and a deconvolution layer (DECONV layer). The input layer 51a is a layer to which data of each pixel value related to the image data of the inspection target is input. The convolution layer of the intermediate layer 51b is a layer for dimensionally compressing image data. The feature quantity of the inspection object is extracted by the dimensional compression. The deconvolution layer of the intermediate layer 51b is a layer that restores the dimension-compressed data in the convolution layer to the original dimension. By the restoration, the original characteristics of the inspection object, that is, the image data showing the characteristics of the inspection object of a non-defective product are restored. The output layer 51c is a layer that outputs data of each pixel value related to the self-reproduced image from which the characteristics of the inspection target are extracted by the convolution layer and the deconvolution layer. The image data output from the output layer 51c is input to the defect degree calculation unit 52.

The learning processing unit 54 is a neural network of the learning model 51 so that the image data of the non-defective inspection object input to the learning model 51 and the output image data are the same as shown by the broken line in FIG. Make the network machine-learn. Specifically, various parameters such as the weighting coefficient (coupling coefficient) between the neurons constituting the neural network are set to the error back propagation method and the steepest descent method so that the input image of the non-defective product and the self-reproduced image are the same. Etc. to optimize.

When the image data of the inspection object of a good product is input to the learning model 51 learned in this way, the image data of the self-reproduced image having substantially the same characteristics as the input image data is output. When the image data of the defective inspection target is input to the learning model 51, the image data of the self-reproduced image having characteristics different from the input image data is output.

The defect degree calculation unit 52 calculates the difference between the image data of the inspection object and the image data of the self-reproduced image output from the learning model 51, and outputs the difference as the defect degree. The degree of defect is input to the quality determination unit 53.

As shown in FIG. 4A, the pass / fail determination unit 53 determines that the inspection object is inspected when the difference between the image data of the inspection object and the image data of the self-reproduced image output from the learning model 51 is less than a predetermined threshold value. Judged as a good product. As shown in FIG. 4B, the quality determination unit 53 determines that the inspection target is a defective product when the difference is equal to or greater than a predetermined threshold value. Then, the pass / fail determination unit 53 outputs the pass / fail determination result.

FIG. 5 is a conceptual diagram showing a defect detection method using time-series data obtained by operating an inspection object. In the above explanation, an example of detecting a defective product of the inspection target using the image data of the inspection target has been described, but the inspection target is inspected using the time series data obtained when the inspection target is operated. It may be configured to judge the quality of an object. If the waveform represented by the time series data is treated as an image, the quality can be determined in the same manner as described above.
As shown in FIG. 5A, the quality determination unit 53 indicates that the inspection target is a good product when the difference between the waveform of the time series data obtained from the inspection target and the self-reproduced waveform is less than a predetermined threshold value. Is determined. As shown in FIG. 5B, the quality determination unit 53 determines that the inspection target is a defective product when the difference is equal to or greater than a predetermined threshold value. Then, the pass / fail determination unit 53 outputs the pass / fail determination result.
Although the example of handling the time series data as an image has been described, the learning model 51 may be configured by using RNN (Recurrent Neural Network) and LSTM (Long short-term memory).

The learning processing unit 54 has a function of verifying the defect detection accuracy of the inspection object by using the learning model 51 learned as described above. For example, the learning processing unit 54 calculates evaluation values such as a correct answer rate, a precision rate, a recall rate, and an F value. Further, the learning processing unit 54 may create a frequency distribution of the degree of defect (see FIG. 8) calculated based on the image data of a plurality of non-defective products input to the learning model 51 and the image data of a plurality of defective products. good. The information processing apparatus 2 transmits the verification result data of the detection accuracy of the learning model 51 to the server 1 at an appropriate timing. The verification result data is data showing the difference between the degree of defect of the inspection object obtained by using the learning model 51 machine-learned by the user and the actual degree of defect of the inspection object, and is information about the inspection object. Does not include. For example, the verification result data is information such as the above index and frequency distribution. Further, the defect degree data calculated from the image data of the plurality of non-defective products and the defect degree data calculated from the image data of the plurality of defective products may be transmitted to the server 1 as verification result data. The information processing apparatus 2 is supposed to transmit the verification result data of the detection accuracy of the learning model 51 to the server 1 at an appropriate timing. However, for example, when the user performs the learning process a predetermined number of times, the user completes the learning. It is possible to enable transmission when the temporary introduction period of the defect detection program 5 has elapsed, or when the temporary introduction period of the defect detection program 5 has expired. Further, the number of transmissions is not limited to once, and may be transmitted a plurality of times. For example, if the user does not perform the present introduction, the transmission may not be performed. Further, it may be possible to change as appropriate depending on the content to be learned and the content requested by the user.

The information processing apparatus 2 includes information such as the type of the learning model 51, the number of intermediate layers 51b constituting the learning model 51, the number of neurons, the type of activation function, the number of training data, the number of learnings, and the verification result data. The number of image data used for verification, the type of data acquisition device used for imaging the inspection target, the method of preprocessing for the image data or time series data to be input to the learning model 51, etc. are included and transmitted to the server 1. It may be configured to do so. By transmitting this information to the server 1, the server 1 can recognize the learning state of the learning model 51 in more detail.

The timing of transmitting the verification result data is not particularly limited. The information processing apparatus 2 transmits, for example, verification result data when a transmission instruction is received from a user. Further, the information processing apparatus 2 may be configured to transmit the verification result data after a predetermined time has elapsed after the start of learning. The information processing apparatus 2 does not transmit the image data itself of the inspection target, the type of the inspection target, and other information that can identify the inspection target to the server 1 as the verification result data.

