JP2021039641A - Re-learning method, and computer program - Google Patents

Abstract

To provide a technique for generating learning data in a region intended by a user when a learning model is re-learned.SOLUTION: A re-learning method includes displaying a feature quantity space and a plurality of first learning data elements arranged on the feature quantity space according to a feature quantity on a display device, extracting a specific region satisfying a region detection condition, generating a plurality of second learning data elements having the feature quantity in the specific region in association with labels, and re-learning a learning model using the plurality of generated second learning data elements and the labels in association with each of the plurality of second learning data elements.SELECTED DRAWING: Figure 4

Description

The present disclosure relates to a technique for retraining a learning model.

Conventionally, a low-density region with a small number of training data is detected in the feature space, new training data in the low-density region is generated, and the training model is retrained using the generated training data. The technique is known (Patent Document 1).

Japanese Unexamined Patent Publication No. 2015-191426

In the conventional technique, when the learning data in the low density region where the number of learning data is small is uniformly generated and the learning model is retrained, the degree of contribution to the improvement of the accuracy of the region not intended by the user, for example, the learning model, is It may be necessary to generate training data even for low density regions.

According to one form of the present disclosure, there is provided a re-learning method for re-learning a learning model learned using a plurality of first learning data elements and labels associated with the plurality of first learning data elements. Will be done. The learning model includes a trained generation model that represents the correspondence between the plurality of first training data elements and the feature amounts of the plurality of first training data elements, and the retraining method uses the feature amounts. Displaying the feature quantity space for representing and the plurality of first learning data elements arranged on the feature quantity space according to the feature quantity on the display device, and a specific region on the feature quantity space. The specific region satisfying the region detection condition for detecting the above is extracted, and a plurality of second learning data elements having the feature amount in the specific region are generated in association with the label. It comprises retraining the training model using the plurality of second training data elements and the labels associated with each of the plurality of second training data elements.

The figure for demonstrating the machine learning system as 1st Embodiment. The figure which shows the inspection object. The figure for demonstrating the judgment system. A flowchart showing how to relearn a learning model. The first figure for demonstrating the re-learning method. The second figure for demonstrating the re-learning method. The third figure for demonstrating the re-learning method. The first figure for demonstrating the second embodiment. The second figure for demonstrating the second embodiment. The first figure for demonstrating a third embodiment. The second figure for demonstrating the third embodiment. FIG. 3 for explaining a third embodiment.

A. First Embodiment:
FIG. 1 is a diagram for explaining a machine learning system 10 as the first embodiment of the present disclosure. FIG. 2 is a diagram showing an inspection object 39. The machine learning system 10 is used in a system for inspecting an inspection object 39 manufactured in a factory or the like, and is executed by a processor of a computer 100. As shown in FIG. 2, the inspection object 39 can be manufactured into three categories, that is, a first-class object 31, a second-class object 32, and a third-class object 33. The first-class object 31 is a normal part having no chips or stains. The second-class object 32 is a part that is chipped but has no stains. The third-class object 33 is a part that is not chipped but is dirty.

As shown in FIG. 1, the machine learning system 10 is a system using a variational autoencoder (VAE), which is one of the models of deep learning. The machine learning system 10 includes an input layer 12, a feature amount calculation unit 14 which is an encoder, a restoration unit 18 which is a decoder, an output layer 19, and a storage unit 106. In the machine learning system 10, when the first learning image 30 is input to the input layer 12, the first learning image 30 is compressed by being represented by the feature amount Z. Further, in the machine learning system 10, the same learning image 30L as the first learning image 30 input from the compressed data is output from the output layer 19. The first learning image 30 is a first learning data element in which the inspection object 39 is detected by a camera which is an example of a sensor.

