Learning support device, learning support method, and learning support program

Abstract

This invention provides a learning support device, a learning support method, and a learning support program that enable the completion of learning. [Solution] The learning support device 100 includes a control unit 20 that trains a machine learning model M based on tagged image data to which tags have been assigned. The control unit 20 includes an error determination unit 22 that performs an error determination process (S515) to determine whether an answer is correct or incorrect based on a preset correct / incorrect threshold, a count determination unit 23 that counts the number of incorrect answers n determined to be incorrect by the error determination unit 22 and performs a count determination process (S517) by comparing the number of incorrect answers n with a preset loop count threshold Nth, and a termination processing unit 24 that performs a learning termination process (S518) when the number of incorrect answers n reaches the loop count threshold Nth.

Description

Technical Field

[0001] The present invention relates to a learning support device, a learning support method, and a learning support program.

Background Art

[0002] Conventionally, when training a machine learning model used for appearance inspection, by a user assigning good or defective tags to each of the pre-collected image group, there is a method of creating tagged image data serving as training teacher data. Further, there is a machine learning system that trains a machine learning model so that it can determine good or defective products according to the tags assigned by the user for the image data. In the machine learning system, as a condition for completing machine learning, there is a case where the machine learning model correctly answers for all of the image data in the image group according to the tags assigned by the user. (For example, Patent Document 1).

[0003] In the conventional method of creating tagged image data, since the user visually observes the images to determine good or defective products and then assigns tags to the image data, when there are a plurality of exactly the same image data in the image group, there is a possibility that different tags are assigned to those same image data. And in the conventional machine learning system, when trying to train a machine learning model based on teacher data including an abnormal teacher data group in which different tags are assigned to the same image data, it becomes impossible to correctly answer for all of the teacher data, and learning may not be completed.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] The present invention aims to provide a learning support device, a learning support method, and a learning support program that enable the completion of learning. [Means for solving the problem]

[0006] (1) A learning support device according to one aspect of the present invention includes a control unit that performs learning of a machine learning model based on tagged image data to which tags have been assigned, wherein the control unit includes an error determination unit that performs an error determination process to determine whether an answer is correct or incorrect based on a preset correct / incorrect threshold, a count determination unit that counts the number of incorrect answers determined to be incorrect by the error determination unit and performs a count determination process by comparing the number of incorrect answers with a preset loop count threshold, and a termination processing unit that performs a termination process for learning when the number of incorrect answers reaches the loop count threshold. (2) In (1) above, the termination processing unit may calculate the average value for each image data and perform an average value determination process to determine whether there are any groups of images with the same average value. If there are no groups of images with the same average value, the termination processing unit may perform the termination process. (3) In (1) or (2) above, the termination processing unit may perform an average value calculation process to calculate the average value for each image data, perform an average value determination process to determine whether there are groups of images with the same average value, and if, as a result of the average value determination process, there are groups of images with the same average value, perform a tag determination process to determine whether the type of tag included in the group of images is the same, and if, as a result of the tag determination process, there is a single type of tag included in the group of images, perform the termination process. (4) In (3) above, if the type of tag included in the group of images is not uniform, a grayscale histogram calculation process is performed to calculate the grayscale histogram of the image data, a grayscale histogram matching determination process is performed to determine whether or not there is a group of images whose grayscale histograms match, if as a result of the grayscale histogram calculation process there is no group of images whose grayscale histograms match, a termination process is performed, and if there is a group of images whose grayscale histograms match, an output process is performed to output the identifier of the image data and the combination of the tag attached to the image data. (5) A learning support method according to one aspect of the present invention includes: a data acquisition process step of acquiring tagged image data to which tags have been assigned; a model learning process step of learning a machine learning model based on the tagged image data; an error determination process step of determining whether an answer is correct or incorrect based on a preset correct / incorrect threshold; a count determination process step of counting the number of incorrect answers determined to be incorrect by the error determination process step, comparing the number of incorrect answers with a preset loop count threshold to perform a count determination; and a termination process step of performing a termination of the learning process when the number of incorrect answers reaches the loop count threshold. (6) A learning support program according to one aspect of the present invention causes a computer to perform the following: a data acquisition process to acquire tagged image data in which tags have been assigned to the image data; a model learning process to learn a machine learning model based on the tagged image data; an error determination process to determine whether an answer is correct or incorrect based on a preset correct / incorrect threshold; a count determination process to count the number of incorrect answers determined to be incorrect by the error determination process and to perform a count determination by comparing the number of incorrect answers with a preset loop count threshold; and a termination process to perform a learning termination process when the number of incorrect answers reaches the loop count threshold. [Effects of the Invention]