FIG. 6 is a conceptual diagram showing the record layout of the user DB6. The user DB 6 is, for example, a relational database and has a plurality of fields for storing information about the user. For example, the user DB 6 stores information such as a record number, a user ID, a software ID, a verification result of the learning model 51, a recommended plan, an inspection target, a company name, a user name, an e-mail address, and a contract date. The software ID is an ID for identifying the defect detection program 5 delivered to the user. The verification result is data showing the result of verifying the detection accuracy of the learning model 51. The recommended plan is data indicating the main introduction plan of the defect detection program 5 recommended to the user. The contract date is the date data on which the contract for the main introduction of the defect detection program 5 is concluded.

FIG. 7 is a flowchart showing a processing procedure of the program providing method according to the first embodiment. The processing unit 11 of the server 1 includes the unlearned learning model 51, and distributes the defect detection program 5 for learning the learning model 51 to detect the defect of the inspection object to the information processing apparatus 2 of the user (step). S11). When the defect detection program 5 is transmitted to the information processing apparatus 2 of a specific user, the server 1 acquires the basic information of the user and records the acquired user information in the user DB 6. The user's basic information can be obtained, for example, by having the user input when applying for the provision of the defect detection program 5. The user information is, for example, information such as a user ID, a user name, and an e-mail address. Further, when the defect detection program 5 is distributed, the server 1 may record the software ID of the defect detection program 5 in the user DB 6 in association with the user ID. Further, when the defect detection program 5 is distributed, the server 1 may transmit information on the method of using the defect detection program 5 or the method of machine learning to the information processing apparatus 2 of the user. The information regarding the method of use or the method of machine learning is, for example, information indicating a method of preprocessing for image data or time series data input to the learning model 51, an example of preprocessing, and the like. The server 1 may transmit the URL in which the information is stored to the information processing apparatus 2 of the user. Further, when performing machine learning in step S13 described later, information regarding the method of using the defect detection program 5 or the method of machine learning described above may be displayed or notified to the user.

The user's information processing apparatus 2 downloads the defect detection program 5 transmitted from the server 1 and installs the downloaded defect detection program 5 (step S12). The defect detection program 5 downloaded here has some functions limited. By installing the defect detection program 5 whose functions are restricted in the information processing apparatus 2, the defect detection program 5 is temporarily introduced. The defect detection program 5 may be used only for a predetermined period from the time when the learning described later is started. In this embodiment, it can be used for one month. Further, the period of use of the defect detection program 5 may not be limited. In this case, after a predetermined period (for example, one month) has elapsed from the start of use, the information processing apparatus 2 of the user is notified by the program provider (server 1). , Notifications such as status confirmation may be given.

Next, the information processing apparatus 2 in which the defect detection program 5 is installed performs machine learning of the learning model 51 (step S13). Specifically, the user prepares a plurality of image data of good products of the inspection target. The information processing apparatus 2 extracts an image portion of an inspection object from image data and executes preprocessing such as adjusting the image size. Then, the learning processing unit 54 of the information processing apparatus 2 inputs the preprocessed good image data into the learning model 51, and sets the learning model 51 so that the output self-reproduced image and the input image are the same. Let them learn by machine.

Next, the learning processing unit 54 of the information processing apparatus 2 verifies the detection accuracy of the learning model 51 after learning (step S14), and displays the verification result (step S15). The user confirms the verification result, and repeatedly executes the machine learning and verification processing of the learning model 51 as necessary. The machine learning and verification loop processing may be configured so that the information processing apparatus 2 automatically executes the machine learning and verification loop processing based on the verification result. In the present embodiment, as a limitation of a part of the functions of the program, the machine learning and verification processes delete the parameters of the learning model 51 when the verification is completed, and when learning again, learn from scratch. There is. That is, the learned learning model 51 cannot be held. It should be noted that additional learning may be possible after the verification is completed without setting such functional restrictions.

FIG. 8 is a frequency distribution showing the verification result of the detection accuracy of defective products using the learning model 51. In FIG. 8, the horizontal axis indicates the degree of defect, and the vertical axis indicates the frequency (number) of the inspection object having a specific degree of defect. The white bar indicates the inspection target of the non-defective product, and the black bar indicates the inspection target of the defective product.
FIG. 8A shows a state in which the learning model 51 has high detection accuracy of defective products and machine learning is properly performed. The degree of defect output when the image data of a non-defective product is input is generally concentrated around the degree of defect X, and the degree of defect output when the image data of a defective product is input is generally the degree of defect. It is concentrated around Y. In this case, by setting the threshold value as shown in FIG. 8A, it becomes possible to determine the quality of the inspection object. That is, if the degree of defect output from the learning model 51 is equal to or greater than the threshold value, it can be determined to be a defective product, and if the degree of defect is less than the threshold value, it can be determined to be a non-defective product.

On the other hand, FIG. 8B shows a state in which the accuracy of detecting defective products by the learning model 51 is low and machine learning is not properly performed. The degree of defect output when image data of a non-defective product is input is also distributed in a portion distant from the periphery of the degree of defect X. Further, the degree of defect output when the image data of the defective product is input is also distributed in the portion distant from the degree of defect Y. In this case, as shown in FIG. 8B, it is not possible to set a threshold value for discriminating between a non-defective product and a defective product, and it is impossible to accurately determine the quality of the inspection target. The frequency distribution shown in FIG. 8B is an example, and the shape of the frequency distribution differs depending on the learning state of the learning model 51.

The information processing apparatus 2 that has completed the process of step S15 transmits the verification result data and the user information to the server 1 at an appropriate timing (step S16). In the present embodiment, when the verification result is obtained, the verification result data is transmitted to the server 1. That is, even if the user does not apply for this introduction plan, the verification result data is transmitted when the verification result is output. The user information includes at least a user ID. Further, as user information, the software ID, the type of inspection target, the company name, the user name, and the e-mail address may be transmitted to the server 1.

