JP7755261B2 - Training Data Generation System - Google Patents

Description

The present invention relates to a training data generation system.

A teacher data generation device has been disclosed that generates training data with teacher signals for training a deep neural network (DNN), a deep learning device (see, for example, Patent Document 1).

International Publication No. 2020/217425

Training a deep learning machine requires a huge amount of training data with teacher signals. To generate such a large amount of training data with teacher signals, an expert with extensive knowledge of the subject being classified must use computer technology to generate the training data and then spend a long time adding teacher signals to the training data. This is an extremely burdensome task. For this reason, there is a need for a system that allows the above-mentioned expert users to generate a huge amount of training data with teacher signals in a short amount of time and with little effort.

The present invention was made in light of the above-mentioned circumstances, and aims to provide a training data generation system that enables users to generate a huge amount of training data with teacher signals in a short amount of time and with little burden, without requiring users to have computer skills.

In order to achieve the above object, the training data generation system according to the present invention comprises:
A training data generation system that generates training data with a teacher signal used for learning a deep learning device that classifies objects to be classified based on the presence or absence of features,
a training data generation model that defines an individual as a vector in a parameter space whose coordinate axes are parameters defining the characteristics, generates a population according to a probability distribution defined in the parameter space, and generates training data defined by setting values of parameters corresponding to each individual in the population;
a determination unit that allows a user to determine how appropriately each piece of training data generated by the training data generation model represents a feature;
an optimization unit that performs an interactive optimization process to estimate the probability distribution while repeating the generation of training data by the training data generation model and the determination by the determination unit, so that the number of training data that are determined to appropriately represent the features increases or the degree to which the features in the training data are appropriately represented increases;
Equipped with
The optimization unit
As the optimization process, interactive evolutionary computation using a covariance matrix adaptive evolution strategy is performed .

The training data generation model is
generating characteristic image data indicating characteristics defined by the setting values of the parameters corresponding to each of the individuals;
generating the training data by superimposing the feature image data on reference image data to be classified that does not have the feature;
This may also be the case.

the classification object is the outer surface of the inspection object,
The feature is a scratch or crack that appears on the outer surface,
the reference image data is image data of an outer surface of the inspection object that does not have the feature;
The feature image data is image data of the feature defined by a set value of a parameter corresponding to the individual.
This may also be the case.

The training data generation model is
generating image data to be classified that represent features defined by the parameter setting values corresponding to each of the individuals;
This may also be the case.

The characteristics are characteristics related to the shape, structure, combination of parts, pattern, color, or a combination thereof of the classification object.
This may also be the case.

the training data is sequence data to be classified,
The training data generation model is
generating the sequence data representing features defined by parameter setting values corresponding to each of the individuals;
This may also be the case.

According to the present invention, training data with teacher signals can be generated by repeatedly generating training data and having the user determine whether the training data is likely to represent a feature, and then performing an interactive optimization process to estimate the probability distribution of parameters that define the features so that either the number of training data determined to be likely to represent a feature increases, or the likelihood of the training data representing a feature increases.This allows the user to generate a huge amount of training data with teacher signals in a short amount of time with little effort, without having to have computer skills.

1A shows image data of a painted surface of a vehicle body, and FIG. 1B shows a plurality of image data obtained by capturing images of the same painted surface under different lighting conditions. 1 is a block diagram showing a functional configuration of a training data generation system according to a first embodiment of the present invention; FIG. 10 is a schematic diagram showing the composition of image data. FIG. 10 is a schematic diagram of a quadratic curve approximating a scratch. 10A to 10D are diagrams showing aspects related to the brightness of scratches, and FIG. 10E is a schematic diagram showing the interval between bright and dark lines. 1A is a schematic diagram showing an example of a change in luminance of a flaw due to a Gaussian mixture distribution, and FIG. 1B is a diagram showing an example of a change in luminance in image data of the flaw. FIG. 10 is a schematic diagram showing how image data is linked to a user's judgment result. FIG. 10 is a diagram illustrating how the probability distribution changes for each generation in the variance matrix adaptive evolution strategy. FIG. 3 is a block diagram showing the hardware configuration of the training data generation system of FIG. 2. 3 is a flowchart of a training data generation process of the training data generation system of FIG. 2 .

Embodiments of the present invention will now be described in detail with reference to the drawings. In each drawing, the same or equivalent parts are designated by the same reference numerals.