[0007] According to the present invention, a learning support device, a learning support method, and a learning support program that enable the completion of learning can be provided. [Brief explanation of the drawing]

[0008] [Figure 1] This is a block diagram illustrating the functions of a learning support device. [Figure 2] This diagram illustrates the hardware configuration of the learning support device. [Figure 3] This is a diagram illustrating the flow of learning support methods. [Figure 4] This diagram illustrates the termination process flow. [Modes for carrying out the invention]

[0009] (Embodiment) Embodiments of the present invention will be described in detail below with reference to the drawings. Figure 1 is a block diagram illustrating the overview of the learning support device 100. Figure 2 is a diagram illustrating the hardware configuration of the learning support device. Figure 3 is a diagram illustrating the flow of the learning support method. In the following, parts having common functions may be denoted by the same reference numerals or symbols.

[0010] [Learning support device] The learning support device 100 according to this embodiment can be used, for example, to collect training data D used for training the machine learning model M in an inspection that uses a machine learning model M to determine whether an image data of an electronic component such as a capacitor is good or defective.

[0011] As shown in Figure 1, the learning support device 100 includes a control unit 20 having a data acquisition unit 21, an error determination unit 22, a count determination unit 23, and a termination processing unit 24. The control unit 20 may also have a display unit 26, an input unit 27, and a modification unit 28. The learning support device 100 may be linked to an external device equipped with a machine learning model M to be learned, may be equipped with the machine learning model M, or may be equipped with a model that is a replica of the machine learning model M. The learning support device 100 may acquire training data D by referring to the internal storage medium of the learning support device 100 via communication from an external storage medium or a data server, etc. The learning support device 100 may directly acquire training data D to which tags (labels) have been assigned by the user to image data obtained by a camera, etc.

[0012] The data acquisition unit 21 acquires at least one training data D (hereinafter sometimes referred to as tagged image data) to which either a good product tag or a defective product tag is attached. Training data D is image data, such as an image of a workpiece (product), to which tags have been pre-attached by an annotator (user). Training data D is used to train the machine learning model M. Training data D consists of one or more data. In the following, unless otherwise specified, training data D is assumed to include multiple data that contain both data with good product tags and data with defective product tags.

[0013] The data acquisition unit 21 is a functional unit that acquires training data D. The data acquisition unit 21 may acquire training data D from a data server (not shown) or the like via communication, or it may acquire training data D by referring to an external storage medium connectable to the learning support device 100 or a storage medium provided by the learning support device 100. The data acquisition unit 21 may also acquire data to which a user has added tags, obtained from a camera or the like.

[0014] The model learning processing unit 25 learns the teacher data D and adjusts the weight coefficients in the neural network of the machine learning model M. The model learning processing unit 25 is composed of an arithmetic processing unit that performs arithmetic processing (S512), an error arithmetic processing unit that performs error arithmetic processing (S513), and an error propagation processing unit that performs error propagation processing (S514). Each processing will be described later.