The server 1 receives the verification result data and the user information transmitted from the information processing apparatus 2 (step S17). Then, the server 1 generates the main introduction plan regarding the main introduction plan of the defect detection program 5 based on the verification result data (step S18). This introduction plan includes a plurality of plans in which the burden on the user related to the main introduction of the program is different. For example, an in-house plan (an example of the first plan) in which the user himself / herself performs the main introduction of the defect detection program 5 and a defect detection. It includes a subscription plan (an example of the second plan) in which the defect detection program 5 is actually introduced while receiving the support of the subscription format by the person in charge who supports the introduction of the program 5. In this case, the in-house production plan has a higher burden on the user related to the main introduction of the defect detection program 5 than the subscription plan. The server 1 recommends an in-house production plan when the discrepancy between the degree of defect of the inspection object obtained by using the learning model 51 and the actual degree of defect of the inspection object is small (when the accuracy of the output data is high). Generate the main introduction plan as a plan, and if the deviation is large (when the accuracy of the output data is low), generate the main introduction plan with the subscription plan as the recommended plan.
The in-house plan provides the user with a release key for releasing the functional restriction of the defect detection program 5, and the user himself performs learning, verification, defect inspection, and maintenance of the learning model 51 using the defect detection program 5. Is.
The subscription plan is a plan that provides the user with a release key for releasing the functional restriction of the defect detection program 5, and also provides the user with a certain amount of support regarding learning, verification, defect inspection, and maintenance of the learning model 51. .. For example, the user can send a question or consultation to the information processing apparatus 2 by using an e-mail, a chat box, or the like, and can receive answers, advice, and the like from an expert or engineer of the defect detection program 5. The subscription plan includes support by telephone and other means of communication by the person in charge.
In this embodiment, the server 1 generates the introduction plan based on the verification result data, but a person (for example, a person in charge) creates the introduction plan based on the verification result data and creates the introduction plan. Then, the information processing device 2 or the user, which is the transmission source of the verification result data, may be notified.

In addition, the introduction plan generated in step S18 may include information indicating whether to recommend an in-house plan or a subscription plan. The information is textual information such as "(recommended)". When the deviation between the degree of defect of the inspection object obtained by using the learning model 51 machine-learned by the user and the actual degree of defect of the inspection object is small (when the accuracy of the output data is high), the server 1 is used. Generate this introduction plan with the in-house production plan as the recommended plan. When the deviation is large (when the accuracy of the output data is low), the server 1 generates the introduction plan with the subscription plan as the recommended plan.

The server 1 may select a data acquisition device suitable for acquiring image data based on the verification result data, and may include information on the selected data acquisition device in the present introduction plan. The user selects whether or not he / she wants to use the selected data acquisition device, and the information processing device 2 transmits the presence / absence of an application for the data acquisition device to the server 1.

The server 1 transmits the present introduction plan generated according to the verification result data to the information processing apparatus 2 that is the transmission source of the verification result data (step S19). The server 1 can specify the destination of the introduction plan from the user information transmitted together with the verification result data. When the user ID is transmitted from the information processing apparatus 2 as the user information, the server 1 can specify the e-mail address corresponding to the user ID by referring to the user information registered in the user DB 6, and is specified. You can send this introduction plan using the email address provided. The e-mail address is an example of the destination of the introduction plan, and the transmission route and transmission method to the user of the introduction plan are not particularly limited. As another example, the e-mail address is stored in a recording medium or the like and mailed. It may be sent by such as.

The user's information processing apparatus 2 receives the present introduction plan transmitted from the server 1 (step S20), and displays the received present introduction plan (step S21). If the user does not introduce the introduction plan, the process ends in step S21. Further, in step S21, the option of not introducing the present introduction plan may be displayed together with the display of the present introduction plan. In this case, if the user selects the option of not introducing this introduction plan, the defect detection program 5 may be disabled. Further, if the defect detection program 5 is not actually introduced, the server 1 may be configured to delete the verification result data. The server 1 may be configured to delete the verification result data when the introduction plan is generated in step S18. Further, it may be configured to notify the user that the verification result data has been deleted.

FIG. 9 is an example of a display screen of the present introduction plan according to the first embodiment. The introduction plan display screen 8 includes characters that introduce the subscription plan and characters that introduce the in-house production plan. Further, the introduction plan display screen 8 includes a radio button 8a for selecting a subscription plan or an in-house production plan.
FIG. 9A is an example of the introduction plan display screen 8 that recommends the subscription plan, and the subscription plan is displayed at the upper part and includes the characters “(recommended)”.
FIG. 9B is an example of the main introduction plan display screen 8 that recommends the in-house production plan, and the in-house production plan is displayed at the upper part and includes the characters “(recommended)”.

Next, the information processing apparatus 2 receives the selection and application information of the present introduction plan from the user (step S22). The user can select this introduction plan by selecting the radio button 8a. The information processing device 2 transmits the received application information to the server 1 (step S23). The application information includes information indicating the type of the introduction plan, that is, either the in-house plan or the subscription, the type of the inspection target, the user information that has not been acquired, and the like. There is a charge for services related to this introduction plan.
In the present embodiment, the plan can be selected by the radio button 8a, but the user may notify the server 1 or the person in charge of which plan to select by e-mail or the like.

The server 1 receives the application information received from the information processing device 2 (step S24). The information processing apparatus 2 records the received application information in the user DB 6, and transmits the function restriction release key of the defect detection program 5 to the information processing apparatus 2 (step S25).