The training data generation system according to this embodiment generates training data with teacher signals used in training a deep learning device that classifies objects based on the presence or absence of features. In this embodiment, the objects to be classified are the outer surfaces of the objects to be inspected, and the features are scratches that appear on the outer surfaces of the objects to be inspected. The training data is image data of the outer surfaces of the objects to be inspected, and the teacher signals are information indicating the presence or absence of features, i.e., the presence or absence of scratches. In other words, the training data generation system according to this embodiment generates image data with information on the presence or absence of scratches that is used in training a deep learning device that classifies the outer surfaces of the objects to be inspected based on the presence or absence of scratches that appear in the image data of the outer surfaces of the objects to be inspected.

In particular, the training data generation system according to this embodiment uses the paint surface of an automobile body painted at an automobile production plant as the exterior surface to be inspected, and paint scratches (hereinafter simply referred to as "scratches") as the feature used as the basis for classification. As shown in FIG. 1(A), image data of a painted surface, for example, 2 cm square, is captured and input into a deep learning machine, which then classifies the painted surface as scratched or not, thereby inspecting the painted surface. It has been confirmed that when a deep learning machine is trained using image data with teacher signals, in which a teacher signal has been manually added by an expert to thousands of images of scratched painted surfaces, the deep learning machine can detect more than 95% of scratches.

Compared to preparing image data of a scratch-free painted surface, preparing a large amount of image data of a painted surface that contains scratches is extremely difficult. Furthermore, as shown in Figure 1(B), even the same painted surface of a vehicle can appear completely different depending on how the light hits it, and in some cases, actual scratches may not be visible. This makes it difficult to inspect for scratches using a deep learning system. The training data generation system of this embodiment generates, in a short amount of time and with little overhead, the enormous amount of image data with teacher signals required for training a deep learning system that automatically inspects painted surfaces for scratches.

[Functional configuration]
As shown in FIG. 2, the training data generation system 1 includes, as its functional configuration, a training data generation model 10, a determination unit 11, and an optimization unit 12.

[Training data generation model]
The training data generation model 10 generates the image data with the teacher signal described above. The training data generation model 10 generates composite image data by superimposing the image data of the flaw on the reference image data of the painted surface, as shown in Fig. 3. That is, in this embodiment, the reference image data of the painted surface corresponds to the image data of the painted surface without a flaw, and the image data of the flaw corresponds to the feature image data.

The image data of the flaw is an image generated by computer graphics. The training data generation model 10 generates the image data of the flaw based on preset parameter settings. Such parameters include, for example, the following:
(1) Parameters of the approximation curve; (2) Parameters of the line brightness, combination, and positional relationship; (3) Parameters related to the image smoothing filter; (4) Parameters related to image synthesis; (5) Parameters related to the brightness change of the flaw due to the mixed Gaussian distribution.

(1) Parameters of Approximation Curve Most scratches are linear and can be approximated by a quadratic curve. Therefore, image data of the scratches can be generated based on the set values of parameters that define the quadratic curve. In this embodiment, the parameters of this quadratic curve: σ _c and a are used as parameters for generating image data of the scratches.

As shown in Figure 4, let r be the vector that indicates the locus of the quadratic curve. Let _r1 and _r3 be the endpoint vectors of the quadratic curve, and let _r2 be the vector of the point (midpoint) where the tangency is parallel to the line segment connecting the endpoints. In this case, vector r is expressed as a linear combination of vectors _r1 , _r2 , and _r3 using coefficients _k0 , _k1 , _k2 , and _k3 .
r=(1/k ₀ )×(k ₁ r ₁ +k ₂ r ₂ +k ₃ r ₃ )...(1)
The midpoint _r2 of this quadratic curve is expressed as follows:
r ₂ = (r ₁ + r ₃ )/2+σ _c μ | r ₁ − _{r 3} |
where σ _c is the variance of the vector r ₂ of the midpoints of the quadratic curve, and μ is a vector of uniform random numbers in [0,1].

The coefficients k ₀ , k ₁ , k ₂ , and k ₃ are expressed by the following formulas.
k ₁ = (1-t) (1-2t)
k ₂ = (4+a)t(1-t)
k ₃ =t(2t-1)
k ₀ =k ₁ +k ₂ +k ₃ =1+at(1-t)
Here, t = 0 to 1. When t = 0, r = _r1 , when t = 0.5, r = _r2 , and when t = 1, r = _r3 .

