JP2024122709A - Distortion area estimation device, distortion area estimation method, machine learning device, and machine learning method - Google Patents

Abstract

[Problem] To provide a technology for detecting visual distortions caused by optical illusions that may occur in a three-dimensional surface model. [Solution] A distortion area estimation device (1) includes a control unit (processor 12) that executes an estimation step of estimating a distortion area that may be included in a three-dimensional surface model and that appears visually distorted due to an optical illusion. [Selected Figure] Figure 1

Description

The present invention relates to a distortion area estimation device and a distortion area estimation method for estimating a distortion area in a three-dimensional surface model. The present invention also relates to a machine learning device and a machine learning method for constructing a trained model for estimating such a distortion area.

Three-dimensional CAD (Computer Aided Design) is widely used to design vehicle parts (including interior and exterior parts) such as passenger cars and work vehicles (see, for example, Patent Document 1). For each part, the exterior design and internal structure that cannot be seen from the outside are designed using 3D CAD. 3D CAD generates data that represents a 3D model of each designed part, and visualizes the 3D model. In addition, during the manufacturing process for each part, each part can be manufactured based on the 3D model data.

In the following, the three-dimensional model that represents the surfaces that make up the exterior will be referred to as a surface model.

Japanese Patent Application Publication No. 9-326046

As mentioned above, when a part is designed based on a 3D model of the part designed using 3D CAD, some of the surfaces that make up the exterior may appear unnaturally distorted, contrary to the intention at the time of design. This type of distortion is thought to be caused by an optical illusion that causes some of the surfaces to appear visually distorted.

In order to eliminate such distortions in parts, it is essential to visually check the surface model on the 3D CAD. This is because the distortions observed after the parts are manufactured also occur, to varying degrees, in the surface model.

Visual inspection of surface models is time-consuming because it must be done manually. Furthermore, the degree of distortion that can be detected by visual inspection is likely to vary depending on the skill level of the worker performing the visual inspection. In this way, visual inspection of surface models can easily lead to increased costs and poor aesthetics for parts and products.

One aspect of the present invention has been made in consideration of the above-mentioned problems, and its purpose is to provide a technology for detecting visual distortions that may occur in three-dimensional surface models and that are caused by optical illusions.

In order to solve the above problem, the distortion area estimation device according to the first aspect of the present invention includes a control unit that executes an estimation step of estimating a distortion area that may be included in a three-dimensional surface model and that appears visually distorted due to an optical illusion.

In order to solve the above problem, a distortion area estimation method according to a ninth aspect of the present invention includes an estimation step of estimating a distortion area that may be included in a three-dimensional surface model and that appears visually distorted due to an optical illusion.

In order to solve the above problem, the machine learning device according to the thirteenth aspect of the present invention includes a control unit that executes a preprocessing step and a construction step of constructing a trained model that estimates a distortion area that may be included in a line included in a three-dimensional surface model and that appears visually distorted due to an optical illusion, by supervised learning using a training dataset, the preprocessing step includes a feature derivation step of setting a plurality of unit intervals in the line included in the surface model and deriving features that indicate the characteristics of the shape of each unit interval, the input of the trained model being the features derived for each unit interval, and the output of the trained model being an estimation result of whether or not each unit interval contains the distortion area.

In order to solve the above problem, the machine learning method according to the 14th aspect of the present invention includes a preprocessing step and a construction step of constructing a trained model that estimates a distortion area that may be included in a line included in a three-dimensional surface model and that appears visually distorted due to an optical illusion by supervised learning using a training dataset, the preprocessing step includes a feature derivation step of setting a plurality of unit intervals in the line included in the surface model and deriving features that indicate the shape characteristics of each unit interval, the input of the trained model being the features derived for each unit interval, and the output of the trained model being an estimation result of whether or not each unit interval contains the distortion area.

The distortion area estimation device according to each aspect of the present invention may be realized by a computer. In this case, the distortion area estimation program of the distortion area estimation device, which causes the computer to operate as each unit (software element) of the distortion area estimation device to realize the distortion area estimation device, and the computer-readable recording medium on which the program is recorded, also fall within the scope of the present invention.

The machine learning device according to each aspect of the present invention may be realized by a computer. In this case, the distortion machine learning program of the machine learning device that realizes the machine learning device on a computer by making the computer operate as each part (software element) of the machine learning device, and the computer-readable recording medium on which it is recorded, also fall within the scope of the present invention.

According to one aspect of the present invention, a technology can be provided for detecting visual distortions that may occur in a three-dimensional surface model and that are caused by optical illusions.

1 is a block diagram showing a configuration of a distorted area estimation system according to an embodiment of the present invention. 2 is a flowchart showing the flow of a distorted area estimation method executed by a distorted area estimation device included in the distorted area estimation system shown in FIG. 1 . 3 is a flowchart showing a flow of pre-processing steps included in the distortion region estimation method shown in FIG. 2 . 2 is a perspective view of a surface model of a part for which the presence or absence of distortion is estimated by the distortion area estimation device shown in FIG. 1 . FIG. 5 is an enlarged perspective view of a portion of the surface model shown in FIG. 4 . 5 is an example of a cross-sectional profile of the surface model shown in FIG. 4. 5 is an enlarged perspective view of a portion of the surface model shown in FIG. 4, illustrating points determined in the process of a feature amount derivation step included in the distortion region estimation method shown in FIG. 2; FIG. 3 is a schematic diagram for explaining an example of a feature amount derived in a feature amount derivation step included in the distortion region estimation method shown in FIG. 2 . 2 is a flowchart showing the flow of a machine learning method executed by a machine learning device included in the distortion region estimation system shown in FIG. 1 . The upper part is a scatter plot in which the PCA feature quantities obtained in a modified example of the estimation step are plotted on a two-dimensional plane, and the lower part is a scatter plot obtained by enlarging a part of the upper scatter plot.

[Distortion area estimation system]
A distorted area estimation system S including a distorted area estimation device 1 according to an embodiment of the present invention will be described with reference to Fig. 1. Fig. 1 is a diagram showing the configuration of the distorted area estimation system S.

The distortion area estimation system S is a system for estimating the presence or absence of visual distortion caused by an optical illusion, which is a distortion that may occur in a three-dimensional surface model. As shown in FIG. 1, the distortion area estimation system S includes a distortion area estimation device 1, a three-dimensional CAD (Computer Aided Design) 2, and a machine learning device 3 (see FIG. 9).

The distortion area estimation system S, which includes the distortion area estimation device 1, is intended to detect visual distortions caused by optical illusions that may occur in three-dimensional surface models of vehicle parts (including interior and exterior parts) such as passenger cars and work vehicles during the design stage of the vehicle parts. Examples of vehicles include construction machinery and agricultural machinery in addition to passenger cars and work vehicles. Furthermore, the vehicle may be a vehicle that is operated by a user or a vehicle that is operated by automatic driving. In this embodiment, the distortion area estimation system S and the distortion area estimation device 1 will be described using an exterior part of a passenger car as an example of a part.