The information processing apparatus 2 receives the release key transmitted from the server 1 (step S26), and releases the function restriction by the received release key (step S27). The release key cancels the function restriction of the defect detection program 5.
When a subscription is applied for, the server 1 provides a subscription-type support service to the user according to a processing procedure (not shown).
Further, the release key may not be transmitted. In this case, a customized program may be provided based on the customer's request, verification result data, and the like. The program is a program in which the above-mentioned functional restrictions are released or the above-mentioned functional restrictions are not present.

According to the program providing method, the program providing device, and the computer program 4 according to the first embodiment configured in this way, it is possible to propose to the user for the user to smoothly introduce the defect detection program 5 including the learning model 51. It becomes.

Specifically, the server 1 can temporarily introduce and present the user with the recommended introduction plan according to the verification result of the learning model 51 learned by the user. The user can select the present introduction plan suitable for the learning situation of the learning model 51 and introduce the defect detection program 5 with reference to the recommended contents. Therefore, the user can introduce the defect detection program 5 with the necessary support. Further, when the user has succeeded in learning the learning model 51, it is not necessary to apply for unnecessary support, and the defect detection program 5 can be introduced at low cost.

Further, since the specific information of the inspection target is not transmitted from the information processing device 2 to the server 1 until the defect detection program 5 is introduced, confidential information about the inspection target is transmitted or leaked. There is no such thing. Only the abstract verification result data is transmitted from the information processing device 2 to the server 1. Therefore, the user can acquire the information on the recommended introduction plan and consider the method of the introduction without worrying about the leakage of the confidential information on the inspection target.

In the first embodiment, as the learning model 51, an autoencoder, RNN, LSTM, etc. that output a self-reproduced image are exemplified, but CNN (Convolution Neural) that has been machine-learned to output the determination results of good and defective products is used. Network), RNN, LSTM, etc. may be used to configure the learning model 51. Further, a learning model 51 having a configuration such as another neural network, SVM (Support Vector Machine), Bayesian network, or regression tree may be configured.

Further, in the first embodiment, an example in which one defect detection program 5 is introduced by one specific plan for one user has been described, but the method of providing the defect detection program 5 is limited to this. It's not a thing. For example, the number of learning models 51 that can use the defect detection program 5 with one license may be limited to a predetermined number. When it is desired to detect defects of a plurality of types of inspection objects using more than a predetermined number of learning models 51, the user applies for a plurality of plans. The server 1 may transmit a plurality of release keys to the information processing device 2 of the user in response to a plurality of applications from one user. In addition, one user may apply for a different introduction plan. For example, the server 1 may be configured to accept a subscription plan applied by one user and an in-house production plan.

Further, if the display of the verification result is not satisfied, the information processing apparatus 2 may be configured to send inquiry information to the server 1.
Further, if the introduction is not applied, the server 1 may be configured to notify the terminal device 3 of the person in charge of that fact and the user information. The person in charge can contact the user, solve the technical problem related to the introduction of the defect detection program 5, and promote the introduction.

(Embodiment 2)
Since the program providing method, the server 1 and the computer program 4 according to the second embodiment are different from the first embodiment in the content and the providing method of the introduction plan, the above differences will be mainly described below. Since other configurations and actions and effects are the same as those in the first embodiment, the corresponding parts are designated by the same reference numerals and detailed description thereof will be omitted.

FIG. 10 is a flowchart showing a processing procedure of the program providing method according to the second embodiment. The processes of steps S31 to S37 are the same as those of steps S11 to S17 of the first embodiment, and the server 1 and the information processing apparatus 2 temporarily introduce the defect detection program 5, learn and verify the learning model 51, and the processing. Executes the process related to sending and receiving verification result data.

The processing unit 11 of the server 1 that has executed the processing of step S37 determines whether or not human support is required based on the verification data (step S38). For example, when the processing unit 11 has a deviation between the degree of defect of the inspection object obtained by using the learning model 51 machine-learned by the user and the actual degree of defect of the inspection object is equal to or more than the first predetermined value. Judge that human support by consulting is necessary. Human support by consulting is more generous than support by subscription, and experts and technicians of defect detection program 5 select an image pickup device for an inspection object, an image pickup method, collect learning data, and a learning model. This is to fully support the main introduction work such as learning and verification of 51.

When it is determined that human support is necessary (step S38: YES), the processing unit 11 executes the arrangement process of the person in charge (step S39). For example, the processing unit 11 transmits to the terminal device 3 of the person in charge of the defect detection program 5 that the information and support of the user who is the transmission source of the verification data are necessary. Further, the processing unit 11 records in the user DB 6 that human support is required in association with the user ID. Further, the processing unit 11 may register the person in charge ID of the person in charge in association with the user ID.
The person in charge who receives the notification contacts based on the user information, proposes a consulting service, and provides support.
In the present embodiment, it is determined whether or not the processing unit 11 of the server 1 needs human support based on the verification result data, but a person (for example, a person in charge) provides human support based on the verification result data. You may decide whether or not you need it and execute the arrangement process of the person in charge. Further, the information processing apparatus 2 or the user may be notified of the determination result of whether or not the processing unit 11 or the person needs human support before the transmission of the present introduction plan.