The input layer 12 receives the input of the first learning image 30 to be compressed. The first learning image 30 input to the input layer 12 is associated with the type data element of the inspection object 39. The type data element is a data element for identifying whether the inspection object 39 is a type 1 object 31, a type 2 object 32, or a type 3 object 33. The feature amount calculation unit 14 is a neural network that extracts a plurality of features from the first learning image 30 input to the input layer 12 and calculates a feature amount Z which is a latent variable using the extracted plurality of features. The feature amount Z is represented by a dimension smaller than the dimension of the learning image 30. For example, the feature amount Z is represented in 100 dimensions.

The restoration unit 18 is a neural network that restores the learning image 30L as the original first learning image 30 by inputting the feature amount Z. The output layer 19 outputs the learning image 30L restored by the restoration unit 18. The feature amount calculation unit 14 and the restoration unit 18 are trained so as to output the same learning image 30L as the original first learning image 30. The feature amount calculation unit 14 and the restoration unit 18 are not limited to the neural network, and other algorithms such as linear regression and Bayes may be used.

The storage unit 106 is composed of a ROM, a RAM, and the like. The storage unit 106 stores the first learning image 30 input to the input layer 12, the image 30I described later, and the feature amount Z calculated by the feature amount calculation unit 14. Further, the input layer 12, the feature amount calculation unit 14, the restoration unit 18, and the output layer 19 of the machine learning system 10 are stored in the storage unit 106.

FIG. 3 is a diagram for explaining the determination system 20. The determination system 20 is a system that outputs a determination result Ru determined using predetermined determination conditions based on the input image 30I of the inspection object 39, and is executed by the processor of the computer 100. The image 30I is acquired by photographing the inspection object 39 with a camera. Judgment result Ru is classified into two categories, "non-defective product" indicating a state in which it can be shipped as a part, and "defective product" indicating a state in which it cannot be shipped. As a predetermined determination condition, there is a first determination condition. The first determination condition is a condition for determining the type 1 object 31 as a "non-defective product" and determining the type 2 object 32 and the type 3 object 33 as a "defective product".

The determination system 20 includes an input layer 12 of the machine learning system 10 and a feature amount calculation unit 14. Further, the determination system 20 includes a determination unit 22 instead of the restoration unit 18. The determination unit 22 is a neural network that classifies the determination result Ru based on the feature amount Z calculated by the feature amount calculation unit 14 and outputs it to the external display device 50. The display device 50 is a device having a display function such as a liquid crystal monitor.

The learning of the determination system 20 is executed before the determination using the image 30I is performed. The learning of the determination system 20 is executed by using the plurality of first learning images 30 in which the inspection object 39 is captured and the sensory evaluation results associated with the plurality of first learning images 30 as teacher data. .. The sensory evaluation result is a result label indicating the result of the sensory determination performed by the judge using the learning image 30 and the first determination condition. As described above, the result label as a label using the first determination condition includes a "non-defective product label" indicating that it is a "non-defective product" and a "defective product label" indicating that it is a "defective product".

The determination unit 22 learns to output the sensory evaluation result associated with the input first learning image 30 as the determination result Ru based on the input first learning image 30. The determination system 20 learned by using the first determination condition and the plurality of first learning images 30 is also called a learning model. The learning model includes a trained generative model 29 representing the correspondence between the plurality of first learning images 30 and the feature quantities Z of the plurality of first learning images 30. The trained generative model 29 includes an input layer 12 and a feature amount calculation unit 14. The storage unit 106 stores the teacher data used for learning the learning model.

FIG. 4 is a flowchart showing a re-learning method of the learning model. FIG. 5 is a first diagram for explaining a re-learning method. FIG. 6 is a second diagram for explaining the re-learning method. FIG. 7 is a third diagram for explaining the re-learning method. The re-learning method shown in FIG. 4 is a re-learning method for re-learning a learning model learned using a plurality of first learning images 30 and labels associated with the plurality of first learning images 30.