[0015] The error determination unit 22 performs an error determination process (S515) for determining a correct answer or an incorrect answer based on a preset correct / incorrect threshold value. Here, a correct answer means that the error calculated by the error arithmetic processing unit of the model learning processing unit 25 in the error arithmetic processing (S513) is less than a predetermined correct / incorrect threshold value. Also, an incorrect answer means that the error calculated in the error arithmetic processing (S513) is not less than (exceeds) a predetermined correct / incorrect threshold value. Specifically, as will be described later, the error determination unit 22 determines whether the error calculated in the error arithmetic processing (S513) is less than a predetermined correct / incorrect threshold value. If it is determined that the error is not less than the predetermined correct / incorrect threshold value (S515: NO), that is, if the error determination unit 22 determines an incorrect answer, the processes from S512 to S515 are repeated again. If it is determined that the error is less than the predetermined threshold value (S515: YES), that is, if the error determination unit 22 determines a correct answer, the process proceeds to the completion determination process (S516).

[0016] The count determination unit 23 counts the number of incorrect answers n determined as incorrect answers by the error determination unit 22, and performs a count determination process by comparing the number of incorrect answers n with a preset loop count threshold value Nth. Thereby, when there is such a case that different tags are assigned to the same image data in the image group, it is possible to prevent the learning from being repeatedly determined as an incorrect answer and never ending.

[0017] When the number of incorrect answers n reaches the loop count threshold value Nth, the end processing unit 24 performs an end process of learning (S518). The end process (S518) will be described later.

[0018] The display unit 26 can notify the user of the determination results by displaying the determination results of each part of the control unit 20 on the screen. When different tags are assigned to the same image data, the display unit 26 can display the combination of the identifier (name) of each image data and the tag on a display or the like. Thereby, it is possible to prompt the user to delete the image data that may have been erroneously tagged or to correct the tag that may have been erroneously assigned.

[0019] The display unit 26 may display the acceptable distance, defective distance, evaluation value, number of teacher data, etc. calculated for each image data. The display unit 26 may display a graph in which the feature amounts are plotted in a space of a predetermined dimension. Thus, by being visualized by the display unit 26, for the user, it is possible to confirm the variation in the quality of the teacher data D, and it becomes easy to confirm the tag, acceptable distance, defective distance, evaluation value, or number of teacher data.

[0020] The input unit 27 receives the input of user operations. A user operation is an operation (instruction) by the user that activates the input unit 27, and as an example, it is a selection operation or an input operation.

[0021] When a user operation for changing the tag assigned to the image data displayed on the display unit 26 is input via the input unit 27, the change unit 28 changes the tag assigned to the image data. The change unit 28 may cause the display unit 26 to display a screen for confirming to the user whether there is an error in the tag previously assigned to the image data. When it is determined that there is an error in the tag of the image data, the user can change the tag of the teacher data D from the acceptable tag to the defective tag or from the defective tag to the acceptable tag via the input unit 27 by the change unit 28.

[0022] [Hardware Configuration of the Learning Support Device] Figure 2 is a diagram illustrating the hardware configuration of the learning support device 100 shown in Figure 1. As shown in Figure 2, the learning support device 100 is configured as a normal computer system, including a CPU (Central Processing Unit) 301, RAM (Random Access Memory) 302, ROM 303 (Read Only Memory), a graphics controller 304, an auxiliary storage device 305, an external connection interface 306 (hereinafter referred to as "I / F"), a network I / F 307, and a bus 308.

[0023] The CPU 301 consists of arithmetic circuits and provides overall control of the learning support device 100. The CPU 301 reads programs stored in the ROM 303 or auxiliary storage device 305 into the RAM 302. The CPU 301 executes various processes in the programs read into the RAM 302. The ROM 303 stores system programs used to control the learning support device 100. The graphics controller 304 generates a screen for display on the display unit 26. The auxiliary storage device 305 functions as a storage device. The auxiliary storage device 305 stores application programs that execute various processes. The auxiliary storage device 305 is composed of, for example, an HDD (Hard Disk Drive), SSD (Solid State Drive), etc. The external connection I / F 306 is an interface for connecting various devices to the learning support device 100. The external connection I / F 306 connects, for example, the learning support device 100, a display, a keyboard, a mouse, etc. The network interface 307 communicates with the learning support device 100 and other devices via the network, based on the control of the CPU 301. Each of the above-mentioned components is connected to the bus 308 so as to be able to communicate.