The part is designed using 3D CAD 2. That is, 3D CAD 2 generates a three-dimensional surface model of the part and stores data representing this surface model in a storage provided in 3D CAD 2 or in a storage provided outside 3D CAD 2. In this embodiment, it is assumed that data representing the surface model is stored in the storage of 3D CAD 2. Note that a surface model is one aspect of a three-dimensional model, and is a model that expresses the appearance (design) of a part using surfaces whose thickness is not defined.

The distortion area estimation device 1 includes a control unit (processor 12, described later) that executes an estimation step of estimating a distortion area that may be included in a three-dimensional surface model and that appears visually distorted due to an optical illusion. This estimation step may be configured to estimate a distortion area that may be included in a line included in the surface model and that appears visually distorted due to an optical illusion using a trained model constructed by supervised learning, or may be configured to estimate a distortion area using a trained model constructed by unsupervised learning, or may be configured to estimate a distortion area using a predetermined algorithm (e.g., a principal component analysis (PCA) algorithm) executed according to a program.

<Outline of the distortion area estimation device>
The distorted area estimation device 1 of this embodiment is a device for executing the distorted area estimation method M1. The distorted area estimation method M1 is a method for estimating a distorted area that may be included in a line included in a surface model using a learned model LM constructed by machine learning based on data representing a surface model provided from a three-dimensional CAD 2. Here, the distorted area means an area that appears visually distorted due to an optical illusion. In addition, in this embodiment, a cross-sectional profile of the surface model is used as the line included in the surface model. However, the line included in the surface model is not limited to the cross-sectional profile, and may be, for example, a constituent line on a surface that constitutes the surface model. The constituent line can also be referred to as a boundary line or a ridge line on a surface that constitutes the surface model.

Details of the algorithm of the learned model LM, the configuration of the distortion region estimation device 1, and the flow of the distortion region estimation method M1 will be described later.

The input of the learned model LM is the feature values derived for each of the multiple unit sections set in the cross-sectional profile of the surface model. These feature values can be appropriately selected to indicate the shape characteristics of each unit section.

The output of the trained model LM is an estimation result of whether or not each unit section of the cross-sectional profile described above contains a distorted area. For example, if it is determined that a certain unit section contains a distorted area, the trained model LM outputs "1", and if it is determined that it does not contain a distorted area, the trained model LM outputs "0".

The inventors of the present application have found that, although it is difficult to explicitly specify as a relational equation between the feature values that are input to the trained model LM and the estimation result of whether or not a distorted region is included that is the output of the trained model LM, there is a certain relationship. Therefore, by using the trained model LM that takes feature values as input, it is possible to accurately estimate whether or not each unit section of the cross-sectional profile includes a distorted region.

<Overview of machine learning device>
The machine learning device 3 is a device for executing the machine learning method M3. The machine learning method M3 is a method for creating a learning data set DS using data representing a surface model provided from the 3D CAD 2, and for constructing a learned model LM by machine learning using the learning data set DS. The configuration of the machine learning device 3 and the flow of the machine learning method M3 will be described in detail later.

The distortion region estimation system S goes through a preparation phase and a trial phase before arriving at a practical phase. The contents of the preparation phase, trial phase, and practical phase can be briefly explained as follows.

(1) Preparation Phase In the preparation phase, an operator determines whether or not a distortion region is included in the surface model. Each time the operator determines whether or not a distortion region is included in the surface model, the machine learning device 3 creates learning data (teacher data) from data representing the surface model provided by the three-dimensional CAD 2, and adds the created learning data to the learning dataset DS. The preparation phase may end when a predetermined period (e.g., one week, one month, or one year) has elapsed since the start of the preparation phase, or when the number of times the distortion region is estimated in the preparation phase reaches a predetermined number (e.g., 100 times, 1000 times, or 10,000 times). When the preparation phase ends, the machine learning device 3 constructs a trained model LM by machine learning using the training data DS. The constructed trained model LM is transferred from the machine learning device 3 to the distortion region estimation device 1.

(2) Trial Phase In the trial phase, the operator judges whether or not the surface model includes a distorted area, and the distorted area estimation device 1 estimates whether or not the line included in the surface model includes a distorted area. Every time the operator performs a measurement to judge whether or not the surface model includes a distorted area, the distorted area estimation device 1 estimates whether or not the surface model includes a distorted area using the learned model LM based on data representing the surface model provided from the three-dimensional CAD 2. The trial phase may end when a predetermined period (e.g., one week, one month, or one year) has elapsed since the start of the trial phase, or when the number of estimations of whether or not the surface model includes a distorted area in the trial phase reaches a predetermined number (e.g., 100 times, 1000 times, or 10,000 times). When the trial phase ends, the operator compares the judgment result he or she made with the estimation result of whether or not the distorted area is included estimated by the distorted area estimation device 1, and evaluates the estimation accuracy of the distorted area estimation device 1. If the estimation accuracy is insufficient, the process returns to the preparation phase. If the estimation accuracy is sufficient, the process proceeds to the practical phase. Note that the estimation accuracy may be confirmed by using a part of the learning dataset DS at the end of the preparation phase. In this case, the trial phase can be omitted.

(3) Practical Phase In the practical phase, the distorted region estimation device 1 estimates whether or not the surface model includes a distorted region. The learned model LM used by the distorted region estimation device 1 in the practical phase has been confirmed to have sufficient estimation accuracy in the trial phase. In the practical phase, the determination by the operator as to whether or not the surface model includes a distorted region can be omitted. This relieves the operator from the trouble of determining whether or not the surface model includes a distorted region, and makes it possible to estimate with high accuracy whether or not the surface model includes a distorted region.

<Configuration of the distortion region estimation device>
The configuration of the distorted region estimation device 1 will be described with reference to Fig. 1. The upper part of Fig. 1 is a block diagram showing the configuration of the distorted region estimation device 1.

As shown in FIG. 1, the distortion region estimation device 1 includes a memory 11, a processor 12, and a storage 13. The memory 11, the processor 12, and the storage 13 are connected to each other via a bus (not shown). This bus may further include an input/output interface (not shown) and a communication interface (not shown). This input/output interface is used, for example, to input data representing a surface model from the 3D CAD 2 to the distortion region estimation device 1, or to output a judgment result from the distortion region estimation device 1 to an external device (e.g., the 3D CAD 2). In addition, this communication interface is used, for example, to acquire a learned model LM from an external device (e.g., the machine learning device 3).

The memory 11 is configured to store the learned model LM constructed by machine learning. The learned model LM is an algorithm that inputs the feature amount of each unit section of a line included in a surface model (in this embodiment, the cross-sectional profile of the surface model), which is a feature amount derived by a preprocessing step described later, and outputs an estimation result of whether or not each unit section includes a distortion area. Note that, for example, a semiconductor RAM (Random Access Memory) can be used as the memory 11. Gradient boosting can be preferably used as the machine learning algorithm of the learned model LM. In addition, algorithms such as a Bayesian Gaussian mixture model and a support vector machine can be used as the machine learning algorithm of the learned model LM. Hereinafter, the machine learning algorithm is also simply referred to as the algorithm.