As shown in FIG. 4, in step S10, the computer 100 uses the trained generative model 29 to output the feature amount Z of the training image 30 in association with the first training image 30. Specifically, the plurality of first learning images 30 stored in the storage unit 106 and the feature amounts Z corresponding to the plurality of first learning images 30 are output to the display device 50. Then, in step S20, the computer 100 displays the learning screen 60 on the display device 50. As shown in FIG. 5, the learning screen 60 includes a feature amount space 62 for representing the feature amount Z, and a plurality of first learning images 30 arranged on the feature amount space 62 according to the feature amount Z. It is provided with an input screen 71 that accepts input of area search conditions. That is, the process of step S20 includes a process in which the computer 100 displays the input screen 71 on the display device 50.

The number of first learning images 30 generally used for machine learning is enormous. Therefore, as shown in FIG. 5, the computer 100 extracts a part of the plurality of first learning images 30 used for learning the learning model, and extracts the extracted plurality of first learning images 30. The feature amount Z of the first learning image 30 may be output to the display device 50. As a result, the user can easily confirm the learning image 30 arranged in the feature amount space 62. The computer 100 may output all of the plurality of learning images 30 used for learning the learning model to the display device 50 together with the feature amount Z. Further, the computer 100 compresses the multidimensional feature amount Z into two dimensions in order to visualize the feature amount Z on the feature amount space 62 and display it on the display device 50. This compression is performed, for example, using an algorithm called t-SNE. The computer 100 may compress the multidimensional feature amount Z into one dimension or three or more dimensions.

As shown in FIG. 5, the feature space 62 is a two-dimensional space having dimensions x and y. The feature amount space 62 is divided into a plurality of block regions Rxy according to the feature amount Z. Each of the plurality of block areas Rxy has the same range amount of the dimension x and the dimension y. In the present embodiment, each of the plurality of block regions Rxy has a dimension x range amount of 0.125 and a dimension y range amount of 0.125. The plurality of training images 30 are arranged in any of the plurality of block regions Rxy according to the feature amount Z. For example, in the feature amount Z of the learning image 30a located on the upper left side, when the x-dimension is 0.100 and the y-dimension is 0.200, the learning image 30a is placed in the block region Rx1y1 including the feature amount Z. Be placed. In order to improve the visibility of the user, it is preferable that the computer 100 selects the learning image 30 of the storage unit 106 so that three or less learning images 30 are located in each block area Rxy. In the present embodiment, the computer 100 selects the learning image 30 of the storage unit 106 so that one or less learning images 30 are located in each block area Rxy. Further, the computer 100 may display the determination result Ru under the first determination condition in association with each learning image 30. By doing so, the user can easily determine the determination result Ru of each learning image 30 located in the feature amount space 62.

The input screen 71 has a first condition screen 70 for accepting the input of the area detection condition shown in FIG. 5, a second condition screen 72 for accepting the detailed condition of the relearning area shown in FIG. 6, and a first condition. A decision screen 75 for receiving the input contents of the screen 70 and the second condition screen 72 and causing the computer 100 to execute the input contents is provided. In the process of step S20, the first condition screen 70 and the determination screen 75 shown in FIG. 5 are displayed on the display device 50. Details of the first condition screen 70 and the second condition screen 72 will be described later.

As shown in FIG. 4, the computer 100 receives the input of the area search condition via the input screen 71 in step S30. Specifically, as shown in FIG. 5, the computer 100 accepts the input of the area search condition on the first condition screen 70. The first condition screen 70 includes a first detail screen 70a, a second detail screen 70b, and a third detail screen 70c for inputting area search conditions. The first detail screen 70a, the second detail screen 70b, and the third detail screen 70c form a condition screen for inputting area search conditions.

The first detailed screen 70a is a screen that accepts the input of the range of the specific area SR in the feature amount space 62. On the first detail screen 70a, the x-dimensional range, the y-dimensional range, and one block area Rxy of the specific area SR are input as units. For example, as shown in FIG. 5, when "3" is entered in the first blank of the first detail screen 70a and "3" is entered in the second blank, the three block areas Rxy arranged in the x dimension. A total of nine block regions Rxy forming a grid-like region formed by the three block regions Rxy arranged in the y dimension are the range of the specific region SR.

