Learning device, learning method, and learning program

Abstract

A learning apparatus includes: a data expansion processing section configured to create enhanced learning candidate data based on original image data; a confidence evaluation unit for adding a confidence evaluation value Es to the learning candidate data; the learning data extraction part is used for extracting learning data from the learning candidate data according to the sequence of the confidence evaluation values Es; and a learning processing unit for performing learning based on the learning data. The learning data extraction unit extracts the learning data until a confidence evaluation value Es attached to the learning candidate data reaches a preset confidence evaluation threshold alpha.

Description

TECHNICAL FIELD

[0001] The present application relates to a learning device, a learning method, and a learning program. BACKGROUND

[0002] In the past, there has been a learning method that uses a neural network to perform deep learning, which selects images effective for model learning from prepared images as teacher data for expansion (for example, Patent Literature 1). The conventional learning method is ingenious in data expansion and data selection, and is suitable for learning of large-scale data.

[0003] In the conventional learning method, by adding a condition that the same image as before the data expansion processing is not selected in the data selection processing, the case of repeated learning of similar images due to the increase in the data expansion processing is reduced, and deviation of learning images is avoided. However, in the conventional learning method, since the number of teacher data images used for each learning is small, it is often difficult to sufficiently learn the features of the images, and the number of learning cycles increases, and sometimes it takes a long time to complete the learning. PRIOR ART DOCUMENT PATENT LITERATURE

[0004] Patent Literature 1: Japanese Patent Application Laid-Open No. 2024-002431 SUMMARY PROBLEMS TO BE SOLVED BY THE INVENTION

[0005] An object of the present application is to provide a learning device, a learning method, and a learning program that have a fast learning speed.

[0006] SOLUTION TO PROBLEM (1) A learning device according to an aspect of the present application includes: a data expansion processing section that creates enhanced learning candidate data based on original image data; a confidence evaluation section that attaches a confidence evaluation value to the learning candidate data; a learning data extraction section that extracts learning data from the learning candidate data in order of the confidence evaluation value; and a learning processing section that performs learning based on the learning data. The learning data extraction section extracts the learning data until the confidence evaluation value attached to the learning candidate data reaches a predetermined confidence evaluation threshold value. (2) In the above (1), a representative image selection processing section that selects a representative of an expanded image group common to one original image data from the learning candidate data can be included. (3) In the above (2), the learning data extraction section extracts the learning data until the number of types of the representative images reaches a default number of types of representative images threshold value. (4) In the above (2), the learning data extraction section can include learning candidate data having a smaller confidence evaluation value than the learning candidate data having the largest confidence evaluation value among the representative images in the learning data. (5) A learning method according to an aspect of the present application includes: a data expansion processing step of creating enhanced learning candidate data based on original image data; a confidence evaluation step of attaching a confidence evaluation value to the learning candidate data; a learning data extraction step of extracting learning data from the learning candidate data in order of the confidence evaluation values; and a learning processing step of performing learning based on the learning data. In the learning data extraction step, the learning data is extracted until the confidence evaluation value attached to the learning candidate data reaches a predetermined confidence evaluation threshold. (6) A learning program according to an aspect of the present application causes a computer to execute: a data expansion processing function of creating enhanced learning candidate data based on original image data; a confidence evaluation function of attaching a confidence evaluation value to the learning candidate data; a learning data extraction function of extracting learning data from the learning candidate data in order of the confidence evaluation values; and a learning processing function of performing learning based on the learning data. In the learning data extraction function, the learning data is extracted until the confidence evaluation value attached to the learning candidate data reaches a predetermined confidence evaluation threshold. Effects of the Invention

[0007] According to the present application, it is possible to provide a learning device, a learning method, and a learning program that have a fast learning speed. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a diagram showing an outline of a learning device. Figure 2 is a diagram showing an outline of a flow of a learning method. Figure 3 is a diagram showing a learning method at the time of limiting extraction of learning data by the number of representative images. Figure 4 is a diagram showing a learning method at the time of limiting extraction of learning data by a confidence evaluation threshold. DETAILED DESCRIPTION

[0009] (Embodiment) An embodiment of the present application will be described in detail below with reference to the accompanying drawings. Figure 1 is a diagram showing an outline of a learning device 100. Figure 2 is a diagram showing an outline of a flow of a learning method. Figure 3 is a diagram showing a learning method at the time of limiting extraction of learning data by the number of representative images.Figure 4 is a schematic diagram of a learning method in which the extraction of learning data is limited by a confidence evaluation threshold. Furthermore, hereinafter, portions having common functions can be given the same symbols or reference numerals.