The processor 12 is configured to execute a distortion region estimation method M1, which will be described later, using the learned model LM stored in the memory 11. The processor 12 may be, for example, a central processing unit (CPU), a graphic processing unit (GPU), a microprocessor, a digital signal processor, a microcontroller, an application specific integrated circuit (ASIC) such as a tensor processing unit (TPU), or a combination of these. The processor 12 functions as a control unit in one aspect of the present invention, and may also be referred to as a "computing device."

The storage 13 is configured to store (non-volatilely store) the learned model LM. When estimating whether each unit section includes a distortion region, the processor 12 deploys the learned model LM stored in the storage 13 on the memory 11 and uses it. Note that the storage 13 may be, for example, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination of these.

In this embodiment, a configuration is adopted in which the distortion region estimation method M1 is executed using a single processor (processor 12), but the present invention is not limited to this. That is, a configuration in which the distortion region estimation method M1 is executed using multiple processors may be adopted. For example, the preprocessing step S11 (described later) may be executed in a first processor, and the estimation step S12 (described later) may be executed in a second processor. In this case, the multiple processors that cooperate to execute the distortion region estimation method M1 may be provided in a single computer and configured to be able to communicate with each other via a bus, or may be provided in a distributed manner in multiple computers and configured to be able to communicate with each other via a network. As an example, a processor built in a computer that constitutes a cloud server and a processor built in a computer owned by a user of the cloud server may cooperate to execute the distortion region estimation method M1.

In addition, in this embodiment, a configuration is adopted in which the learned model LM is stored in memory 11 built into the same computer as the processor (processor 12) that executes the distortion area estimation method M1, but the present invention is not limited to this. In other words, a configuration may be adopted in which the learned model LM is stored in memory built into a computer different from the processor that executes the distortion area estimation method M1. In this case, the computer with the built-in memory that stores the learned model LM is configured to be able to communicate with the computer with the built-in processor that executes the distortion area estimation method M1 via a network. As an example, a mode can be considered in which the learned model LM is stored in memory built into a computer that constitutes a cloud server, and a processor built into a computer owned by a user of the cloud server executes the distortion area estimation method M1.

In addition, in this embodiment, a configuration is adopted in which the learned model LM is stored in a single memory 11, but the present invention is not limited to this. That is, a configuration may be adopted in which the learned model LM is distributed and stored in multiple memories. In this case, the multiple memories that store the learned model LM may be provided in a single computer (which may or may not be a computer with a built-in processor that executes the distortion area estimation method M1), or may be distributed and provided in multiple computers (which may or may not include a computer with a built-in processor that executes the distortion area estimation method M1). As an example, a configuration may be considered in which the learned model LM is distributed and stored in memories built into each of the multiple computers that make up the cloud server.

<Flow of the distortion area estimation method>
The flow of the distortion area estimation method M1 will be described with reference to FIG. 2 to FIG. 8. FIG. 2 is a flowchart showing the flow of the distortion area estimation method M1. FIG. 3 is a flowchart showing the flow of the pre-processing step S11 included in the distortion area estimation method M1. FIG. 4 is a perspective view of a surface model SM of a part for which the presence or absence of distortion is estimated by the distortion area estimation device 1. FIG. 5 is a perspective view of an enlarged portion of the surface model SM. FIG. 6 is an example of a cross-sectional profile of the surface model SM. FIG. 7 is a perspective view of an enlarged portion of the surface model SM, illustrating a point determined in the process of the feature amount derivation step S112 included in the distortion area estimation method M1. FIG. 8 is a schematic diagram for explaining an example of a feature amount derived in the feature amount derivation step S112. As shown in FIG. 4, FIG. 5, and FIG. 7, in this embodiment, an orthogonal coordinate system defined by the x-axis, y-axis, and z-axis in the space in which the surface model SM is handled is used. However, the coordinate system used in the distortion area estimation method M1 and the machine learning method M3 described later is not limited to the orthogonal coordinate system. The coordinate system used in the machine learning method M3 may be, for example, an oblique coordinate system, a polar coordinate system, or a general coordinate system. An example of the general coordinate system is a cylindrical coordinate system.

The distortion area estimation method M1 includes a preprocessing step S11 and an estimation step S12.

(Pre-processing step)
The pre-processing step S11 is a step in which the processor 12 derives features for each unit section of the cross-sectional profile of the surface model to be input to the learned model LM. As shown in FIG. 3, the pre-processing step S11 includes a cross-sectional profile generation step S111 and a feature derivation step S112.

The cross-sectional profile generating step S111 is a step of generating a plurality of cross-sectional profiles from the surface profile of the surface model SM.
The multiple cross-sectional profiles generated in the cross-sectional profile generating step S111 are preferably cross sections expressed in an oblique coordinate system formed by three mutually intersecting axes, and are cross-sectional profiles in multiple cross sections parallel to a plane defined by two of the three axes. In addition, it is more preferable that the intervals between adjacent cross-sectional profiles are equal.

As described above, in this embodiment, an orthogonal coordinate system is used with three axes, the x-axis, the y-axis, and the z-axis, which are orthogonal to each other. In addition, as the multiple cross-sectional profiles, (1) multiple cross sections C _xy parallel to the xy plane and equally spaced, (2) multiple cross sections C _yz parallel to the yz plane and equally spaced, and (3) multiple cross sections C _zx parallel to the zx plane and equally spaced are used (see FIGS. 4 and 5). In addition, the interval d _yz between adjacent cross sections C _yz is set to 20 mm (see FIG. 5). The interval between adjacent cross sections C _xy and the interval between adjacent cross sections C _zx are also set to 20 mm, similar to the cross section C _yz . However, these intervals can be appropriately determined according to the size of the expected distortion region, etc.

In addition, when it is known from the shape of the part which of the cross _{sections Cxy} , _Cyz , and _Czx is the cross section that is easy to estimate the distortion area, it is also possible to select one of the cross sections _Cxy , _Cyz , and _Czx and use only the cross section profile of that cross section. However, by using all the cross section profiles of the cross sections _Cxy , _Cyz , and _Czx , it becomes unnecessary to be concerned about the direction of the coordinate axes of each part, and it is possible to reduce overlooking of distortion areas.

The feature amount derivation step S112 is a step of setting a plurality of unit sections in the line included in the surface model SM, and deriving a feature amount indicating a feature of the shape of each unit section. In this embodiment, the cross-sectional profiles in each of the cross sections C _xy , C _yz , and C _zx shown in Figs. 4 and 5 are adopted as the lines included in the surface model SM. Fig. 6 shows an example of the cross-sectional profile in the cross section C _yz . Taking the cross-sectional profile shown in Fig. 6 as an example, the feature amount derivation step S112 sets a plurality of unit sections _Sei (i is an integer of 3≦i≦N, where N is a positive integer) in the cross-sectional profile in the cross section C _yz , and derives a feature amount of each unit section _Sei .
More specifically, the cross-sectional profile of the surface model SM is formed by serially connecting N curves, each of which is formed by a single spline curve. The feature amount deriving step S112 derives the feature amount of each unit interval, with the unit interval being a first interval formed by the i-th (i is an integer between 2 and N-1) spline curve, a second interval formed by the i-1-th spline curve, and a third interval formed by the _{i+1-th spline curve. For example, when the unit interval Sei+1} shown in FIG. 6 and FIG. 7 is the first interval, the second interval is the unit interval _Sei , and the third interval is the unit interval Sei ₊₂ .
In this embodiment, the first section is set as the center, and the second and third sections located before and after the first section are set as both ends to derive 20 feature amounts described below. Hereinafter, the 20 feature amounts derived using the three sections, with the first section set as the center, and the second and third sections located before and after the first section as both ends, are referred to as feature amounts of the first section.