[0024] The learning support device 100 may have hardware other than those described above. For example, the learning support device 100 may include a GPU (Graphics Processing Unit), FPGA (Field-Programmable Gate Array), DSP (Digital Signal Processor), etc. The learning support device 100 does not need to be housed in a single chassis as hardware, and may be separated into several devices.

[0025] The functions of the learning support device 100 shown in Figure 1 are realized by the hardware shown in Figure 2. The data acquisition unit 21 and the modification unit 28 are realized by the CPU 301 executing a program stored in RAM 302, ROM 303, or auxiliary storage device 305, and processing the data stored in RAM 302, ROM 303, or auxiliary storage device 305, or the data acquired via the external connection I / F 306 or network I / F. The display unit 26 is a display device. The input unit 27 is a mouse, keyboard, touch panel, etc. The functions of the modification unit 28 may be further realized using a graphics controller 304.

[0026] [Learning support methods] Next, a learning support method using the learning support device 100 according to this embodiment will be described. Figure 3 is a diagram illustrating the flow of the learning support method. The learning support method by the learning support device 100 includes data acquisition processing (S511) and model learning processing (S512 to S514). The learning support method may also include display processing (S560), input judgment processing (S570), change processing (S580), and notification processing (S590).

[0027] In the learning support method, first, the data acquisition unit 21 of the control unit 20 in the learning support device 100 acquires, for example, training data D from a data server that includes good product data OK, to which good product tags have been attached, and defective product data NG, to which defective product tags have been attached, as part of the data acquisition process (S511).

[0028] The control unit 20, as part of the model learning process (S512 to S514) in the learning support method, learns the training data D and adjusts the weight coefficients in the neural network of the machine learning model M. The model learning process (S512 to S514) consists of an arithmetic process (S512), an error calculation process (S513), and an error propagation process (S514).

[0029] The control unit 20 performs a calculation process (S512) in which it trains the neural network of the machine learning model M using the training data D. In this calculation process (S512), the neural network outputs good product scores and defective product scores for the training data D.

[0030] The control unit 20 performs an error calculation process (S513) to calculate the error between the tags assigned to the training data D and the score output for the training data D.

[0031] The control unit 20 performs an error propagation process (S514), adjusting the weight coefficients of the hidden layers of the neural network using the error calculated in the error calculation process (S513).

[0032] The control unit 20 performs an error determination process (S515) to determine whether the error calculated in the error calculation process (S513) falls below a predetermined correct / incorrect threshold. If it is determined that the error does not fall below the predetermined correct / incorrect threshold (S515: NO), the process from S512 to S515 is repeated. If it is determined that the error falls below the predetermined correct / incorrect threshold (S515: YES), the process proceeds to the completion determination process (S516).

[0033] As a concrete example from the calculation process (S512) to the error determination process (S515), we will explain the case where good product data "OK" with a good product tag "1" is input.

[0034] When the first computational processing (S512) is performed on the training data D, the neural network of the machine learning model M outputs values ​​such as "0.9" and "0.1" as the good product score and defective product score, respectively.

[0035] Next, in the error calculation process (S513), the difference of "0.1" between the good product tag "1" and the good product score "0.9" is calculated. If the defective product data is NG and a defective product tag is attached, the difference with the defective product score is calculated.

[0036] Next, in the error propagation process (S514), the weight coefficients of the hidden layers of the neural network of the machine learning model M are adjusted so that the error calculated in the error calculation process (S513) becomes smaller.

[0037] In the error determination process (S515), the weight coefficients are repeatedly adjusted until the error calculated in the error calculation process (S513) falls below a predetermined correct / incorrect threshold (correct answer), thereby performing machine learning on the neural network of the machine learning model M. As a result, the machine learning model M acquires the ability to classify the target data into either good product tags or defective product tags.