By changing the parameter a, the quadratic curve is defined as follows:
a>0: hyperbola a=0: parabola a<0: ellipse In this way, the quadratic curve that approximates the flaw is determined by the variance σ _c of the midpoint _r2 and the parameter a of the curve. In this embodiment, σ _c and a are included as parameters to be optimized.

(2) Line Brightness, Combination, and Positional Relationship Parameters As shown in Figures 5(A) to 5(D), the brightness of a scratch varies depending on how the light hits it. For example, in the image shown in Figure 5(A), the scratch is composed of white bright lines, while in the image shown in Figure 5(B), the scratch is composed of black dark lines. In the image shown in Figure 5(C), the scratch has a bright line at the bottom edge and dark lines elsewhere. In Figure 5(D), the scratch has a bright line at the top edge and dark lines elsewhere.

In this embodiment, the parameter to be optimized includes the line type T _line of the flaw, which can be set to whether the lines constituting the flaw are only light lines, a combination of light and dark lines, or only dark lines. Furthermore, as shown in FIG. 5E, the distances Ox and Oy between the light and dark lines on the image can also be included as parameters to be optimized. The number of light and dark lines in a single flaw can also be included as a parameter.

(3) Parameters Related to Image Smoothing Filters Since the sharpness of a flaw changes depending on the focus adjustment, the lighting conditions, etc., parameters of a filter that smooths the image can be included as parameters to be optimized. For example, the smoothing strength _Tbulr is a parameter. Such filters include, but are not limited to, a Gaussian filter.

(4) Parameters Related to Image Synthesis As shown in Fig. 3, the composite image data is formed by superimposing flawless reference image data and flawed image data, i.e., by embedding the flawed image data in the reference image data. In this case, each pixel of the composite image data is a weighted linear sum of a pixel of the reference image data and a pixel of the flawed image data. Each of these weights can be included as a parameter to be optimized.

(5) Parameters Related to Luminance Changes Due to Gaussian Mixture Distribution As shown in FIG. 5(D), in an image of an actual scratch, the brightness of the scratch changes depending on factors such as the unevenness of the scratch depth and the way the light hits it. The change in brightness of the scratch can be expressed using a Gaussian mixed distribution, as shown in FIG. 6. Therefore, in this embodiment, the brightness value of the scratch in the image data of the scratch shown in FIG. 3 is corrected using a Gaussian mixed distribution. The Gaussian mixed distribution is expressed by the following equation:
The Gaussian mixture distribution is determined by the mean μ _k of each distribution, the maximum and minimum standard deviations σ _max and σ _min , the stability s, and the weight π _k of each distribution. Therefore, these are included as parameters to be optimized. Note that the weight π _k has the following constraints:
Σ _k Π _k =1

The image data of the flaw shown in Figure 3 is generated based on the parameters described above. These parameters are collectively referred to as xk (k = a natural number from 1 to K). The parameters xk become the coordinate axes of the parameter space PS shown in Figure 1. In other words, the parameter space PS is a K-dimensional space. The vector Vn (n = a natural number from 1 to N) representing each point in this parameter space PS represents a combination of the setting values of the parameter xk. In this embodiment, this vector Vn is considered to be one individual. Hereinafter, the vector Vn will also be referred to as the individual Vn.

In the parameter space PS, a probability distribution PD of vector Vn, which is a combination of parameters xk, is defined. The probability distribution PD is a multivariate normal distribution. When the probability distribution PD is determined, the training data generation model 10 generates populations V1 to VN of vector Vn according to this probability distribution PD. The populations V1 to VN are extracted randomly according to the probability distribution PD. In this case, the extracted individuals Vn become denser towards the center of the ellipse.

In other words, the training data generation model 10 regards each individual as a vector Vn (n = 1 to N) in a parameter space PS, whose coordinate axes are the parameter xk that defines the feature, and generates populations V1 to VN according to a probability distribution PD defined in the parameter space PS. Furthermore, the training data generation model 10 generates image data TDn defined by the set value of the parameter xk corresponding to each individual Vn in the populations V1 to VN.

As described above, in this embodiment, the training data generation model 10 generates scratch image data TDn indicating scratches defined by the set value of parameter xk corresponding to each individual Vn, and generates image data TDn as training data by overlaying the feature image data (scratch image data) on reference image data of the object to be classified that does not have scratches, as shown in Figure 3.