Next, in the feature derivation step S112, M pseudo line segments obtained by dividing each of the first section (e.g., unit section Se _i+1 ), the second section (e.g., unit section Se _i ), and the third section (e.g., unit section Se _i+2 ) into M equal parts (M is an integer equal to or greater than 2) are defined as line segments L1j, L2j, and L3j (j is an integer equal to or greater than 1 and equal to or less than M). In the cross-sectional profile shown in FIG. 6 , the start point and the end point of each unit section Se _i+1 are defined as points P _i and P _i+1 . Then, five points obtained by dividing the line segment P _i P _i+1 into six equal parts are defined as points P _ij (see FIG. 7 ). Therefore, the line segment P _i P _ij of the unit section Se _i+1 , which is the first section, corresponds to the line segment L1j. Similarly, the line segment P _i P _ij of the unit section _Sei, which is the second section, corresponds to the line segment L2j, and the line segment P _i P _ij of the unit section Sei _+2, which is the third section, corresponds to the line segment L3j. Note that in the cross-sectional profile shown in FIG. 7, M=6 is used as described above. However, M is not limited to M=6, and can be appropriately selected depending on the size, design, etc. of the surface model SM. A typical example of M is M=50, but M may be about 10 or about 100.

In this embodiment, the method of deriving the 20 feature amounts will be described by taking as an example a case where a unit section Sei ₊₁ generalized using i is set as the first section in a cross-sectional profile in a cross-section _Cyz parallel to the yz plane. However, in the actual feature amount deriving step S112, the 20 feature amounts are repeatedly derived for a case where each unit section _Sei (i is an integer of 2 to N-1) in the cross-sectional profile is set as the first section. Also, similar to the case of the cross-sectional profile in the cross-section _Cyz , the 20 feature amounts are repeatedly derived for each unit section of the cross-sectional profile in each of the cross-sections _Cxy and _Czx .

The feature derivation step S112 derives, as features for each unit section, at least the maximum amount of change in the angle between adjacent line segments L1j and L1j+1 in the first section, the maximum amount of change in the angle between adjacent line segments L2j and L2j+1 in the second section, and the maximum amount of change in the angle between adjacent line segments L3j and L3j+1 in the third section.
Then, in the feature amount derivation step S112, the maximum amount of change in the angle between adjacent line segments L1j and L1j+1 in the first section is determined as the first feature amount. Also, in the feature amount derivation step S112, the maximum amount of change in the angle between adjacent line segments L2j and L2j+1 in the second section is compared with the maximum amount of change in the angle between adjacent line segments L3j and L3j+1 in the third section, and the smaller of these is determined as the second feature amount and the larger of these is determined as the third feature amount.

Here, the angle θ _ij between line segment Lij (line segment P _i P _ij ) and line segment Lij+1 (line segment P _i P _ij+1 ) in unit interval _Sei , which is the second interval, can be an acute angle or an obtuse angle, but in this embodiment, an acute angle is used as the angle θ _ij . Fig. 8 illustrates angles θ _i1 to θ _i4 . The change in angle θ _ij is obtained by dividing angle θ _ij by length _Li , which is obtained by dividing the length of unit interval _Sei into 6 equal parts.

The fourth feature is the magnitude relationship between the maximum angle change amount in the first interval, the maximum angle change amount in the second interval, and the maximum angle change amount in the third interval, expressed as one of the values 0, -1, or 1. If the absolute value of the maximum angle change amount in the first interval is small and the absolute values of the maximum angle change amount in the second and third intervals are large, 0 is adopted. Furthermore, in cases other than those in which 0 is adopted as described above, if the maximum angle change amount is positive, -1 is adopted. Examples of this case include (1) the maximum values of the angle change in the first to third sections are all positive, and the maximum values of the angle change become smaller in the order of the second section, the first section, and the third section; (2) the maximum values of the angle change in the second section and the third section are similar, and the maximum value of the angle change in the first section is greater than the maximum value of the angle change in the second section and the third section; and (3) the maximum value of the angle change in the second section is greater than the maximum value of the angle change in the first section and the third section, and the maximum values of the angle change in the first section and the third section are similar. The range in which the maximum values of the angle change are compared and determined to be similar can be appropriately determined. In addition, when the maximum values of the angle change are compared to determine whether they are similar, the magnitude relationship between the maximum values of the angle change to be compared does not matter. In addition, in cases that do not fall under the above-mentioned case of adopting 0, if the minimum value of the angle change is negative, 1 is adopted. Examples of this case include (4) the minimum values of the angle change in the first to third sections are all negative, and the minimum values of the angle change decrease in the order of the second section, the first section, and the third section; (5) the minimum values of the angle change in the second and third sections are similar, and the minimum value of the angle change in the first section is smaller than the minimum value of the angle change in the second and third sections; and (6) the minimum value of the angle change in the second section is smaller than the minimum value of the angle change in the first and third sections, and the minimum value of the angle change in the first and third sections is similar. In addition, further criteria may be set for each of the cases of adopting 0, adopting -1, and adopting 1.

In addition, the feature derivation step S112 sets the angle in the first section (D1) as the fifth feature, and sets the smaller (Ds) and larger (Db) of the angles in the second section and the third section as the sixth and seventh features, respectively.

The difference between Ds and D1 is the eighth feature, and the difference between Db and D1 is the ninth feature.

The absolute value of the difference between Db and Ds is the tenth feature.

The sum of the angles in the second section and the sum of the angles in the third section is the eleventh feature.

The sum of the angles in the first section, the second section, and the third section is defined as the twelfth feature.

The thirteenth feature is a value that expresses the magnitude relationship between the maximum angle in the first section, the maximum angle in the second section, and the maximum angle in the third section, using any one of the values 0, -1, and 1. If the absolute value of the maximum angle in the first section is small and the absolute values of the maximum angle in the second and third sections are large, 0 is adopted. If the maximum angle is positive in cases that do not fall under the above-mentioned cases of adopting 0, -1 is adopted. An example of this case is when the maximum angle values in the first to third sections are all positive, and the maximum angle values decrease in the order of the second section, the first section, and the third section. If the minimum angle value is negative in cases that do not fall under the above-mentioned cases of adopting 0, 1 is adopted. An example of this case is when the minimum angle values in the first to third sections are all negative, and the minimum angle values decrease in the order of the second section, the first section, and the third section. As with the fourth feature, further criteria may be set for each of the cases where 0 is adopted, -1 is adopted, and 1 is adopted.