[0010] (Learning device) Referring to Figure 1 The learning device 100 according to the present embodiment will be described. As shown in Figure 1 the learning device 100 of the present embodiment includes a data expansion processing section 10, a confidence evaluation section 20, a representative image selection processing section 30, a learning data extraction section 40, a learning processing section 50, and a control section 60.

[0011] The learning device 100 is a device that causes a machine learning model (hereinafter referred to as "model") to learn so as to be able to make an appropriate evaluation of target data (for example, image data). Here, the model is a mechanism (algorithm) in machine learning that is responsible for receiving input values that can be discriminated by a computer and outputting evaluation values.

[0012] The model identifies and evaluates the content of the target data. For example, when the target data is image data, the quality of the evaluation object (whether it is a good product or a defective product, etc.).

[0013] The model has a structure that includes a neural network and parameters. The neural network has a structure in which a plurality of neurons are coupled. The neural network may, for example, be a multi-layer neural network of a hierarchical structure formed by grouping a plurality of neurons. The neural network is defined by the number of neurons and their coupling relationship. The coupling strength between neurons or between layers is defined by parameters such as weight coefficients. In the neural network, input target data is evaluated based on the operation results of a plurality of neurons and parameters. The model has, for example, a function of classifying target data (learning data) into a good product label or a defective product label, or a function of calculating a good product score.

[0014] The learning device 100 has a computer (not shown) as a main hardware. The computer includes a processor such as a CPU, a GPU, a storage device such as a RAM, a ROM, a hard disk, or a solid state drive (SSD), and a communication section for connecting to a network such as the Internet. In addition, the storage device of the computer also stores a learning program for causing the computer to operate as the learning device 100 of the present embodiment. Note that the computer includes a cloud computing system. Each functional section included in the learning device 100 is realized by the computer as hardware and the learning program as software.

[0015] The data expansion processing section 10 is a functional section that generates enhanced learning candidate data (a learning candidate data set) based on the original image data. Through such enhancement processing, the quality and quantity of the image data can be improved, and the data can be enriched, and a high-quality, efficient, and sufficiently large amount of data for learning of a machine learning model can be generated from a small amount of data, thereby improving the performance of the model. The enhancement of the original image data is achieved through various processing methods, such as: flip processing that flips the original image data left and right, up and down, or 180 degrees to create new data; translation processing that shifts the coordinates of the original image data in the X or Y direction to create data; slight rotation processing that slightly rotates the original image data to the right or left to create data; and filter processing that adds a median filter or a Gaussian filter to the original image data to create data.

[0016] The confidence evaluation section 20 is a functional section that attaches a confidence evaluation value Es calculated by a predetermined formula to the learning candidate data. The confidence evaluation value Es is an index indicating the degree of expectation of how much the learning of the data contributes to the improvement of the model determination performance. In other words, the confidence evaluation value Es is an index indicating the degree of confidence with which the model evaluates (judges) the data when learning the data.

[0017] [Confidence Evaluation Value] The learning device 100 (the confidence evaluation section 20) calculates, for each learning candidate data, the distance in the feature space between the learning candidate data and good product data, i.e., the good product distance, based on the corresponding feature amount when calculating the confidence evaluation value Es, where the good product data is learning data with a good product label (teacher data constituting the basis of machine learning of the model). Therefore, the confidence evaluation value Es can be a parameter calculated based on the distance in the feature space between the learning candidate data and the learning data. Note that the "distance" used below to express the good product distance and the defective product distance, for example, can adopt the Euclidean distance between data projected in the feature space. As long as the distance in the feature space can be calculated, it is not limited to the Euclidean distance, and the Mahalanobis distance or the like can also be used. For example, the distance between learning data that is one data in a learning data group and learning candidate data that is one data in a learning candidate data group can be calculated by the following formula 1. [Formula 1] Where q(k,i) is the coordinate of the learning data k in a certain dimension i of the feature space, and p(s,i) is the coordinate of the learning candidate data in a certain dimension i of the feature space. d(k,s) is the distance between the learning data k and the learning candidate data. Vector qk is the set of coordinate data of the learning data k in the feature space, and vector pk is the set of coordinate data of the learning candidate data in the feature space. It should be noted that k is an integer not exceeding the number of data points in the learning data (m+n: m and n are integers), i is an integer not exceeding the predetermined number of dimensions (j) (j is an integer), and s is an integer not exceeding the number of data points in the learning candidate data (t) (t is an integer).