The second detail screen 70b is a screen that accepts the input of the number of classifications of the label associated with the first learning image 30 within the range of the specific area SR received by the first detail screen 70a. That is, the second detail screen 70b accepts the condition of the number of classifications of the label of the first learning image 30 having the feature amount within the range of the specific area SR. The first learning image 30 that is the target of whether or not the condition of the number of classifications is satisfied is the learning image 30 stored in the storage unit 106 used for learning the first learning model. That is, not only a part of the learning images 30 arranged on the feature amount space 62 displayed on the display device 50 is the target. In the present embodiment, the label classification is two classifications, "non-defective product label" and "defective product label". In the present embodiment, the user inputs the number of classifications in the blank of the second detail screen 70b as one of the conditions of the specific area SR. In FIG. 5, since the number of classifications is input as “2”, the learning image 30 to which the “good product label” is associated is associated with the “defective product label” within the range input to the first detail screen 70a. The existence of the learned image 30 is one of the conditions for the area search condition.

The third detail screen 70c is a screen that accepts the input of the density information of the first learning image 30 as the first learning data element within the range of the specific area SR received by the first detail screen 70a. In the present embodiment, the third detailed screen 70c is a screen that accepts input of the number of data elements in the block area Rxy, which is an example of density information, that is, the number of the first learning images 30 in the block area Rxy. The first learning image 30 that is the target of whether or not the condition of the density information is satisfied is the learning image 30 stored in the storage unit 106 used for learning the first learning model. In the example shown in FIG. 5, one condition of the area search condition is that the number of the first learning images 30 having the feature amount within the range of the specific area SR is four or less. The density information is not limited to the number of data elements. For example, the density information is an element indicating a plurality of stages for expressing the density of the first training data element within the range of the specific region SR such as "dense", "dense", "sparse", and "extremely sparse". Alternatively, it may be composed of the ratio of the number of data elements of the first learning image 30 within the range of the specific area SR to the total number of data elements of the first learning image 30. When input is accepted on the third detail screen 70c with an element indicating a plurality of stages, the third detail screen 70c contains a plurality of density information such as "dense", "dense", "sparse", and "extremely sparse". Character images divided into paragraphs are displayed, and the user selects a desired one from the displayed character images. The density of "dense", "dense", "sparse", and "extremely sparse" is, for example, the number of data elements of the first learning image 30 within the range of the specific area SR with respect to the total number of data elements of the first learning image 30. The ratio may be predetermined, or the number of the first learning images 30 within the range of the specific area SR may be predetermined.

After the user inputs the input on the first condition screen 70, the determination screen 75 is pressed by using the graphical user interface 74 such as a cursor, so that the area search condition input on the first condition screen is the computer 100. Will be accepted. As a result, in step S30 of FIG. 4, the computer 100 extracts the specific area SR satisfying the received area search condition and displays it in the feature amount space 62 displayed on the display device 50. For example, as shown in FIG. 6, when the computer 100 extracts the first specific area SR1, the second specific area SR2, and the third specific area SR3 as the specific area SR, the extracted first specific area SR1 to third Each of the specific areas SR3 is displayed by surrounding them with a bold image.

When the area search condition input on the first condition screen 70 is accepted, the computer 100 displays the second condition screen 72 on the display device 50 instead of the first condition screen 70 as shown in FIG. .. As shown in FIG. 4, the computer 100 receives the input of the specific area SR used for re-learning the learning model, that is, the learning specific area LSR in step S40 after step S30. The input of the learning specific area LSR is performed on the second condition screen 72 shown in FIG. The second condition screen 72 includes an area input screen 72a for inputting a specific area LSR for learning, and a label setting screen 72b for inputting labels constituting teacher data. The area input screen 72a accepts the input of the area for generating the second learning data used for re-learning the learning model among the first specific area SR1 to the third specific area SR3 extracted in step S40. It is a screen of. The user inputs the area used for re-learning the learning model from the first specific area SR1 to the third specific area SR3 displayed on the feature amount space 62 on the area input screen 72a. In FIG. 6, the first specific area SR1 is input to the area input screen 72a. The area input screen 72a can input a plurality of specific area SRs.