[Judgment section]
The determination unit 11 inputs the image data TDn generated by the training data generation model 10. The determination unit 11 has the user S determine the likelihood of each piece of image data TDn representing a feature (flaw), i.e., how appropriately the feature (flaw) is represented. The likelihood of the data representing a feature indicates the degree to which the user S sees the classification target as actually having a feature, even if the synthesized image data is obtained by combining image data representing a feature (flaw) with image data of the classification target that does not have the feature.

In this embodiment, the judgment unit 11 has the user S judge whether or not a feature (blemish) is present. Specifically, as described below, the training data generation system 1 has an operation unit 34 and a display unit 35 as a man-machine interface. The judgment unit 11 causes the display unit 35 to display an image based on image data (synthetic image data in FIG. 3) TDn generated by the training data generation model 10. The user S looks at the image based on the image data TDn displayed on the display unit 35, judges whether or not the image has a feature (blemish), and inputs the judgment result to the operation unit 34. As shown in FIG. 7, the judgment unit 11 links the image data and the judgment result of the user S as a teacher signal and sends it to the optimization unit 12.

[Optimization section]
The optimization unit 12 performs an interactive optimization process to estimate a probability distribution PD so that the number of image data TDn determined to appropriately represent the feature (flaw) increases or the degree to which the feature (flaw) in the image data TDn is appropriately represented increases, while repeating the generation of image data TDn by the training data generation model 10 and the judgment of the image data TDn by the judgment unit 11. The interactive optimization process is a process that uses an optimization method that searches for an optimal value using a metaheuristic, multi-point search method through repeated dialogue with a user.

In this embodiment, the optimization unit 12 performs interactive evolutionary computation using the covariance matrix adaptive evolution strategy (CMA-ES) as the optimization process. Unlike general evolution strategies, CMA-ES does not directly store populations V1 to VN that are solution candidates, but instead generates a population of solution candidates using a multivariate normal distribution PD(m, σ ² , C), and searches for a solution by updating the mean m, standard deviation σ, and covariance matrix C based on candidates that receive high evaluations using an objective function.

In CMA-ES, for example, we will explain how to update the probability distribution PD from generation g to generation g+1, as shown in Figure 8. As the generation progresses, the mean m of the individual Vn (referred to as the "elite") determined to have a defect among the generated populations V1 to VN in that probability distribution PD is updated to the mean m of the next generation. Furthermore, the covariance matrix C is updated to a covariance matrix C that combines a Rank-μ-update obtained by adding a covariance matrix based on the difference between the current elite and the current generation's m to the covariance matrix of the current generation, and a Rank-One-Update obtained based on the trajectory of the covariance matrix C up to that point. Furthermore, the standard deviation σ is updated by decreasing it as the length of the inter-generational shift of the probability distribution PD becomes shorter, and increasing it as the length of the shift becomes longer.

In this way, the optimization unit 12 searches for the multivariate normal distribution PD(m, σ ² , C) while updating the mean m, covariance matrix C, and standard deviation σ of the multivariate normal distribution PD(m, σ ² , C) so that the number of image data TDn determined to be likely to include a scratch increases or the likelihood that the image data TDn represents a scratch increases. Basically, in the initial stage, a search is performed over a wide area, so a large standard deviation σ is used. As the multivariate normal distribution PD(m, σ ² , C) approaches the area where the optimal solution exists, the standard deviation σ decreases and the movement of the mean m also decreases. This makes it possible to obtain an optimal multivariate normal distribution PD(m, σ ² , C). The generation of the solution candidate is updated, and the optimization unit 12 terminates the optimization process when the generation has been updated a predetermined number of times, for example.

The multivariate normal distribution PD(m, σ ² , C) of the final generation finally defined by the optimization unit 12 becomes the optimized probability distribution PD. The training data generation model 10 generates image data with a teacher signal defined by the setting values of the parameter xk corresponding to each individual in the population according to the optimized probability distribution PD. The deep learning device performs learning using the image data with the teacher signal generated in this way. Note that the number of individuals in the population does not need to be the same as N during search, and image data of a larger number of individuals can be generated.

[Hardware configuration]
The training data generation system 1 shown in Fig. 1 is realized, for example, by a computer having the hardware configuration shown in Fig. 9 executing a software program. Specifically, the training data generation system 1 is composed of a CPU (Central Processing Unit) 21 that controls the entire device, a main memory unit 22 that operates as a work area for the CPU 21, an external memory unit 23 that stores the operating program of the CPU 21, an operation unit 24, a display unit 25, an input/output unit 26, and an internal bus 28 that connects these units.