Moreover, the length L _i in the first section (hereinafter, referred to as length L1) is defined as a fourteenth feature amount.

In addition, the length _Li in the second section (hereinafter referred to as length L2) is compared with the length _Li in the third section (hereinafter referred to as length L3), and the smaller one (Ls) and the larger one (Lb) are defined as the 15th feature and the 16th feature, respectively.

The ratio of Ls to L1 (ln(Ls/L1)) is the 17th feature.

The ratio of Lb to L1 (ln(Lb/L1)) is the 18th feature.

The ratio of Lb to Ls (ln(Lb/Ls)) is the 19th feature.

The 20th feature is the magnitude relationship between L1, L2, and L3 expressed as one of the values 0, -1, and 1. If L1 is small and L2 and L3 are large, or if L1 is large and L2 and L3 are small, 0 is adopted. If L2, L1, and L3 monotonically decrease in this order, -1 is adopted. If L2, L1, and L3 monotonically increase in this order, 1 is adopted. Note that further criteria may be set for each of the cases of adopting 0, -1, and 1.

As described above, in this embodiment, the feature derivation step S112 derives 20 feature values. However, the feature values derived by the feature derivation step S112 are not limited to the above 20. Some of the above feature values may be omitted, or other feature values may be added to the above feature values. The feature values derived by the feature derivation step S112 may be determined appropriately depending on the shape of the part, the response to possible distortion, etc.

(Estimation step)
In the estimation step S12, the processor 12 uses the learned model LM to estimate the presence or absence of a distorted area that may be included in a line (in this embodiment, a cross-sectional profile) included in the surface model SM and that appears visually distorted due to an optical illusion. In the estimation step S12, the processor 12 reads data representing the feature amount derived in the preprocessing step S11 from the memory 11, and inputs the data representing the read feature amount to the learned model LM. Then, the processor 12 writes the estimation result output from the learned model LM, that is, data representing the estimation result of whether or not each unit section includes the distorted area, to the storage 13. The form of the data representing the estimation result can be appropriately determined, but in this embodiment, if the unit section includes a distorted area, it is set to "1", and if the unit section does not include a distorted area, it is set to "0".

In the actual estimation step S12, for all cross-sectional profiles of cross sections C _xy , C _yz , and C _zx , 20 feature amounts are input for each unit section _Sei (i is an integer between 2 and N-1) as the first section, and an estimation result as to whether or not each unit section includes the distortion region is output.

The distortion region estimation method M1 may further include an output step of outputting data representing the estimation result estimated in the estimation step S12. In this output step, the processor 12 may be configured to read data representing the estimation result from the storage 13 and provide the read data representing the estimation result to the three-dimensional CAD 2. This allows the three-dimensional CAD 2 to differentiate the distortion region included in each unit section constituting each cross-sectional profile of the surface model SM from other regions. Also, in this output step, the processor 12 may be configured to present the estimation result read from the storage 13 to the user by outputting it to a display. This allows the user to know whether or not a distortion region is included in each unit section constituting each cross-sectional profile of the surface model SM.

<Machine learning device and machine learning method>
The configuration of the machine learning device 3 will be described with reference to Fig. 1 and Fig. 9. The lower part of Fig. 1 is a block diagram showing the configuration of the machine learning device 3. Fig. 9 is a flowchart showing the flow of a machine learning method M3 executed by the machine learning device 3.

The machine learning device 3 is a device that executes a machine learning method M3 for constructing a trained model LM. As shown in FIG. 5, the machine learning device 3 includes a storage 31, a processor 32, and a memory 33. The storage 31, the processor 32, and the memory 33 are connected to each other via a bus (not shown). This bus may further be connected to an input/output interface (not shown) and a communication interface (not shown). This input/output interface is used, for example, to input training data from an external device (e.g., a sensor and a keyboard) to the machine learning device 3. In addition, this communication interface is used, for example, to provide the trained model LM to an external device (e.g., the above-mentioned distortion area estimation device 1).

The storage 31 is configured to store the learning dataset DS. The learning dataset DS is a collection of training data in which data representing multiple feature amounts in each unit section constituting each cross-sectional profile of the surface model SM is labeled with an estimation result indicating whether or not the unit section contains a distortion area.

Here, attaching a label representing an estimation result to data representing multiple feature quantities refers to associating the data representing multiple feature quantities with the estimation result in an arbitrary manner. Examples of methods for associating the data representing multiple feature quantities with the estimation result include a method of creating a table that associates the data representing multiple feature quantities with the estimation result on a one-to-one basis, and a method of storing the data representing multiple feature quantities in a directory corresponding to the estimation result. Note that, for example, a flash memory, HDD, SSD, or a combination of these can be used as the storage 31.

As shown in FIG. 9, the machine learning method M3 includes a preprocessing step S31 and a construction step S32.
The pre-processing step S31 is the same as the pre-processing step S11 shown in Fig. 2. That is, the pre-processing step S31 is a step in which the processor 32 derives a plurality of feature amounts in each unit section of the cross-sectional profile of the surface model SM to be input to the learned model LM. In the pre-processing step S31, the processor 32 sets a set of data representing a plurality of feature amounts in each unit section as a learning dataset DS.

Each learning dataset DS contains data representing multiple feature quantities in each unit interval. The data representing multiple feature quantities in each unit interval contained in each learning dataset DS is data representing multiple feature quantities derived for each of the multiple unit intervals set in the cross-sectional profile of the surface model, similar to the data representing multiple feature quantities in each unit interval input to the trained model LM.

In the pre-processing step S31, a label representing the estimation result described above is further attached to the multiple feature amounts derived for each unit interval.

In the construction step S32, the processor 32 is configured to execute a construction process in which the learning data set DS stored in the storage 31 is expanded on the memory 33, and a learned model LM is constructed by supervised machine learning using the learning data set DS. As described above, the learned model LM is an algorithm that inputs a plurality of feature quantities derived for each of a plurality of unit sections set in the cross-sectional profile of the surface model, and outputs an estimation result of whether or not each unit section of the cross-sectional profile contains a distortion region. As the processor 32, for example, a CPU, a GPU, a microprocessor, a digital signal processor, a microcontroller, an ASIC such as a TPU, or a combination thereof can be used. The processor 32 functions as a control unit in one aspect of the present invention, and is sometimes called a "calculation device".
The memory 33 is configured to store the trained model LM obtained by executing the construction process by the processor 32. For example, a semiconductor RAM can be used as the memory 33. The trained model LM stored in the memory 33 may be stored (non-volatilely stored) in the storage 31 described above.

Note that, although the configuration has been described here in which the machine learning method M3, including the construction process for constructing the trained model LM, is executed by a single processor 32 provided in a single computer, the present invention is not limited to this. In other words, it is also possible to adopt a configuration in which the machine learning method M3 is executed jointly by multiple processors provided in a single computer or distributed across multiple computers.