Further, in step S40, the computer 100 receives the input of the label setting within the range of the specific area SR input on the area input screen 72a on the label setting screen 72b. Specifically, the user sets the label set on the label setting screen 72b to "defective product" and presses the block area Rxy in the specific area SR with the graphical user interface 74 such as a cursor. A defective label is associated with the feature amount of the block area Rxy. On the other hand, when the user sets the label to "good product" on the label setting screen 72b and presses the block area Rxy in the specific area SR on the graphical user interface 74, the feature amount of the pressed block area Rxy is A good product label is associated. In the example shown in FIG. 7, the computer 100 attaches a single hatched image to the block area Rxy associated with the “non-defective product label” and attaches a cross-hatched image to the block area Rxy associated with the “defective product label”. doing. When the determination screen 75 is pressed in the state shown in FIG. 7, the computer 100 generates a plurality of second learning images as a plurality of second learning data elements in association with the label in step S60 as shown in FIG. .. Specifically, the restoration unit 18 is used to generate a second learning image having a feature amount in the designated first specific region SR1. For example, the computer 100 generates a plurality of second learning images having different feature quantities for each block region Rxy of the designated first specific region SR1. A label as teacher data is associated with the plurality of generated second learning images. In step S60, the computer 100 may generate a plurality of second learning images for the block area Rxy in which the number of the first learning images 30 is equal to or less than a predetermined threshold value in the designated specific area SR. Good. Further, the user may specify the number of second learning images to be generated for each block area Rxy of the specific area SR.

Next, in step S70, the computer 100 relearns the learning model using the generated second learning image and the label associated with each of the plurality of second learning images. Specifically, the generated plurality of second learning images and the labels associated with each of the plurality of second learning images are input to the input layer 12 of the determination system 20 as teacher data, and the input second learning is input. Learn to output the label associated with the image as the determination result Ru.

According to the first embodiment, by accepting the input of the region detection condition and generating the second learning image for performing re-learning using the specific region SR satisfying the region detection condition, the feature quantity space 62 It is possible to relearn the learning model by generating a second learning image of the area intended by the user. Further, according to the first embodiment, the user can more accurately specify the intended re-learning area by inputting the area detection condition using the input screen 71. In particular, according to the first embodiment, the user refers to the feature amount space 62 in which the first learning image 30 displayed on the display device 50 is arranged, and refers to the first detailed screen 70a to the first condition screen 70. By using the third detail screen 70c, the area for the intended re-learning can be specified more accurately.

B. Second embodiment:
FIG. 8 is a first diagram for explaining a second embodiment. FIG. 9 is a second diagram for explaining the second embodiment. The difference from the first embodiment is that the type information of the first learning image 30 is added as one of the region detection conditions. Since the other configurations are the same as those in the first embodiment, the same reference numerals are given and the description thereof will be omitted.

In the second embodiment, as shown in FIG. 8, in step S30, the input of the area search condition is accepted via the input screen 71A. The input screen 71A has a fourth detail screen 70d in addition to the first detail screen 70a, the second detail screen 70b, and the third detail screen 70c. The fourth detail screen 70d accepts input of information regarding the type of the first learning image 30, which is a factor for associating the label. The information regarding the type of the first learning image 30 is information indicating which of the first-class object 31 to the third-class object 33 is an image. In the present embodiment, the information regarding the type of the first learning image 30 is "normal" indicating the first learning image 30 of the first type object 31 and "normal" indicating the first learning image 30 of the second type object 32. It includes three types of information: "chip" and "dirt" shown in the first learning image 30 of the third type object 33. The user inputs the type of the first learning image 30 included in the specific area SR on the fourth detail screen 70d as one of the area detection conditions. For example, when "normal" and "missing" are input to the fourth detail screen 70d, the computer 100 has a "normal" first learning image 30 and a "missing" first learning image 30. Extract a specific region SR in which both and are present. In the present embodiment, under the first determination condition, the first learning image 30 of "normal" or "dirt" is associated with a "good product label", and the first learning image 30 of "missing" is associated with a "defective product label". Is associated. Therefore, in the feature quantity space 62, the determination accuracy of the determination system 20 can be improved by increasing the number of data elements used for learning of the determination system 20 as the area where different labels are associated. Therefore, the user inputs an area to which different labels are associated as one of the area detection conditions, and in step S40, the area contributing to the determination accuracy of the determination system 20 is extracted as the specific area SR. When the area detection condition input to the first condition screen 70A shown in FIG. 8 is executed, the computer 100 extracts only the first specific area SR1 and displays it on the display device 50, for example, as shown in FIG.