When it is determined in step S38 that human support is unnecessary (step S38: NO), or when the processing of step S39 is completed, the processing unit 11 is based on the verification result data, and the defect detection program 5 is used. Generate this introduction plan for the introduction plan (step S40). This introduction plan includes, for example, an in-house production plan (first plan) in which the user himself / herself performs the main introduction of the defect detection program 5, and defect detection with the support of a subscription format by a person in charge of supporting the introduction of the defect detection program 5. This includes all or part of the subscription plan (second plan) for the main introduction of the program 5 and the consulting plan (third plan) for the main introduction of the defect detection program 5 with the support of the consulting format by the person in charge. The subscription plan has a higher burden on the user related to the main introduction of the defect detection program 5 than the consult plan. That is, the consulting plan is the plan with the lowest burden on the user related to the main introduction of the defect detection program 5 among the three plans.
The subscription plan and the consulting plan may be regarded as the first plan and the second plan of the present invention, and the in-house production plan and the consulting plan may be regarded as the first plan and the second plan of the present invention. , The in-house production plan may be regarded as the first plan of the present invention, and the subscription plan and the consulting plan may be regarded as the second plan of the present invention.

When it is determined that human support is necessary, the processing unit 11 may generate the present introduction plan including at least a consulting plan. If it is determined that human support is not required, the processing unit 11 may generate the introduction plan including at least an in-house plan or a subscription plan.

In addition, the introduction plan generated in step S40 includes information indicating whether to recommend an in-house plan, a subscription plan, or a consulting plan. The server 1 provides a consulting plan when the deviation between the degree of defect of the inspection object obtained by using the learning model 51 machine-learned by the user and the actual degree of defect of the inspection object is equal to or more than the first predetermined value. Generate this introduction plan as the recommended plan. If the above deviation is less than the first predetermined value and is greater than or equal to the second predetermined value, the introduction plan with the subscription plan as the recommended plan is generated. The second predetermined value is a value smaller than the first predetermined value. If the above deviation is less than the second predetermined value, the introduction plan with the in-house production plan as the recommended plan is generated.

The server 1 transmits the present introduction plan generated according to the verification result data to the information processing apparatus 2 that is the transmission source of the verification result data (step S41).

The user's information processing apparatus 2 receives the present introduction plan transmitted from the server 1 (step S42), and displays the received present introduction plan (step S43).

FIG. 11 is a screen display example of the present introduction plan according to the second embodiment. The introduction plan display screen 8 shown in FIG. 11 shows an example when human support is required. The introduction plan display screen 8 includes characters that introduce a consulting plan, characters that introduce a subscription plan, and characters that introduce an in-house production plan. Further, the introduction plan display screen 8 includes a radio button 8a for selecting a consulting plan, a subscription plan, or an in-house production plan.
FIG. 11 is a screen display example of the present introduction plan that recommends the consulting plan, and the consulting plan is displayed at the upper part and includes the characters “(recommended)”. The subscription plan is displayed below the consulting plan, and the in-house production plan is displayed further below.

Hereinafter, the server 1 and the information processing apparatus 2 execute the same processing as in steps S22 to S27 of the first embodiment.

According to the program providing method, the program providing device, and the computer program 4 according to the second embodiment configured in this way, a person in charge can be arranged according to the verification result related to the detection accuracy of the learning model 51, which is defective. It is possible to support the smooth introduction of the detection program 5.

Further, the introduction plan including the consulting plan can be transmitted to the information processing apparatus 2 and displayed on the information processing apparatus 2.

Furthermore, if the verification result is poor, it is possible to display that a consulting plan is recommended, propose the utilization of the person in charge, and support the smooth introduction of the defect detection program 5 by the person in charge of the operation.

(Embodiment 3)
Since the program providing method according to the third embodiment, the server 1 and the computer program 4 differ from the first embodiment in the content of the functional restriction of the defect detection program 5, the above differences will be mainly described below. Since other configurations and actions and effects are the same as those in the first embodiment, the corresponding parts are designated by the same reference numerals and detailed description thereof will be omitted.

The defect detection program 5 according to the third embodiment has a function of deleting the parameter of the learning model 51 when a predetermined time elapses after the machine learning of the learning model 51, and limits the function of the defect detection program 5. There is. The parameters are weights between neurons constituting the neural network of the learning model 51 and the like.

FIG. 12 is a flowchart showing a processing procedure of the program providing method according to the third embodiment. The processing unit 11 of the server 1 distributes the defect detection program 5 to the information processing apparatus 2 of the user (step S51). The user's information processing apparatus 2 downloads the defect detection program 5 transmitted from the server 1 and installs the downloaded defect detection program 5 (step S52).

Next, the information processing apparatus 2 in which the defect detection program 5 is installed performs machine learning of the learning model 51 (step S53). The information processing apparatus 2 determines whether or not a predetermined time has elapsed after learning the learning model 51 (step S54). The predetermined time is, for example, 1 to 3 days. When it is determined that the predetermined time has elapsed (step S54: YES), the information processing apparatus 2 deletes the parameter of the learning model 51 (step S55), and returns the process to step S53. That is, the learning model 51 is in an unlearned state.

When it is determined that a predetermined time has not elapsed after learning the learning model 51 (step S54: NO), the learning processing unit 54 of the information processing apparatus 2 verifies the detection accuracy of the learning model 51 after learning (step S54: NO). S56), the verification result is displayed (step S57).

Hereinafter, the server 1 and the information processing apparatus 2 execute the same processing as in steps S16 to S27 of the first embodiment. In the process of step S25, the server 1 transmits a release key for releasing the function restriction to the information processing apparatus 2. The information processing apparatus 2 can hold the parameters of the learning model 51 by receiving the release key and releasing the function restriction. That is, the learning model 51 can be held in a trained state, and the defect detection program 5 can be practically utilized. The release key in the present embodiment is an example of information for enabling the defect detection program 5 whose function of deleting the parameter is canceled, in other words, information for enabling the program whose function is canceled.