Although the configuration in which the learning dataset DS is stored in a single storage 31 provided in a single computer has been described here, the present invention is not limited to this. In other words, it is also possible to adopt a configuration in which the learning dataset DS is stored in a single computer or in multiple storages distributed across multiple computers. Furthermore, the learning dataset DS does not need to be stored in the storage 31 built into the computer together with the processor 32 and memory 33, but may be stored in a cloud server configured to be able to communicate with the computer via a network.

In addition, although a configuration in which the trained model LM is stored in a single memory 33 provided in a single computer has been described here, this is not limiting. In other words, it is also possible to adopt a configuration in which the trained model LM is stored in a single computer or in multiple memories provided in a distributed manner across multiple computers.

The program for causing the processor 32 to execute the machine learning method M3 is recorded, for example, on a computer-readable, non-transitory, tangible recording medium. The processor 32 executes the determination method S1 by executing instructions contained in this program. This recording medium may be the storage 31, the memory 33, or another recording medium. For example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit may be used as the other recording medium.

According to the machine learning device 3, it is possible to construct a learned model LM used by the above-mentioned distortion region estimation device 1. Moreover, the teacher data used for machine learning is data representing a plurality of feature amounts in each unit section constituting each cross-sectional profile of the surface model SM, to which a label indicating an estimation result of whether or not the unit section contains a distortion region has been attached. Conventionally, the determination of whether or not the surface model SM contains a distortion region has been performed by an operator visually checking the surface model SM after displaying the surface model SM on a display. Therefore, when creating teacher data, an operator can easily determine whether or not a distortion region exists. Alternatively, the results of visual confirmation that have already been accumulated can be used as teacher data. Therefore, according to the machine learning device 3, it is possible to easily create highly accurate teacher data. The same can be said about the machine learning method M3 implemented using the machine learning device 3 and the program that operates a computer as the machine learning device 3.

In this embodiment, the machine learning method M3 for constructing the learned model LM and the estimation method for estimating the presence or absence of a distorted area using the learned model LM are described as being executed in two devices, respectively, but the present invention is not limited to this. In other words, the scope of the present invention also includes an embodiment in which the machine learning method M3 for constructing the learned model LM and the estimation method for estimating the presence or absence of a distorted area using the learned model LM are executed in one device.

[Modification of the Estimation Step]
The estimation step S12 of the distortion area estimation method M1 shown in FIG. 2 is configured such that the processor 12 uses the learned model LM to estimate the presence or absence of a distortion area that may be included in a line (in this embodiment, a cross-sectional profile) included in the surface model SM and that appears visually distorted due to an optical illusion.

However, a modified example of the estimation step S12 may be configured to estimate a distortion area that may be included in the surface model SM without using the learned model LM. In this case, the processor 12 performs a principal component analysis (PCA) on the multiple features (in this embodiment, 20 features) derived in the preprocessing step S11, and converts them into two-dimensional features (hereinafter referred to as PCA features).

The top part of Fig. 10 is a scatter plot in which the PCA features are plotted on a two-dimensional plane. The bottom part of Fig. 10 is a scatter plot that is an enlarged version of a portion of the top part of Fig. 10.

In a modified example of estimation step S12, a region R (see the lower part of Figure 10) that the processor 12 estimates to contain a distortion region for each unit section of the cross-sectional profile of the surface model SM on the two-dimensional plane on which the PCA features are plotted is predefined. The modified example of estimation step S12 utilizes the tendency that when features of unit sections containing distortion regions are converted into PCA features, they tend to gather in specific regions on the two-dimensional plane. In this way, by configuring estimation step S12 using PCA features, it is possible to estimate distortion regions that may be included in the surface model SM without using the learned model LM.

In this modified example, PCA is performed on the multiple feature quantities derived in the preprocessing step S11. However, the multiple feature quantities used in this modified example can be changed as appropriate. In addition, the region R that defines the feature quantities of the unit section including the distortion region can also be changed as appropriate.

The machine learning device 3 and the distortion area estimation device 1 included in the distortion area estimation system S shown in FIG. 1 were used to estimate distortion areas that may be included in the cross-sectional profile of the surface model SM. In this embodiment, the Bayesian Gaussian Mixture model, the Support Vector Machine, and the Gradient Boosting (LightGBM) were used as algorithms for the trained model LM. Hereinafter, the Bayesian Gaussian Mixture model, the Support Vector Machine, and the Gradient Boosting will be referred to as the first embodiment, the second embodiment, and the third embodiment, respectively.

In the first to third examples, the same data set was used in each of the learning phase, trial phase (verification phase), and practical phase (estimation phase). These data sets correspond to 12,214 unit sections obtained from multiple cross-sectional profiles of the surface models SM of 13 parts. In this example, of the 12,214 unit sections, in each of the first to third examples, 7,816 (64% of the total) were used in the learning phase, 1,955 (16% of the total) were used in the trial phase, and 2,443 (20% of the total) were used in the practical phase. Note that the machine learning device 3 was used in the learning phase, and the learned model LM obtained by the machine learning device 3 executing the machine learning method M3 was provided to the distortion region estimation device 1. The distortion region estimation device 1 was used in the trial phase and practical phase.

In the estimation step S12 using the trained model LM, as described above, 20 feature quantities corresponding to each unit interval were input, and the estimation result of whether or not the unit interval contained a distorted area was output. The output estimation result was set to "1" if the unit interval contained a distorted area, and "0" if the unit interval did not contain a distorted area.

In addition, the estimation results using the learned model LM were evaluated in the following three ways.
Correct answer: The area that the operator judged to be distorted was also estimated to be distorted by the trained model LM.
- Undetected: The learned model LM did not estimate an area that the operator judged to be distorted as distorted.
Overdetection: The trained model LM estimated that a part was distorted in a way that the operator did not determine.

The higher the percentage of correct answers (correct answer rate), the higher the evaluation will be. The lower the percentage of undetected answers (undetected rate), the higher the evaluation will be. It is considered acceptable for the percentage of overdetected answers (overdetected rate) to not reach 0%.

However, even an operator may have difficulty determining whether or not there is a distorted area in a part, and the determination also depends on the operator's level of skill. Therefore, if the overdetection rate is extremely low, it is likely that suspicious cases that may be considered to be distortion are overlooked, and this is not desirable as a trained model LM.

The trained model LM using the Bayesian Gaussian mixture model, the first embodiment, had 32 correct answers, 11 undetected cases, and 161 overdetected cases out of 2,443 cases in the practical phase. The trained model LM using the support vector machine, the second embodiment, had 34 correct answers, 9 undetected cases, and 10 overdetected cases. The trained model LM using gradient boosting, the third embodiment, had 38 correct answers, 5 undetected cases, and 98 overdetected cases.

The first example had the most non-detections and overdetections. There were also parts with zero correct answers. Furthermore, when workers checked the unit sections that were estimated to be overdetected, they found that there were many areas that did not appear to be distorted.

In the second example, the number of correct answers and undetected cases was not bad, but there were few overdetections. Therefore, there is a high possibility that it will not be possible to extract areas that appear to be distorted and may be overlooked by some workers.