According to the second embodiment and according to the second embodiment, the user can specify the intended re-learning area with higher accuracy by using the fourth detail screen 70d. For example, the user sets as one of the area detection conditions that the type data elements of the first learning image 30 that affect the determination result Ru of the determination system 20, for example, two types of data elements whose determination results are divided are mixed. By doing so, the determination accuracy of the retrained learning model can be improved.

C. Third Embodiment:
FIG. 10 is a first diagram for explaining a third embodiment. FIG. 11 is a second diagram for explaining the third embodiment. FIG. 12 is a third diagram for explaining the third embodiment. The difference from the first embodiment is that a threshold value of the match rate [%] is added as one of the region detection conditions. Since the other configurations are the same as those in the first embodiment, the same reference numerals are given and the description thereof will be omitted.

In the third embodiment, as shown in FIG. 10, in step S30, the input of the area search condition is accepted via the input screen 71B. The first condition screen 70B has a fifth detail screen 70e in addition to the first detail screen 70a, the second detail screen 70b, and the third detail screen 70c. The fifth detail screen 70e inputs information on the matching rate between the content of the label used when learning the learning model and the determination result Ru using the learning model within the range input on the first detail screen 70a. Accept. That is, the user inputs a numerical value from 0 to 100 on the fifth detail screen 70e as one of the area detection conditions. In the example shown in FIG. 10, that the matching rate is 90% or less is input as one of the region detection conditions.

In step S40, the computer 100 extracts the specific area SR satisfying the area search condition input by the input screen 71B and displays it on the display device 50. The computer 100 determines whether or not the match rate, which is one of the area detection conditions, is satisfied as follows. First, the computer 100 extracts the first candidate specific area CSR1, the second candidate specific area CSR2, and the third candidate specific area CSR3, which are areas satisfying the conditions input to the first detailed screen 70a to the third detailed screen 70c. To do. Then, the computer 100 determines whether or not the match rate condition satisfied on the fifth detail screen 70e is satisfied for each of the extracted candidate specific areas CSR1 to CSR3, and sets the area satisfying the match rate condition as the specific area SR. Extract as. For example, as shown in FIG. 11, for a plurality of first learning images 30 stored in the storage unit 106 having a feature amount in the first candidate specific area CSR1, the contents of the label and the determination result Ru by the learning model match. Calculate the rate. Since the match rate in the first candidate specific region CSR1 is 95.2%, the condition of the match rate is not satisfied. Therefore, the computer 100 does not extract the first candidate specific area CSR1 as the specific area SR. On the other hand, for example, when the matching rate of each of the second candidate specific area CSR2 and the third candidate specific area CSR3 is 90% or less, the computer 100 uses the second candidate specific area CSR2 or The third candidate specific area CSR3 is extracted as the specific areas SR2 and SR3 and displayed on the display device 50.

According to the third embodiment, the same effect is obtained in that it has the same configuration as the first embodiment. For example, according to the third embodiment, it is possible to generate a second learning image of a region intended by the user and relearn the learning model. Further, according to the third embodiment, the user can further accurately specify the intended re-learning area by using the fifth detail screen 70e. For example, the user can suppress the generation of the second learning image in the region having a high matching rate by setting the region having a low matching rate as one of the region detection conditions. That is, it is possible to efficiently execute re-learning that can improve the determination accuracy while reducing the data elements for re-learning the learning model.