According to the program providing method, the program providing device, and the computer program 4 according to the third embodiment, the introduction of the defect detection program 5 is effectively promoted by limiting the period for holding the learning model 51 in the learned state. Can be done.

(Embodiment 4)
Since the program providing method, the server 1 and the computer program 4 according to the fourth embodiment have different introduction plan provision timings from the first embodiment, the above differences will be mainly described below. Since other configurations and actions and effects are the same as those in the first embodiment, the corresponding parts are designated by the same reference numerals and detailed description thereof will be omitted.

FIG. 13 is a flowchart showing a processing procedure of the program providing method according to the fourth embodiment. The processing unit 11 of the server 1 distributes the defect detection program 5 to the information processing apparatus 2 of the user (step S71). The user's information processing apparatus 2 downloads the defect detection program 5 transmitted from the server 1 and installs the downloaded defect detection program 5 (step S72). Next, the information processing apparatus 2 in which the defect detection program 5 is installed performs machine learning of the learning model 51 (step S73). Next, the learning processing unit 54 of the information processing apparatus 2 verifies the detection accuracy of the learning model 51 after learning (step S74), and the information processing apparatus 2 transmits the verification result data and the user information to the server 1 (step). S75).

The server 1 receives the verification result data and the user information transmitted from the information processing apparatus 2 (step S76). Then, the server 1 generates the main introduction plan for the main introduction plan of the defect detection program 5 based on the verification result data (step S77), and the main introduction plan generated according to the verification result data is the verification result. The data is transmitted to the information processing device 2 that is the source of the data (step S78).

The user's information processing device 2 receives the present introduction plan transmitted from the server 1 (step S79). Then, the information processing apparatus 2 displays the verification result obtained in step S74 (step S80), and displays the received main introduction plan (step S81).

According to the program providing method, the program providing device, and the computer program 4 according to the fourth embodiment, the present introduction plan and the verification result are displayed in the learning process of the learning model 51, and the recommended introduction plan of the defect detection program 5 is selected by the user. It is possible to support the efficient introduction of the defect detection program 5.

In the fourth embodiment, an example of displaying the present introduction plan each time the machine learning and verification process of the learning model 51 is performed has been described, but the present introduction plan is informed each time the learning and verification process are performed a predetermined number of times. It may be configured to transmit to the processing device 2. Further, the server 1 determines whether or not the verification result has converged, and when the verification result converges to a predetermined value, that is, when the possibility that the accuracy of the learning model 51 is further improved is low, the present introduction is made. The plan may be configured to be transmitted to the information processing apparatus 2.

The server 1 generates information about this introduction plan based on the learning state of the learning model 51, that is, the verification result data, the scale of the user, the past program introduction record, the user's AI or the skill level of machine learning, and the like. And may be output.

(Embodiment 5)
Since the program providing method, the server 1 and the computer program 4 according to the fifth embodiment are different from the first embodiment in that they output improvement points in addition to the introduction plan providing method based on the verification result data, they are mainly described below. The above differences will be described. Since other configurations and actions and effects are the same as those in the first embodiment, the corresponding parts are designated by the same reference numerals and detailed description thereof will be omitted.

The server 1 according to the fifth embodiment stores the plan learning model 9 in the storage unit 12.
FIG. 14 is a conceptual diagram showing an example of the plan learning model 9 according to the fifth embodiment. The plan learning model 9 is, for example, a model having a deep neural network. When the verification result data, the type and configuration of the learning model 51, the skill level of the user, and the AI introduction record by the user are input to the plan learning model 9, the plan suitable for the user and the information indicating the improvement points in machine learning are output. do.

The plan learning model 9 has an input layer 9a, an output layer 9c, and an intermediate layer 9b. The input layer 9a has a plurality of nodes in which the verification result data, the type of the learning model 51, the configuration of the learning model 51, the skill level of the user, and the AI introduction record by the user are input. In the intermediate layer 9b, the feature amount of the input information is extracted. The output layer 9c has the accuracy that the first plan is appropriate for the user, the accuracy that the second plan is appropriate, the accuracy that the third plan is appropriate, the first improvement point, the second improvement point, ..., the Nth improvement point. Has multiple nodes that output information indicating the accuracy with which to improve. N is an integer, indicating that there are multiple improvements. The first to third plans are an in-house production plan, a subscription plan, and a consulting plan. The number of plans is not particularly limited to three. The first to Nth improvement points include improvement points regarding preprocessing and improvement points regarding the configuration of the learning model. Improvements regarding preprocessing are used, for example, parameters for preprocessing the captured image, a threshold value, a filter size, a method for setting an area for extracting image data to be input to the learning model 51 from the captured image, and preprocessing. Items such as libraries that should be considered for improvement.

After machine learning of the learning model 51, the information processing apparatus 2 has verification result data and user information, a type of the learning model 51, a configuration of the learning model 51, a user's skill level, an AI introduction record by the user, and the like. Information is transmitted to the server 1. The processing unit 11 receives the verification result data, user information, and the like transmitted from the information processing device 2. Then, the processing unit 11 inputs information such as received verification result data, the type and configuration of the learning model 51, the user's skill level, and the AI introduction record by the user into the plan learning model 9. The processing unit 11 determines an appropriate plan for the user based on the accuracy information output from the node of the output layer 9c of the plan learning model 9, and generates the present introduction plan.