[0018] If the distance between the learning candidate data and the good data OKg, which is a data point in the good data OK, is denoted as d(OKg,s), then d(OKg,s) is expressed using Formula 1 as Formula 2 below. It should be noted that in OKg, OK is the symbol for good, and g is an integer not exceeding the number (m) of good data OK. [Formula 2] q(OKg,i) is the coordinate of the good data OKg in the training data set on a certain dimension i in the feature space, and the vector qOKg is the set of coordinate data of the good data OKg in the feature space.

[0019] If the distance set between the learning candidate data and each good product data OK is represented as a vector d(OK,s), then the vector d(OK,s) is expressed by Formula 2 as the following Formula 3. [Formula 3]

[0020] The good product distance E(OK,s) in the learning candidate data is, for example, the minimum value among the elements of the vector d(OK,s). That is, the good product distance E(OK,s) is the minimum value among the elements of the vector d(OK,s), which is the set of distances between the learning candidate data and each good product data OK. The good product distance E(OK,s) is expressed by Equation 3 as Equation 4 below. Here, the smaller the good product distance E(OK,s), the closer the learning candidate data is to any good product data OK in the feature space. [Formula 4]

[0021] For example, we can extract *a* elements from the vector *d(OK,s)* in ascending order and take the average of these *a* elements as the good product distance *E(OK,s)* in the learning candidate data. *a* is a natural number, such as 3. The good product distance *E(OK,s)* in this case is expressed as Equation 5 below. The smaller the good product distance *E(OK,s)*, the closer the learning candidate data is to multiple (*a*) good product data sets in the feature space, and the closer the learning candidate data is to the group of good product data sets (good product cluster). [Formula 5]

[0022] Furthermore, the learning device 100 (confidence evaluation unit 20) calculates the distance between the learning candidate data and the defective data in the feature space, i.e., the defective distance, based on the corresponding feature quantity, for each learning candidate data, where the defective data is the learning data with a defective label. If the distance between the learning candidate data and a defective data NGh in the defective data NG is represented as d(NGh,s), then d(NGh,s) is expressed using Formula 1 as Formula 6 below. It should be noted that in NGh, NG is the symbol for defective, and h is an integer not exceeding the number (n) of defective data NG. [Formula 6] It should be noted that q(NGh,i) is the coordinate of the defective data NGh in the learning data on a certain dimension i in the feature space, and the vector qNGh is the set of coordinate data of the defective data NGh in the feature space.

[0023] If the set of distances between the learning candidate data and each defective data NG is represented as a vector d(NG,s), then the vector d(NG,s) is expressed by Formula 6 as the following Formula 7. [Formula 7]

[0024] The defective distance E(NG,s) in the learning candidate data is, for example, the minimum value among the elements of the vector d(NG,s). That is, the defective distance E(NG,s) is the minimum distance between the learning candidate data and each defective data NG. The defective distance E(NG,s) is expressed by Equation 7 as Equation 8 below. In this case, the smaller the defective distance E(NG,s), the closer the learning candidate data is to any defective data NG in the feature space. [Formula 8]

[0025] For example, we can extract *a* elements from the vector d(NG,s) in ascending order and take the average of these *a* elements as the defect distance E(NG,s) in the learning candidate data. In this case, the defect distance E(NG,s) is expressed by Formula 7 as Formula 9 below. The smaller the defect distance E(NG,s), the closer the learning candidate data is to multiple (*a*) defective data sets NG in the feature space, and the closer the learning candidate data is to the group of defective data NG (defect cluster). [Formula 9]