D. Other embodiments:
D-1. The first other embodiment:
In each of the above embodiments, the input screen 71 is displayed on the display device 50 to accept the input of the area detection condition from the user, but the present invention is not limited to this. For example, the computer may accept input of the area detection condition from the user by voice recognition. Further, the user may input the specific area SR on the feature amount space 62 by a straight line, a curve, a circle, or the like without going through the input screen 71.

D-2. Second other embodiment:
The second embodiment and the third embodiment may be combined and executed. That is, the input screen 71 may have the first detail screen 70a to the fifth detail screen 70e.

D-3. Third other embodiment:
In each of the above embodiments, the learning data element is the learning image 30 obtained by capturing the inspection object 39 with the camera, but it may be a data element acquired by various sensors. For example, the learning data element may be a sound data element, a vibration data element, or the like. The sound data element is used, for example, for a tapping sound inspection, is acquired by a microphone or the like, and is represented as a sound wave waveform. The vibration data element is used in a vibration method used in, for example, equipment diagnosis technology, is acquired by an acceleration sensor, a gyro sensor, or the like, and is represented as a vibration waveform.

D-4. Fourth Other Embodiment:
In each of the above embodiments, the area detection condition is accepted by the computer 100 by being input by the user via the input screens 71, 71A, 71B, but is not limited thereto. For example, the area detection condition may be set in the computer 100 in advance.

E. Other forms:
The present disclosure is not limited to the above-described embodiment, and can be realized by various configurations within a range not deviating from the gist thereof. For example, the technical features in the embodiments corresponding to the technical features in each form described in the column of the outline of the invention may be used to solve some or all of the above-mentioned problems, or one of the above-mentioned effects. It is possible to replace or combine as appropriate to achieve part or all. Further, if the technical feature is not described as essential in the present specification, it can be deleted as appropriate.

(1) According to one form of the present disclosure, re-learning for re-learning a learning model learned using a plurality of first learning data elements and labels associated with the plurality of first learning data elements. A method is provided. The learning model includes a trained generation model that represents the correspondence between the plurality of first training data elements and the feature amounts of the plurality of first training data elements, and the retraining method uses the feature amounts. Displaying the feature quantity space for representing and the plurality of first learning data elements arranged on the feature quantity space according to the feature quantity on the display device, and a specific region on the feature quantity space. The specific region satisfying the region detection condition for detecting the above is extracted, and a plurality of second learning data elements having the feature amount in the specific region are generated in association with the label. It comprises retraining the training model using the plurality of second training data elements and the labels associated with each of the plurality of second training data elements. According to this form, the user intended in the feature space by accepting the input of the area detection condition and generating the second learning data element for re-learning using the specific area satisfying the area detection condition. The second training data of the region can be generated and the learning model can be retrained.

(2) In the above-mentioned form, it may further have to accept the input of the area detection condition. According to this form, since the user can input the area detection condition, the area for the intended re-learning can be specified more accurately.

(3) In the above-described embodiment, accepting the input may include displaying an input screen for accepting the input of the area detection condition on the display device. According to this form, the user can more accurately specify the intended re-learning area using the input screen.

(4) In the above embodiment, the input screen is a first detail screen that accepts input of a range of the specific area in the feature amount space, and a label associated with the first learning data element within the range. It may have a second detail screen that accepts the input of the number of classifications, and a third detail screen that accepts the input of the density information of the first learning data element within the range. According to this form, the user can more accurately specify the intended re-learning area by using the first detail screen, the second detail screen, and the third detail screen.

(5) In the above embodiment, the input screen may further have a fourth detail screen that accepts input of information regarding the type of the first learning data element.