Further, the processing unit 11 identifies improvement points based on the accuracy information output from the node of the output layer 9c of the plan learning model 9, and provides verification result data, user information, information such as the type and configuration of the learning model. At the same time, the information indicating the improvement point is transmitted to the terminal device 3 of the person in charge of the user. The information processing apparatus 2 stores the user ID and the contact information of the person in charge of the user in association with each other. The processing unit 11 may be configured to transmit information or the like indicating the improvement point to the person in charge when the consulting plan is selected by the user. Further, the processing unit 11 may be configured to notify the information processing apparatus 2 or the user of information indicating the improvement point or the like when the in-house production plan or the subscription plan is selected by the user.
The method of generating the plan learning model 9 is as follows. Here, a method of generating the plan learning model 9 by using the server 1 will be described, but the computer for machine learning the plan learning model 9 is not particularly limited.

First, the verification result data obtained by learning and verification of an arbitrary learning model 51, the type of the learning model 51, the configuration of the learning model 51, the skill level of the user, the AI introduction record by the user, and whether or not a plurality of plans are appropriate. Prepare training data in which the evaluation value shown and the evaluation value indicating which of the multiple improvement points should be focused on are associated with each other.

Then, the processing unit 11 inputs the verification result data included in the training data, the type of the learning model 51, the configuration of the learning model 51, the skill level of the user, the AI introduction record by the user, and the like into the unlearned plan learning model 9. The weight coefficient between the neurons constituting the neural network of the plan learning model 9 so that the difference between the probability output from the plan learning model 9 based on the input information and the evaluation value included in the training data becomes small ( The plan learning model 9 is machine-learned by optimizing various parameters such as the coupling coefficient) using the error back propagation method, the steepest descent method, and the like.

When verification result data or the like is input to the machine-learned plan learning model 9 in this way, accuracy information indicating an appropriate plan and improvement points is obtained from the plan learning model 9 according to the input verification result data or the like. It will be output.

According to the program providing method, the program providing device, and the computer program 4 according to the fifth embodiment configured in this way, the present introduction plan is generated in consideration of the type and configuration of the learning model 51, the skill level of the user, and the like. be able to.

Further, it is possible to notify the user or the person in charge of the improvement points in introducing the learning model 51 by the user, and the user or the person in charge can efficiently support the main introduction of the learning model 51.

An example of specifying the main introduction plan and improvement points using the plan learning model 9 which is a machine learning model has been described. However, the main introduction plan is generated based on a rule using a table or the like, and the improvement points are specified. It may be configured as follows.

(Modification example)
The installed program including the unlearned learning model 51 may be made available to other than the installed information processing apparatus 2, or may be made available only to the installed information processing apparatus 2. In the former case, the convenience of the user is improved, and in the latter case, there is an advantage that the program provider can be easily managed.
The modified example and the information processing device 2 described below mean one specific computer, but the information processing device 2 described in the above embodiments 1 to 5 and the "information" according to the present invention. The "processing device" does not necessarily mean that it is limited to one composed of one computer. The information processing device 2 may be a multi-computer composed of a plurality of computers or a device on the cloud as long as the program providing method has the same effect.

Further, the data related to the accuracy of the output data of the learning model 51 machine-learned by the user may be transmitted only from the information processing device 2 in which the program is installed, and the data other than the information processing device 2 in which the program is installed may be transmitted. It may be possible to transmit from the information processing device 2 to the server 1. In the former case, it is easy to manage the program provider, and in the latter case, for example, when the installed program can be used by other than the installed information processing apparatus 2, the convenience of the user can be improved.

The program may be used only for a predetermined period from the time it is installed, or the program may be disabled after a predetermined number of learnings and verifications.

The verification result data may be finally delivered to the server 1, and may be automatically transmitted from the user's information processing device 2 to the server 1, and may be sent to the program provider by e-mail, recording medium, or the like. It may be submitted and provided to the server 1.

The information for canceling the function of deleting the parameter of the learning model 51 may be provided to the user, may be transmitted to the information processing apparatus 2 in which the program is installed, or the information in which the program is installed. It may be transmitted to other than the processing device 2. Further, it may be sent by e-mail or stored in a recording medium and delivered.

Further, instead of distributing the release key, it may be possible to reinstall the program in which the parameters of the learning model 51 are not deleted by the application from the user. For example, information on the link destination of the program in which the parameters of the learning model 51 are not deleted. May be provided to the user. In this case, when the application information is received (step S24), instead of transmitting the function restriction release key, information regarding the link destination of the program in which the parameter of the learning model 51 is not deleted is transmitted to the user (for example, the information processing device 2). The information regarding the link destination is an example of information for enabling the defect detection program 5 whose function of deleting the parameter is canceled, in other words, information for enabling the program whose function is canceled.

The server 1 may include a plurality of programs that can be provided to the user, a program for executing a process related to the program, a user DB 6, and the like, based on information registered in advance by the user and application contents of the user. It may be possible to provide an appropriate program among a plurality of programs that can be provided.

The user may be notified of the remaining period of the temporary introduction at an appropriate timing during the temporary introduction or during the temporary introduction period.

During the temporary introduction period, the temporary introduction period may be extended at the request of the user. The extension period and the number of extensions can be set as appropriate. If there is an extension application, the verification result data of the detection accuracy of the learning model 51 up to that point is acquired, and based on the verification result data, a different program, simple advice, or a paid plan different from this introduction plan, For example, a plan for giving advice on verification may be provided.

Regarding the deletion of the parameter of the learning model 51, the user may be notified by displaying the time until the deletion on the information processing apparatus 2 of the user. Further, the time until the parameter is deleted may be changed according to the user's request or the like.
The parameter may be deleted when the launched application is closed instead of the time. Further, the parameter may be deleted after a predetermined period (for example, one year) without terminating the application.