[0038] Next, the control unit 20 performs a completion determination process (S516) to determine whether processing has been completed for all training data D. If it is determined that processing has not been completed for all training data D (S516: NO), the process proceeds to the count determination process (S517). If it is determined that processing has been completed for all training data D (S516: YES), the learning support method is terminated as shown in the flowchart of Figure 3 (END).

[0039] In the count determination process (S517), the control unit 20 counts the number of loops (number of incorrect answers) n in which the steps from the calculation process (S512) to the completion determination process (S516) have been executed for the same training data D. If the number of incorrect answers n (number of loops) is less than or equal to a preset threshold Nth (for example, 1500 times), the system returns to the calculation process (S512) and repeats the steps from the calculation process (S512) to the completion determination process (S516). If the number of incorrect answers n (number of loops) reaches the threshold Nth (for example, 1500 times), the system proceeds to the termination process (S518).

[0040] Next, the termination process (S518) will be explained. Figure 4 is a diagram illustrating the flow of the termination process. As shown in Figure 4, the termination processing unit 24 of the control unit 20 performs an average value calculation process (S610) in the termination process (S518) to calculate the average value μ for each image data. The average value μ calculated for each image data becomes the representative value for each image data.

[0041] The termination processing unit 24 performs an average value determination process (S620) to determine whether there are any groups of images (multiple image data) with the same average value μ, following the average value calculation process (S610). If the result of the average value determination process (S620) is that there are no groups of images (multiple image data) with the same average value μ (S620: NO), the unit displays a message to that effect as appropriate and performs the termination process (END).

[0042] Specifically, for example, the average value μ can be calculated for an image f(i,j) of size (M×N) using the following formula.

number

[0043] If the termination processing unit 24 determines, based on the result of the average value determination process (S620), that there are groups of images with the same average value μ (S620: YES), it executes a tag determination process (S630) to determine whether or not the type of tag included in the group of images is the same.

[0044] Then, if the tag determination process (S630) shows that the image group contains only one type of tag (S630:YES), the system terminates (END). This is because even if there are image groups with the same average value μ, if the image group contains only one type of tag, there is no problem with tag assignment, and therefore the image group is not extracted as a problematic image group. If the learning support device 100 finds that the image group contains only one type of tag (S630:YES), it may, as appropriate, display a message on the display or the like indicating that there is no problem with tag assignment.

[0045] If the result of the tag determination process (S630) indicates that the type of tag included in the image group is not uniform (S630: NO), the termination processing unit 24 executes a grayscale histogram calculation process (S640) to calculate a grayscale histogram of the image data.

[0046] In the grayscale histogram calculation process (S640), the grayscale histogram is a graph that shows the brightness (luminance) distribution of each pixel in the image data. For example, the grayscale histogram may have grayscale values ​​(from 0 (black) to 255 (white)) on the horizontal axis and the number of pixels for each grayscale value on the vertical axis.

[0047] Following the grayscale histogram calculation process (S640), the termination processing unit 24 executes a grayscale histogram matching determination process (S650) to determine whether there are any groups of images (multiple image data) with matching grayscale histograms. In this way, since the grayscale histogram matching determination process (S650) is executed after the average value determination process (S620), it is possible to extract identical image data with different tags with higher accuracy compared to extracting identical image data with different tags based on the result of the average value determination process (S620).

[0048] If, as a result of the grayscale histogram matching determination process (S650), no image group with matching grayscale histograms is found, the process terminates (END).

[0049] If the grayscale histogram matching determination process (S650) finds a group of images with matching grayscale histograms, the output process (S660) is executed to output the identifier of the image data and the combination of the tags assigned to that image data.

[0050] In output processing (S660), the control unit 20 outputs to the display an identifier for the image data and a combination of the tags assigned to that image data, for example, two identical image data images with different tags. This allows the user to see identical image data with different tags, prompting them to delete the incorrectly tagged image data or correct the incorrectly assigned tags.