[0026] Furthermore, the learning device 100 (confidence evaluation unit 20) uses the calculated good product distance E(OK,s) and defective product distance E(NG,s) to calculate the confidence evaluation value Es in the learning candidate data. The confidence evaluation value Es is, for example, the value obtained by dividing the good product distance E(OK,s) by the defective product distance E(NG,s), and is expressed as the following formula 10. [Formula 10]

[0027] For example, a confidence score Es less than 1 indicates a smaller distance between good and defective products, E(OK,s), compared to the distance between defective and good products, E(NG,s). This means the candidate data is closer to the good product cluster than the defective cluster. Therefore, if the candidate data is labeled as defective, a confidence score Es less than 1 indicates that, in the current learning data set, the candidate data is more difficult to classify as either good or defective in the model, resulting in a higher learning effect for the model.

[0028] On the other hand, for example, a confidence score Es greater than 1 indicates that the distance between defective products E(NG,s) and good products E(OK,s) is smaller, meaning the learning candidate data is closer to the defective product cluster than the good product cluster. Therefore, if the learning candidate data is labeled as good, a larger confidence score Es indicates that in the current learning data set, it is more difficult for the learning candidate data to be classified as good or defective in the model, resulting in a higher learning effect for the model.

[0029] The representative image selection processing unit 30 is a functional unit that selects a representative image from a common extended image group of the original image data as the representative image from the learning candidate data. For example, such as Figure 3 and Figure 4As shown, in the learning candidate data, three circular markers represent an expanded image group. The circular marker with the smallest confidence evaluation value Es (when sorted in ascending order) among the three circular markers is the representative image of that expanded image group. When the learning candidate data is sorted in ascending order of confidence evaluation value Es, the representative image selection processing unit 30 selects the first circle, the third triangle, the sixth rhombus, and the nth chestnut shape in sequence according to the order of the learning candidate data.

[0030] The learning data extraction unit 40 is a functional unit that treats the learning candidate dataset as a group, calculates the confidence evaluation value Es for the group, and performs extraction processing to extract learning candidate data whose confidence evaluation value Es is within a predetermined range (e.g., not exceeding a preset confidence evaluation threshold α) as learning data. In other words, the learning data extraction unit 40 extracts learning data from the learning candidate data in order of confidence evaluation value Es.

[0031] The learning processing unit 50 is a functional unit that enables the model to undergo learning processing at a predetermined learning rate. Learning processing involves adjusting the model's parameters to bring them closer to their optimal values.

[0032] The control unit 60 is a functional unit that controls the learning data extraction unit 40 and the learning processing unit 50, enabling the model to perform effective learning processing.

[0033] (Learning methods) The following describes the flow of the learning method according to the embodiment. The learning method can be implemented using a learning device 100. Each step (process or procedure) of the learning method can be executed by a corresponding component (assembly) in the learning device 100. Figure 2 It is a diagram showing an outline of the learning method process. Figure 3 This is a schematic diagram of a learning method that limits the extraction of learning data by restricting the number of representative images. Figure 4 This is a schematic diagram of a learning method that limits the extraction of learning data by using a confidence level assessment threshold.

[0034] (1) The learning device 100 pre-acquires a dataset as the original image data.

[0035] (2) Next, as Figure 2 As shown, the learning device 100, through the data expansion processing unit 10, creates enhanced learning candidate data (learning candidate dataset) based on the original image data (data expansion processing step) (process A). This processing improves the quality and quantity of the image data, enriches the data, and generates high-quality, efficient, and sufficiently large amounts of data from a small amount of data for learning the machine learning model, thereby improving the model's performance.

[0036] (3) Next, the learning device 100, through the confidence evaluation unit 20, attaches a confidence evaluation value Es calculated using a predetermined formula to the learning candidate data (confidence evaluation step). Furthermore, the learning device 100, through the representative image selection processing unit 30, uses the confidence evaluation value Es attached to each learning candidate data (image) to select a representative image from the learning candidate dataset that is a common extended image group of the original image data as the representative image (process B). The representative image can be the image with the smallest confidence evaluation value Es among multiple image data (extended image groups) extended from the common original image data (meaning that images with the same extended source image are not selected in process B). In other words, in process B, when images are arranged in ascending order of evaluation value, the image with the smallest evaluation value is selected as the representative image, and subsequent images with the same original image data are not selected as representative images. For example, such as... Figure 3 As shown, the pie images in the first row, “Learning Candidate Data” column, are selected as representative images, while the pie images in the second and fourth rows, which have the same original image data, are not selected as representative images.