The third example had a low number of undetected cases, and the number of overdetected cases was somewhere between those of the first and second examples, which seemed appropriate. When the workers checked the unit sections that were estimated to be overdetected, they found that they had extracted unit sections that could be seen as distorted depending on how they looked at them.

From the above results, it was found that gradient boosting is a preferable algorithm for the learned model LM used in the distortion region estimation system S.

〔summary〕
A distortion area estimation device according to a first aspect of the present invention includes a control unit that executes an estimation step of estimating a distortion area that may be included in a three-dimensional surface model and that appears visually distorted due to an optical illusion.

The above configuration makes it possible to detect visual distortions that may occur in a three-dimensional surface model and are caused by optical illusions.

In addition to the configuration of the distortion area estimation device according to the second aspect of the present invention, in addition to the configuration of the distortion area estimation device according to the first aspect described above, the control unit further executes a preprocessing step before the estimation step, and the estimation step uses a trained model constructed by supervised learning to estimate a distortion area that may be included in a line included in the surface model and that appears visually distorted due to an optical illusion, the preprocessing step includes a feature derivation step of setting a plurality of unit intervals in the line included in the surface model and deriving a feature that indicates a feature of the shape of each unit interval, the input of the trained model is the feature derived for each unit interval, and the output of the trained model is an estimation result of whether or not each unit interval includes the distortion area.

According to the above configuration, it is possible to output an estimation result of whether or not the trained model to which the feature quantities derived for each unit section are input includes the distorted region for each unit section. Therefore, by using this distorted region estimation device, it is possible to detect visual distortions that may occur in a three-dimensional surface model and are caused by optical illusions.

In addition, in the distortion region estimation device according to the third aspect of the present invention, in addition to the configuration of the distortion region estimation device according to the second aspect described above, the lines included in the surface model are a cross-sectional profile of the surface model, and the pre-processing step sets the multiple unit sections in the cross-sectional profile and derives the feature amount of each unit section.

The cross-sectional profile of the surface model is suitable as the line to use to derive the features.

In addition, in the distortion region estimation device according to the fourth aspect of the present invention, in addition to the configuration of the distortion region estimation device according to the third aspect described above, the pre-processing step further includes a cross-sectional profile generation step of generating a plurality of the cross-sectional profiles from the surface profile of the surface model.

According to the above configuration, it is possible to detect areas of distortion that may occur in a surface model using multiple cross-sectional profiles. This allows the user to grasp in two dimensions the areas in the surface model where distortion occurs.

In addition, in the distortion region estimation device according to the fifth aspect of the present invention, in addition to the configuration of the distortion region estimation device according to the fourth aspect described above, a configuration is adopted in which the multiple cross-sectional profiles generated in the cross-sectional profile generation step are multiple cross sections represented in an oblique coordinate system formed by three mutually intersecting axes, and are cross-sectional profiles in multiple cross sections parallel to a plane defined by two of the three axes.

It is preferable that adjacent cross-sectional profiles of the multiple cross-sectional profiles are parallel to each other. It is more preferable that adjacent cross-sectional profiles are parallel to each other and are equally spaced apart. By having adjacent cross-sectional profiles be equally spaced apart, it is possible to reduce the probability of overlooking a distortion region that is larger than the distance between the adjacent cross-sectional profiles.

In addition, in the distortion region estimation device according to the sixth aspect of the present invention, in addition to the configuration of the distortion region estimation device according to the fifth aspect described above, a configuration is adopted in which the three axes are the x-axis, y-axis, and z-axis, respectively, and the cross-sectional profiles are cross-sectional profiles at (1) multiple cross sections parallel to the xy plane and equally spaced, (2) multiple cross sections parallel to the yz plane and equally spaced, and (3) multiple cross sections parallel to the zx plane and equally spaced.

According to the above configuration, when the cross-sectional profile generation step generates the cross-sectional profile, the pre-processing step can be performed mechanically without considering the orientation of the surface model. Therefore, it is possible to improve the efficiency of estimating the distortion area.

In addition, in the seventh aspect of the present invention, in addition to the configuration of the distortion region estimation device according to any one of the third to sixth aspects described above, each of the cross-sectional profiles is configured by connecting N curves (N is an integer equal to or greater than 3) in series, each of the curves being configured by a single spline curve, and the feature derivation step derives the feature of each unit section, with the section being configured by the first section consisting of the i-th spline curve (i is an integer equal to or greater than 2 and equal to or less than N-1), the second section consisting of the i-1-th spline curve, and the third section consisting of the i+1-th spline curve as the unit section.

When the surface model is CAD data handled by 3D CAD, the surface profile is made up of multiple constituent surfaces. In such cases, the cross-sectional profile is constructed by connecting multiple spline curves, so the above configuration can be used advantageously.

In addition, in the distortion region estimation device according to the eighth aspect of the present invention, in addition to the configuration of the distortion region estimation device according to the seventh aspect described above, the M pseudo line segments obtained by dividing each of the first section, the second section, and the third section into M equal parts (M is an integer equal to or greater than 2), are defined as line segments L1j, L2j, and L3j (j is an integer equal to or greater than 1 and equal to or less than M), respectively, and the feature derivation step derives, as features of each unit section, at least the maximum value of the change in the angle between adjacent line segments L1j and L1j+1 in the first section, the maximum value of the change in the angle between adjacent line segments L2j and L2j+1 in the second section, and the maximum value of the change in the angle between adjacent line segments L3j and L3j+1 in the third section.

Examples of features include the features described above. By using at least these features, it is possible to accurately estimate the distortion occurring in a unit section in which a first section with a relatively small curvature is sandwiched between a second section and a third section with a relatively large curvature.

A distortion area estimation method according to a ninth aspect of the present invention includes an estimation step of estimating a distortion area that may be included in a three-dimensional surface model and that appears visually distorted due to an optical illusion.
Furthermore, a distorted area estimation method according to a tenth aspect of the present invention, in addition to the configuration of the distorted area estimation method according to the ninth aspect described above, further includes a preprocessing step executed before the estimation step, in which the estimation step uses a trained model constructed by supervised learning to estimate a distorted area that may be included in a line included in the surface model and that appears visually distorted due to an optical illusion, the preprocessing step includes a feature derivation step of setting a plurality of unit intervals in the line included in the surface model and deriving features indicating characteristics of a shape of each unit interval, an input of the trained model is the feature derived for each unit interval, and an output of the trained model is an estimation result of whether or not each unit interval includes the distorted area.

In addition, in the distortion region estimation method according to the eleventh aspect of the present invention, in addition to the configuration of the distortion region estimation method according to the tenth aspect described above, the lines included in the surface model are a cross-sectional profile of the surface model, and the pre-processing step sets the multiple unit sections in the cross-sectional profile and derives the feature amount of each unit section.

In addition, in the distortion region estimation method according to the twelfth aspect of the present invention, in addition to the configuration of the distortion region estimation method according to the eleventh aspect described above, the pre-processing step further includes a cross-sectional profile generation step of generating a plurality of the cross-sectional profiles from the surface profile of the surface model.