The main memory unit 22 is composed of RAM (Random Access Memory) and other components. Programs 29 executed by the CPU 21 are loaded into the main memory unit 22 from the external memory unit 23. The main memory unit 22 is also used as a working area (temporary data storage area) for the CPU 21.

The external storage unit 23 is composed of non-volatile memory such as flash memory or a hard disk. The external storage unit 23 pre-stores a program 29 to be executed by the CPU 21.

The operation unit 24 consists of devices such as a keyboard and mouse, and an interface device that connects these devices to the internal bus 28.

The display unit 25 consists of a display device such as a CRT (Cathode Ray Tube) or LCD monitor.

The input/output unit 26 is an interface for sending and receiving data with external devices. Image data 30 (e.g., the reference image data and composite image data shown in Figure 3) used to generate training data is input and output via this input/output unit 26.

The functions of training data generation system 1 can be implemented in a computer system consisting of one or more computers, each including one or more processors and one or more storage devices, including non-transitory storage media. The multiple computers communicate with each other via an interconnected communications network to realize the functions of training data generation system 1. For example, some of the functions of training data generation system 1 may be implemented in one computer, and other parts may be implemented in other computers.

[Operation]
Next, the operation of the training data generation system 1 according to this embodiment will be described. As shown in Fig. 10, first, the training data generation model 10 generates populations V1 to VN according to a probability distribution defined in a parameter space whose coordinate axes are parameters defining flaws (step S1). Next, the training data generation model 10 generates training data defined by the set value of the parameter xk corresponding to each individual Vn in the populations V1 to VN (step S2).

Next, the judgment unit 11 prompts the user to judge whether or not there is a scratch for each piece of training data generated by the training data generation model (step S3).

Next, the optimization unit 12 updates the probability distribution PD (step S4). After that, the optimization unit 12 determines whether the termination condition is met (step S5). If the termination condition is not met (step S5; No), the process returns to step S1. After that, steps S1 to S5 are repeated as long as the termination condition is not met (step S5; No).

If the termination condition is met (step S5; No), the optimization unit 12 stores the final probability distribution PD (step S6) and terminates the process.

The training data generation system 1 according to this embodiment generates image data with teacher signals to be used in training a deep learning device that inspects scratches on the painted surface of a vehicle body. However, this is not limited to this. It can also be applied to a deep learning device that inspects other inspection objects using images. For example, it can be used to inspect cracks in concrete, and scratches on the surface of films and rubber sheets. It can also be applied to the maintenance and inspection of transportation infrastructure and buildings.

[Inspection of the shape of the item]
Furthermore, the training data generation system 1 according to this embodiment can be applied to a deep learning device that classifies objects based on the presence or absence of features. For example, it can also be applied to inspecting the shape of three-dimensional objects. There are no particular restrictions on the three-dimensional objects, and it can be used to inspect any three-dimensional object, including commercial products and stationery. In this case, the training data generation model 10 generates image data of the objects to be classified that represent features defined by the set values of the parameter xk corresponding to each individual Vn.

The image data generated in this case will be image data that reveals the shape of the object to be classified. Such image data may be processed image data, or may be an image created entirely using computer graphics. The feature will be the shape of the object to be classified. In addition to shape, structure, combination of parts, pattern, color, or a combination of these can also be used as features that serve as the basis for classification.

The training data generation system 1 according to this embodiment can also be applied to sequence data, such as time-series data, such as meteorological observation data, oceanographic observation data, or medical data. In this case, the training data generation model 10 may generate sequence data representing characteristics defined by the parameter settings corresponding to each individual Vn.

In this embodiment, the determination unit 11 has the user S determine whether or not a feature (flaw) is present, but this is not limited to this. For example, the user S may be asked to select an image that is likely to contain the feature (flaw), or to score or rank the image's likelihood of containing the feature (flaw), and to determine the degree to which the image appropriately represents the feature (flaw). Alternatively, two pieces of image data may be presented to the user S, and the user may be asked to select which of the images is most likely to contain a genuine feature (flaw). When the user S is asked to score or rank the image's likelihood of containing the feature (flaw), the optimization unit 12 performs optimization so that the score or rank of the image's likelihood of containing the feature (flaw) is high.

As explained in detail above, the training data generation system 1 according to the above embodiment repeats the generation of training data and the user S's determination of the likelihood of image data TDn representing a feature, while performing an interactive optimization process to estimate the probability distribution of parameters that define the features so as to increase the number of image data TDn that are determined to be likely to represent a feature, thereby designing a generation model that generates training data with teacher signals. This makes it possible for the user to generate a huge amount of training data with teacher signals in a short amount of time with little burden, without having any computer skills.