[0037] (3-1) The learning data extraction unit 40 can also extract learning data until the confidence evaluation value Es attached to the learning candidate data reaches the preset confidence evaluation threshold α (learning data extraction step).

[0038] (4) Next, the number of representative images selected in process B is detected. If a representative image is detected, process D is initiated. If no representative image is detected, the learning termination process (process C) is executed.

[0039] (5) The selected representative images are sorted according to the confidence assessment value Es (e.g., ascending order) and are divided into learning type and other type, for example, based on a predetermined value, with the smaller confidence assessment value being the learning type and the larger one being the other type (process D).

[0040] (5-1) In the process D, the learning device 100 can select a representative image until the confidence assessment value Es reaches the preset confidence assessment threshold α (condition 1). In this way, it can avoid extracting inefficient learning candidate data (teacher data) as learning data, and can also extract an appropriate amount of learning data, neither too much nor too little.

[0041] (5-2) Alternatively, in process D, the learning device 100 may perform classification processing until a predetermined number of representative images is reached (condition 2). In other words, the learning data extraction unit 40 of the learning device 100 will continue to extract learning data until the number of representative image types reaches a default threshold β. In this way, an appropriate amount of learning data can be extracted, neither too much nor too little.

[0042] (5-3) Alternatively, in process D, the learning device 100 may continue until all representative images have been classified (condition 3). This allows for the extraction of an appropriate amount of learning data, neither too much nor too little. It should be noted that conditions 1-3 above can be applied individually or in any combination.

[0043] (6) Extract the representative images that were classified as learning data by processing D as learning data (processing E).

[0044] (6-1) In this process, images with a confidence evaluation value Es smaller than that of the representative image are extracted as learning data (processing G). In other words, the learning data extraction unit 40 includes learning candidate data with a confidence evaluation value Es smaller than that of the representative image with the largest confidence evaluation value Es in the learning candidate data. In this way, all extended image groups common to the representative image and the original image data can be extracted, thereby suppressing the insufficiency of learning data and extracting sufficient learning data with good learning effect compared to the case where other learning candidate data common to the representative image and the original image data are not extracted.

[0045] (7) After processing E, learning is performed based on the learning data extracted up to processing E (process F). It should be noted that a termination condition is also set, that is, learning stops when the expected appropriate number of learning iterations is exceeded. In processing F, after learning is completed, the process returns to processing B.

[0046] (8) Appropriately, between processing A and processing B, the pre-learned model is set as the initial value (process H).

[0047] Thus, the learning method according to this embodiment uses the learning device 100 to extract learning data based on representative images in the original image data, which serves as the data expansion source, until predetermined conditions (conditions 1-3 above) are met. This reliably extracts high-quality learning data of different types required for learning, while simultaneously extracting a sufficient amount of learning data without causing excessive redundancy. Therefore, a learning device 100 and a learning method with fast learning speed can be provided.

[0048] (Learning program) The learning program has the functionality to enable the computer to perform each step of the learning method described above. Specifically, corresponding to each step in the learning method, the learning program has data expansion processing capabilities, confidence assessment capabilities, learning data extraction capabilities, and learning processing capabilities. In this way, high-quality, efficient, and sufficiently large amounts of data can be generated from a small amount of data for learning the machine learning model, thereby improving the model's performance. Therefore, a learning program with fast learning speed can be provided.

[0049] It should be noted that the technical scope of the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention.

[0050] Furthermore, without departing from the spirit of the present invention, the constituent elements in the above embodiments may be appropriately replaced with well-known constituent elements. Additionally, without departing from the spirit of the present invention, the above variations may be appropriately combined.

[0051] As described above, the learning apparatus 100 according to the embodiment includes: a data expansion processing unit 10 for creating enhanced learning candidate data based on original image data; a confidence evaluation unit 20 for attaching confidence evaluation values ​​Es to the learning candidate data; a learning data extraction unit 40 for extracting learning data from the learning candidate data in order of confidence evaluation values ​​Es; and a learning processing unit 50 for learning based on the learning data. The learning data extraction unit 40 extracts learning data until the confidence evaluation values ​​Es attached to the learning candidate data reach a preset confidence evaluation threshold α. In this way, high-quality, efficient, and sufficiently large amounts of data for learning the machine learning model can be generated from a small amount of data, thereby improving the performance of the model. Therefore, a learning apparatus 100 with fast learning speed can be provided.