The distortion area estimation methods according to the ninth to twelfth aspects of the present invention each have the same effects as the distortion area estimation devices according to the first to fourth aspects of the present invention.

A machine learning device according to a thirteenth aspect of the present invention includes a control unit that executes a preprocessing step and a construction step of constructing a trained model that estimates a distortion area that may be included in a line included in a three-dimensional surface model and that appears visually distorted due to an optical illusion, by supervised learning using a training dataset, the preprocessing step including a feature derivation step of setting a plurality of unit intervals in the line included in the surface model and deriving features that indicate characteristics of the shape of each unit interval, the input of the trained model being the features derived for each unit interval, and the output of the trained model being an estimation result of whether or not each unit interval includes the distortion area.

The machine learning method according to the fourteenth aspect of the present invention includes a preprocessing step and a construction step in which a control unit constructs a trained model that estimates a distortion area that may be included in a line included in a three-dimensional surface model and that appears visually distorted due to an optical illusion, by supervised learning using a training dataset, the preprocessing step includes a feature derivation step in which a plurality of unit sections are set in the line included in the surface model, and features indicating characteristics of the shape of each unit section are derived, the input of the trained model being the features derived for each unit section, and the output of the trained model being an estimation result of whether or not each unit section includes the distortion area.

The machine learning device according to the thirteenth aspect of the present invention and the machine learning method according to the fourteenth aspect of the present invention each achieve the same effects as the distortion area estimation device according to the second aspect of the present invention.

[Additional Notes]
The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Other embodiments obtained by appropriately combining the technical means disclosed in the above-described embodiments are also included in the technical scope of the present invention.

REFERENCE SIGNS LIST 1 Distortion area estimation device 11 Memory 12 Processor M1 Distortion area estimation method S11 Pre-processing step S12 Estimation step 2 Three-dimensional CAD
3 Machine learning device 31 Storage 32 Processor 33 Memory LM Trained model DS Training data set

Claims (14)

a control unit that executes an estimation step of estimating a distortion area that may be included in a three-dimensional surface model and that appears visually distorted due to an optical illusion;
A distortion area estimation device comprising: The control unit further performs a pre-processing step before the estimation step,
The estimation step estimates a distortion area that may be included in a line included in the surface model and that appears visually distorted due to an optical illusion, using a learned model constructed by supervised learning;
the pre-processing step includes a feature amount deriving step of setting a plurality of unit sections on the line included in the surface model and deriving feature amounts indicating characteristics of a shape of each unit section;
An input of the trained model is the feature derived for each unit interval,
The output of the trained model is an estimation result of whether or not each unit section includes the distortion region.
2. The distortion area estimation device according to claim 1 . the lines included in the surface model are cross-sectional profiles of the surface model;
the pre-processing step includes setting the plurality of unit sections in the cross-sectional profile and deriving the feature amount of each unit section;
3. The distortion area estimation device according to claim 2. The pre-processing step further includes a cross-sectional profile generating step of generating a plurality of the cross-sectional profiles from a surface profile of the surface model.
4. The distortion area estimation device according to claim 3. The plurality of cross-sectional profiles generated in the cross-sectional profile generating step are cross sections represented by an oblique coordinate system formed by three mutually intersecting axes, and are cross-sectional profiles in a plurality of cross sections parallel to a plane defined by two of the three axes.
5. The distortion area estimation device according to claim 4. The three axes are defined as the x-axis, the y-axis, and the z-axis,
The cross-sectional profiles are: (1) a plurality of cross sections parallel to an xy plane and equally spaced apart; (2) a plurality of cross sections parallel to a yz plane and equally spaced apart; and (3) a plurality of cross sections parallel to a zx plane and equally spaced apart.
The distortion area estimation device according to claim 5 . Each of the cross-sectional profiles is configured by connecting N curves in series (N is an integer equal to or greater than 3),
each of the curves is constructed from a single spline curve;
the feature deriving step defines a section as the unit section, the section being composed of a first section composed of an i-th spline curve (i is an integer not less than 2 and not more than N-1), a second section composed of an i-1-th spline curve, and a third section composed of an i+1-th spline curve, and derives the feature of each unit section;
7. The distortion region estimating device according to claim 3, wherein the distortion region estimating device comprises: a first section for estimating a distortion area; M pseudo line segments obtained by equally dividing each of the first section, the second section, and the third section into M sections (M is an integer equal to or greater than 2) are defined as line segments L1j, L2j, and L3j (j is an integer equal to or greater than 1 and equal to or less than M), respectively.
the feature deriving step derives, as features for each unit section, at least a maximum value of an angle change between adjacent line segments L1j and L1j+1 in the first section, a maximum value of an angle change between adjacent line segments L2j and L2j+1 in the second section, and a maximum value of an angle change between adjacent line segments L3j and L3j+1 in the third section;
The distortion region estimating device according to claim 7 . a step of estimating a distortion area that may be included in the three-dimensional surface model and that appears visually distorted due to an optical illusion;
A method for estimating a distorted area comprising: further comprising a pre-processing step performed before said estimation step;
The estimation step estimates a distortion area that may be included in a line included in the surface model and that appears visually distorted due to an optical illusion, using a learned model constructed by supervised learning;
the pre-processing step includes a feature amount deriving step of setting a plurality of unit sections on the line included in the surface model and deriving feature amounts indicating characteristics of a shape of each unit section;
An input of the trained model is the feature derived for each unit interval,
The output of the trained model is an estimation result of whether or not each unit section includes the distortion region.
The method for estimating a distorted region according to claim 9 . the lines included in the surface model are cross-sectional profiles of the surface model;
the pre-processing step includes setting the plurality of unit sections in the cross-sectional profile and deriving the feature amount of each unit section;
The method for estimating a distorted region according to claim 10 . The pre-processing step further includes a cross-sectional profile generating step of generating a plurality of the cross-sectional profiles from a surface profile of the surface model.
The method for estimating a distorted region according to claim 11 . The present invention includes a control unit that executes a preprocessing step and a construction step of constructing a trained model that estimates a distortion area that may be included in a line included in a three-dimensional surface model and that appears visually distorted due to an optical illusion, by supervised learning using a training dataset,
the pre-processing step includes a feature amount deriving step of setting a plurality of unit sections on the line included in the surface model and deriving feature amounts indicating characteristics of a shape of each unit section;
An input of the trained model is the feature derived for each unit interval,
The output of the trained model is an estimation result of whether or not each unit section includes the distortion region.
A machine learning device characterized by: The method includes a preprocessing step, and a construction step of constructing a trained model by supervised learning using a training dataset, the trained model estimating a distortion area that may be included in a line included in a three-dimensional surface model and that appears visually distorted due to an optical illusion,
the pre-processing step includes a feature amount deriving step of setting a plurality of unit sections on the line included in the surface model and deriving feature amounts indicating characteristics of a shape of each unit section;
An input of the trained model is the feature derived for each unit interval,
The output of the trained model is an estimation result of whether or not each unit section includes the distortion region.
A machine learning method comprising:

Images