Note that the training data generation system 1 according to the above embodiment uses a covariance matrix adaptive evolution strategy to search for the probability distribution of optimal parameter settings through interactive evolutionary computation. However, this is not limited to this. Optimization algorithms that can be used include the Estimation of Distribution Algorithm (EDA), Bayesian optimization algorithm, Population-Based Incremental Learning (PBIL), and Compact Genetic Algorithm (CGA).

The hardware and software configurations of the training data generation system 1 are examples and can be changed and modified as desired.

The core processing portion of training data generation system 1, which is composed of CPU 21, main memory 22, external memory 23, operation unit 24, display unit 25, input/output unit 26, and internal bus 28, can be realized using an ordinary computer system rather than a dedicated system. For example, a computer program for executing the above operations may be stored and distributed on a computer-readable recording medium (such as a flexible disk, CD-ROM, or DVD-ROM), and the computer program may be installed on a computer to configure training data generation system 1 for executing the above processing. Alternatively, the computer program may be stored in a storage device of a server device on a communications network such as the Internet, and downloaded by an ordinary computer system to configure training data generation system 1.

When the functions of the training data generation system 1 are realized by sharing the responsibilities between an OS (operating system) and an application program, or by collaboration between an OS and an application program, only the application program portion may be stored on a recording medium or storage device.

It is also possible to superimpose a computer program onto a carrier wave and distribute it over a communications network. For example, the computer program may be posted on a bulletin board system (BBS) on the communications network and distributed over the network. This computer program may then be launched and run under the control of the OS in the same way as any other application program, thereby enabling the above-mentioned processing to be performed.

This invention allows for various embodiments and modifications without departing from the broad spirit and scope of the invention. Furthermore, the above-described embodiments are intended to explain the invention and do not limit the scope of the invention. In other words, the scope of the invention is defined by the claims, not the embodiments. Various modifications made within the scope of the claims and the meaning of the invention equivalent thereto are deemed to be within the scope of the invention.

The present invention can be applied to generating training data used in learning deep learning machines.

1 Training Data Generation System, 10 Training Data Generation Model, 11 Judgment Unit, 12 Optimization Unit, 21 CPU, 22 Main Memory Unit, 23 External Memory Unit, 24 Operation Unit, 25 Display Unit, 28 Internal Bus, 29 Program, 30 Image Data, S User, xk Parameters, Vn Individuals, PD Probability Distribution, PS Parameter Space

Claims (6)

A training data generation system that generates training data with a teacher signal used for learning a deep learning device that classifies objects to be classified based on the presence or absence of features,
a training data generation model that defines an individual as a vector in a parameter space whose coordinate axes are parameters defining the characteristics, generates a population according to a probability distribution defined in the parameter space, and generates training data defined by setting values of parameters corresponding to each individual in the population;
a determination unit that allows a user to determine how appropriately each piece of training data generated by the training data generation model represents a feature;
an optimization unit that performs an interactive optimization process to estimate the probability distribution while repeating the generation of training data by the training data generation model and the determination by the determination unit, so that the number of training data that are determined to appropriately represent the features increases or the degree to which the features in the training data are appropriately represented increases;
Equipped with
The optimization unit
As the optimization process, interactive evolutionary computation using a covariance matrix adaptive evolution strategy is performed.
Training data generation system. The training data generation model is
generating characteristic image data indicating characteristics defined by the setting values of the parameters corresponding to each of the individuals;
generating the training data by superimposing the feature image data on reference image data to be classified that does not have the feature;
The training data generation system of claim 1 . the classification object is the outer surface of the inspection object,
The feature is a scratch or crack that appears on the outer surface,
the reference image data is image data of an outer surface of the inspection object that does not have the feature;
The feature image data is image data of the feature defined by a set value of a parameter corresponding to the individual.
The training data generation system of claim 2 . The training data generation model is
generating image data to be classified that represent features defined by the parameter setting values corresponding to each of the individuals;
The training data generation system of claim 1 . The characteristics are characteristics related to the shape, structure, combination of parts, pattern, color, or a combination thereof of the classification object.
The training data generation system of claim 4 . the training data is sequence data to be classified,
The training data generation model is
generating the sequence data representing features defined by parameter setting values corresponding to each of the individuals;
The training data generation system of claim 1 .