[0052] The learning method according to the implementation includes: a data expansion processing step, creating enhanced learning candidate data based on the original image data; a confidence evaluation step, attaching a confidence evaluation value Es to the learning candidate data; a learning data extraction step, extracting learning data from the learning candidate data in order of the confidence evaluation values ​​Es; and a learning processing step, performing learning based on the learning data. The learning method involves extracting learning data in the learning data extraction step until the confidence evaluation value Es attached to the learning candidate data reaches a preset confidence evaluation threshold α. This allows for the generation of high-quality, efficient, and sufficiently large amounts of data from a small amount of data for learning the machine learning model, thereby improving the model's performance. Therefore, a fast image inspection method can be provided.

[0053] The learning procedure according to the implementation method enables a computer to perform: a data expansion processing function to create enhanced learning candidate data based on the original image data; a confidence evaluation function to attach a confidence evaluation value Es to the learning candidate data; a learning data extraction function to extract learning data from the learning candidate data in order of the confidence evaluation values ​​Es; and a learning processing function to perform learning based on the learning data. In the learning data extraction function, the learning procedure extracts learning data until the confidence evaluation value Es attached to the learning candidate data reaches a preset confidence evaluation threshold α. In this way, high-quality, efficient, and sufficiently large amounts of data can be generated from a small amount of data for learning the machine learning model, thereby improving the model's performance. Therefore, a learning procedure with fast learning speed can be provided.

[0054] Explanation of reference numerals in the attached figures 100: Learning device, 10: Data expansion processing unit, 20: Confidence evaluation unit, 30: Representative image selection processing unit, 40: Learning data extraction unit, 50: Learning processing unit, 60: Control unit, α: Confidence evaluation threshold, β: Representative image type and quantity threshold.

Claims

1. A learning device, comprising: The data extension processing unit is used to create enhanced learning candidate data based on the original image data; A confidence assessment unit is used to attach confidence assessment values ​​to the learning candidate data; The learning data extraction unit is used to extract learning data from the learning candidate data in the order of the confidence evaluation values; A learning processing unit is used to perform learning based on the learning data; The representative image selection processing unit is used to select a representative image from the extended image group common to the original image data from the learning candidate data as the representative image, and extract it as the learning data; The learning data extraction unit extracts the learning data until the confidence evaluation value attached to the learning candidate data reaches a preset confidence evaluation threshold.

2. The learning device according to claim 1, The learning data extraction unit extracts the learning data until the number of representative image types reaches the default threshold for the number of representative image types.

3. The learning device according to claim 1, The learning data extraction unit includes learning candidate data with a smaller confidence evaluation value than the learning candidate data with the largest confidence evaluation value in the representative image into the learning data.

4. A learning method, comprising: The data expansion processing step creates enhanced learning candidate data based on the original image data; The confidence assessment step involves attaching a confidence assessment value to the learning candidate data; The learning data extraction step involves extracting learning data from the learning candidate data according to the order of the confidence evaluation values. The learning process involves learning based on the learning data. The representative image selection processing step involves selecting a representative image from the extended image group common to the original image data from the learning candidate data, and extracting it as the learning data. In the learning data extraction step, the learning data is extracted until the confidence evaluation value attached to the learning candidate data reaches the preset confidence evaluation threshold.

5. A learning program for causing a computer to execute: The data extension processing function creates enhanced learning candidate data based on the original image data; The confidence assessment function adds a confidence assessment value to the learning candidate data; The learning data extraction function extracts learning data from the learning candidate data according to the order of the confidence evaluation values; The learning processing function learns based on the learning data; The representative image selection processing function selects a representative image from the extended image group common to the original image data from the learning candidate data as the representative image and extracts it as the learning data; In the learning data extraction function, the learning data is extracted until the confidence evaluation value attached to the learning candidate data reaches the preset confidence evaluation threshold.

Images