(6) In the above embodiment, the input screen further includes the contents of the label used at the time of learning the learning model and the learning of the first learning data element having the feature amount within the range. It may have a fifth detail screen that accepts input of information regarding the concordance rate with the determination result using the model. According to this form, the user can further accurately specify the intended re-learning area by using the fifth detail screen.

In addition to the above forms, the present disclosure is in the form of a computer program for re-learning a learning model, a non-transient recording medium on which a computer program is recorded, a re-learning device for re-learning a learning model, and the like. It can be realized.

10 ... Machine learning system, 12 ... Input layer, 14 ... Feature amount calculation unit, 18 ... Restoration unit, 19 ... Output layer, 20 ... Judgment system, 22 ... Judgment unit, 29 ... Generation model, 30 ... First learning image, 30I ... image, 30L ... learning image, 30a ... learning image, 31 ... type 1 object, 32 ... type 2 object, 33 ... type 3 object, 39 ... inspection object, 50 ... display device, 60 ... Learning screen, 62 ... Feature space, 70 ... 1st condition screen, 70a ... 1st detail screen, 70b ... 2nd detail screen, 70c ... 3rd detail screen, 70d ... 4th detail screen, 70e ... 5th Detail screen, 71 ... Input screen, 72 ... Second condition screen, 72a ... Area input screen, 72b ... Label setting screen, 74 ... Graphical user interface, 75 ... Decision screen, 100 ... Computer, 106 ... Storage unit, CSR1 ... No. 1 candidate specific area, CSR2 ... 2nd candidate specific area, CSR3 ... 3rd candidate specific area, LSR ... learning specific area, Ru ... judgment result, Rxy ... block area, SR ... specific area, SR1 ... 1st specific area, SR2 ... 2nd specific area, SR3 ... 3rd specific area,

Claims (7)

A re-learning method for re-learning a learning model learned using a plurality of first training data elements and labels associated with the plurality of first training data elements.
The learning model includes a trained generative model that represents the correspondence between the plurality of first training data elements and the feature amounts of the plurality of first training data elements.
The re-learning method is
Displaying the feature amount space for expressing the feature amount and the plurality of first learning data elements arranged on the feature amount space according to the feature amount on the display device.
Extracting the specific region that satisfies the region detection condition for detecting the specific region on the feature space, and
To generate a plurality of second learning data elements having the feature amount in the specific region in association with the label.
A re-learning method comprising re-learning the learning model using the generated plurality of second learning data elements and the labels associated with each of the plurality of second learning data elements. The re-learning method according to claim 1, further
A re-learning method comprising accepting an input of the area detection condition. The re-learning method according to claim 2.
Receiving the input is a re-learning method including displaying an input screen for receiving the input of the area detection condition on the display device. The re-learning method according to claim 3.
The input screen has a first detail screen that accepts input of a range of the specific area in the feature amount space, and a second that accepts input of the number of classifications of the label associated with the first learning data element within the range. A re-learning method including a detail screen and a third detail screen that accepts input of density information of the first learning data element within the range. The re-learning method according to claim 4.
The re-learning method further comprises a fourth detail screen that accepts input of information regarding the type of the first learning data element. The re-learning method according to any one of claims 3 to 5.
The input screen further matches the content of the label used at the time of learning the learning model with the determination result using the learning model for the first learning data element having the feature amount within the range. A re-learning method having a fifth detail screen that accepts input of information about the rate. A computer program for re-learning a learning model learned using a plurality of first training data elements and labels associated with the plurality of first training data elements.
The learning model includes a trained generative model that represents the correspondence between the plurality of first training data elements and the feature amounts of the plurality of first training data elements.
The computer program
A function of displaying the feature amount space for expressing the feature amount and the plurality of first learning data elements arranged on the feature amount space according to the feature amount on a display device.
A function of extracting the specific area satisfying the area detection condition for detecting the specific area in the feature space, and
A function of generating a plurality of second learning data elements having the feature amount in the specific region in association with the label, and
A computer program that causes a computer to execute a function of retraining the learning model using the generated plurality of second training data elements and the label associated with each of the plurality of second training data elements. ..

Images