[0051] As described above, the learning support method according to this embodiment includes a data acquisition process (S511) to acquire tagged image data (training data D) in which tags have been assigned to the image data; a model learning process (S512 to S514) to train a machine learning model M based on the tagged image data; an error determination process (S515) to determine whether an answer is correct or incorrect based on a pre-set correct / incorrect threshold; a count determination process (S517) to count the number of incorrect answers n determined to be incorrect by the error determination process (S515), compare the number of incorrect answers n with a pre-set loop count threshold Nth, and perform a count determination; and a termination process (S518) to perform a learning termination process when the number of incorrect answers n reaches the loop count threshold Nth. This prevents the determination from being repeated indefinitely and learning from not terminating, such as when there is a group of images in the training data D that contain identical image data with different tags. Therefore, the loop from S512 to S517 can be exited and learning can be completed.

[0052] Furthermore, according to the learning support method, the control unit 20 executes the count determination process (S517) and the termination process (S518). This allows for limitations on the number of incorrect answers n determined to be incorrect by the error determination process (S515) by the error determination unit 22, and the number of loops in the model learning process (S512 to S514) executed as a result. This prevents the determination from being repeated indefinitely and prevents the learning from being completed. Therefore, the termination process in S518 can be executed to complete the learning.

[0053] Furthermore, in the process in which a user tags image data to create training data D and inputs it into a learning support device 100 that executes a learning support method for training a machine learning model M, errors may occur, such as identical image data being assigned different tags. In such cases, the learning process will not be completed. Therefore, with the learning support device 100 that executes the learning support method according to this embodiment, the user can confirm whether identical image data was assigned different tags, and if so, which set of image data it was, and can prompt the user to correct the tags assigned to the image data.

[0054] [Learning support methods] Each process in the learning support method described above can be implemented by the learning support program according to this embodiment on the computer provided in the learning support device 100. The learning support program causes the computer to execute a data acquisition process (S511) to acquire tagged image data (training data D) in which tags have been assigned to the image data; a model learning process (S512 to S514) to train a machine learning model M based on the tagged image data; an error determination process (S515) to determine whether an answer is correct or incorrect based on a preset correct / incorrect threshold; a count determination process (S517) to count the number of incorrect answers n determined to be incorrect by the error determination process (S515), compare the number of incorrect answers n with a preset loop count threshold Nth, and perform a count determination; and a termination process (S518) to perform a learning termination process when the number of incorrect answers n reaches the loop count threshold Nth. This prevents the determination from being repeated indefinitely and prevents the learning from terminating. Therefore, learning can be completed.

[0055] It should be noted that the technical scope of the present invention is not limited to the embodiments described above, and various modifications can be made without departing from the spirit of the invention. Furthermore, it is possible to replace the components in the embodiments described above with well-known components as appropriate, without departing from the spirit of the invention. In addition, the above-described modifications can be combined as appropriate, without departing from the spirit of the invention.

[0056] As described above, the learning support device 100 according to the embodiment includes a control unit 20 that trains a machine learning model M based on tagged image data to which tags have been assigned. The control unit 20 includes an error determination unit 22 that performs an error determination process (S515) to determine whether an answer is correct or incorrect based on a preset correct / incorrect threshold, a count determination unit 23 that counts the number of incorrect answers n determined to be incorrect by the error determination unit 22 and performs a count determination process (S517) by comparing the number of incorrect answers n with a preset loop count threshold Nth, and a termination processing unit 24 that performs a learning termination process (S518) when the number of incorrect answers n reaches the loop count threshold Nth. This prevents the determination from being repeated indefinitely and prevents the learning from terminating. Thus, learning can be completed.

[0057] The learning support method according to this embodiment includes a data acquisition process (S511) to acquire tagged image data (training data D) in which tags have been assigned to the image data; a model learning process (S512 to S514) to train a machine learning model M based on the tagged image data; an error determination process (S515) to determine whether an answer is correct or incorrect based on a pre-set correct / incorrect threshold; a count determination process (S517) to count the number of incorrect answers n determined to be incorrect by the error determination process (S515), compare the number of incorrect answers n with a pre-set loop count threshold Nth, and perform a count determination; and a termination process (S518) to perform a learning termination process when the number of incorrect answers n reaches the loop count threshold Nth. This prevents the determination from being repeated indefinitely and learning from not terminating, such as when there is a group of images in the training data D that contain identical image data with different tags. Thus, learning can be completed.