Further, the information processing apparatus 2 may be configured so that the question information regarding the machine learning of the learning model 51 can be transmitted to the server 1 by the operation of the user. The information processing device 2 receives the question content from the user, and together with the information on the received question content, the ID of the defect detection program 5 or the learning model 51 used by the user, the type and configuration of the learning model 51, and information on preprocessing. (Information such as parameters, thresholds, area settings, etc.), machine learning process logs, and verification result data are transmitted to the server 1. The server 1 receives the question information transmitted from the information processing device 2 and information related to the learning model 51, and transmits the received question information and the like to the terminal device 3 of the person in charge associated with the user ID. The information processing device 2 may be configured so that the question information or the like can be directly transmitted to the terminal device 3 of the person in charge.

Although the defect detection program for detecting the defect of the inspection object has been described in the first to fifth embodiments, the present invention can be provided to any program that operates by learning an unlearned learning model. For example, the present invention may be applied to a program including AI used in speech recognition processing, medical image recognition processing, object classification processing, and language translation processing. In addition, a program for counting steel materials, a program for inspecting or manufacturing vehicle parts such as fog lamp covers, a program for inspecting or manufacturing electronic parts such as printed boards, a program for inspecting the condition of food or packaging, photography. The present invention can be applied to a program for classifying photographs and any other program.

An example of an object detection program will be described. For example, the defect detection program 5 in FIG. 2 is an object detection program including a learning model for detecting an unlearned object. The learning model related to object detection is, for example, a convolutional neural network (CNN). The learning model has an input layer into which an image to be detected as an object is input, an intermediate layer for extracting features of the image, and an output layer to output information indicating the types of objects included in the image to be detected as an object. .. The intermediate layer has a configuration in which a convolution layer that folds the pixel value of each pixel of the image input to the input layer and a pooling layer that maps the pixel value convoluted by the convolution layer are alternately connected, and object detection is performed. The feature amount of the image is extracted while compressing the pixel information of the target image.
When the training model is a model that recognizes a plurality of types of objects, the information output from the output layer includes information indicating the probability that the objects included in the image are each of the plurality of types of objects. For example, when the learning model is configured to recognize the first object and the second object, the output layer outputs the probability that the object included in the image is the first object and the probability that the object is the second object. The same applies when recognizing three or more objects.

When the learning model is an object detection model configured as described above, the verification result data is the correct answer rate, precision rate, recall rate, and F value of the detection result obtained by using the learning model machine-learned by the user. be.
The verification result data may include the probability of being output from the output layer and the data indicating the correct answer (data indicating the object included in the image). The server 1 can calculate the correct answer rate and the like based on the data included in the verification result data.

Based on the verification result data, the server 1 generates the introduction plan and provides it to the user's information processing apparatus 2 as in the first embodiment.

In this way, even in the learning model for object detection, this introduction plan and verification results are displayed, the recommended introduction plan for the object detection program is presented to the user, and the efficient introduction of the object detection program is supported. Can be done.

Although the learning model has been described as being CNN in the first to fifth embodiments, the configuration of the learning model is not limited to CNN. The learning model may be, for example, a model having a configuration such as a neural network other than CNN, an SVM (Support Vector Machine), a Bayesian network, or a regression tree.

1 Server 2 Information processing device 3 Terminal device 4 Computer program 5 Defect detection program 6 User DB
7 Recording medium 8 Introduction plan display screen 9 Plan learning model 11 Processing unit 12 Storage unit 13 Communication unit 51 Learning model 51a Input layer 51b Intermediate layer 51c Output layer 52 Defect degree calculation unit 53 Good / bad judgment unit 54 Learning processing unit N Communication network

Claims (10)

Deliver a program containing an unlearned learning model to the user's information processing device,
Acquire the data related to the accuracy of the output data of the learning model machine-learned by the user who provisionally introduced the delivered program.
A program providing method for outputting information regarding the main introduction plan of the program based on the acquired data. The program providing method according to claim 1, wherein the data includes information regarding a discrepancy between the data obtained by using the learning model machine-learned by the user and the target data. The program
The learning model is trained to detect defects in the inspection object.
The data includes data relating to a discrepancy between the degree of defect of the inspection object obtained by using the learning model machine-learned by the user and the actual degree of defect of the inspection object.
The program providing method according to claim 1 or 2, wherein the data does not include information about an inspection object. The introduction plan is
Including the first plan and the second plan in which the burden on the user related to the main introduction of the program is different,
The program providing method according to any one of claims 1 to 3, wherein the first plan or the second plan is output based on the accuracy of the output data. Compared to the second plan, the first plan has a higher burden on the user related to the main introduction of the program.
The introduction plan is
If the accuracy of the output data is high, it contains information indicating that the first plan is recommended.
The program providing method according to claim 4, which includes information indicating that the second plan is recommended when the accuracy of the output data is low. The program to be temporarily introduced is
After machine learning, it has a function to delete the parameters of the learning model.
The program providing method according to any one of claims 1 to 5, wherein when the program is actually introduced, information for enabling the program whose function has been canceled is transmitted to the user. While the learning model is being verified or trained, the data is acquired and the data is acquired.
The program providing method according to any one of claims 1 to 6, wherein the information regarding the present introduction plan corresponding to the acquired data is transmitted to the user. The program providing method according to any one of claims 1 to 7, wherein the data is deleted when the program is not actually introduced. A distribution unit that distributes programs containing unlearned learning models to users,
An acquisition unit that acquires data related to the accuracy of the output data of the learning model machine-learned by the user who temporarily introduced the delivered program, and
A program providing device including an output unit that outputs information about the main introduction plan of the program based on the acquired data. Deliver a program containing an unlearned learning model to users
Acquire the data related to the accuracy of the output data of the learning model machine-learned by the user who provisionally introduced the delivered program.
A computer program for causing a computer to execute a process of outputting information regarding the main introduction plan of the program based on the acquired data.

Images