[0058] The learning support program according to this embodiment causes the computer to perform the following: a data acquisition process (S511) to acquire tagged image data (training data D) in which image data is tagged; a model learning process (S512 to S514) to train a machine learning model M based on the tagged image data; an error determination process (S515) to determine whether an answer is correct or incorrect based on a pre-set correct / incorrect threshold; a count determination process (S517) to count the number of incorrect answers n determined to be incorrect by the error determination process (S515), compare the number of incorrect answers n with a pre-set loop count threshold Nth, and perform a count determination; and a termination process (S518) to terminate learning if the number of incorrect answers n reaches the loop count threshold Nth. This prevents the determination from being repeated indefinitely and prevents learning from terminating. Therefore, learning can be completed. [Explanation of Symbols]

[0059] 100 Learning support devices 20 Control Unit 21 Data Acquisition Unit 22 Error judgment section 23 Count determination unit 24 Termination Processing Unit 25 Model Learning Process 26 Display section 27 Input section 28 Changes 301 CPU 302 RAM 303 ROM 304 Graphics Controller 305 Auxiliary storage 306 External Connection I / F (External Connection Interface) 307 Network I / F 308 Bus μ average value D Training data M Machine Learning Model n Number of incorrect answers Nth Loop Count Threshold

Claims

1. The system comprises a control unit that trains a machine learning model based on tagged image data to which tags have been assigned, The control unit includes an error determination unit that performs an error determination process to determine whether the answer is correct or incorrect based on a preset correct / incorrect threshold, A count determination unit counts the number of incorrect answers determined by the error determination unit, compares the number of incorrect answers with a preset loop count threshold, and performs a count determination process. The system includes a termination processing unit that performs the learning termination process when the number of incorrect answers reaches the loop count threshold, The termination processing unit executes an average value calculation process to calculate the average value for each image data, An average value determination process is performed to determine whether there are groups of images with the same average value. If, as a result of the average value determination process, there are no image groups with the same average value, the termination process is performed. Learning support device.

2. The termination processing unit executes an average value calculation process to calculate the average value for each image data, An average value determination process is performed to determine whether there are groups of images with the same average value. If, as a result of the average value determination process, there are groups of images with the same average value, a tag determination process is performed to determine whether or not there is only one type of tag included in the group of images. If the result of the tag determination process indicates that there is only one type of tag included in the image group, the termination process is performed. The learning support device according to claim 1.

3. If the group of images contains more than one type of tag, a grayscale histogram calculation process is performed to calculate the grayscale histogram of the image data. A grayscale histogram matching determination process is performed to determine whether or not there is a group of images whose grayscale histograms match. If, as a result of the grayscale histogram calculation process, there are no image groups for which the grayscale histogram matches, the process is terminated. If there is a group of images whose grayscale histograms match, an output process is executed to output the combination of the identifier of the image data and the tag assigned to the image data. The learning support device according to claim 2.

4. On the computer, A data acquisition process to obtain tagged image data in which tags have been assigned to the image data, A model training process that trains a machine learning model based on the tagged image data, An error determination process that determines whether an answer is correct or incorrect based on a pre-set correct / incorrect threshold, A count determination process that counts the number of incorrect answers determined by the error determination process, compares the number of incorrect answers with a pre-set loop count threshold, and performs a count determination. A learning support program for causing the program to execute a termination process that terminates the learning process when the number of incorrect answers reaches the loop count threshold, In the termination process described above, an average value calculation process is performed to calculate the average value for each image data. An average value determination process is performed to determine whether there are groups of images with the same average value. If, as a result of the average value determination process, there are no image groups with the same average value, the termination process is performed. Learning support program.

Images