JP2024017172A - Analysis system, analysis method, and program - Google Patents

Abstract

[Challenge] Reflect customer opinions quantitatively in the design. [Solution] Using a trained generator that has learned the characteristics of the target design, a variation design having the characteristics of the target design is generated, and the features in the generated variation design are assigned to each feature vector in a multidimensional feature vector space. a variation design generation unit that generates information associated with the variation design as feature information as a vector value; an opinion quantification unit that quantifies the opinion by converting it into an opinion; an opinion reflection design generation unit that causes the learned generator to generate a design in which the opinion is reflected, based on the quantified opinion; Equipped with. [Selection diagram] Figure 1

Description

The present invention relates to an analysis system, an analysis method, and a program.

There are techniques for reflecting customer opinions in design (for example, Patent Document 1 and Patent Document 2).

Japanese Patent Application Publication No. 2000-285151 Japanese Patent Application Publication No. 2012-178130

However, it is extremely difficult to accurately verbalize the impression received from a design. Customer opinions may include unconscious impressions of a design that are not verbalized, and customer opinions may not necessarily reflect the customer's true intentions. For example, even if the design is changed in accordance with customers' opinions, some customers may not like the changed design.

In view of the above-mentioned problems, an object of the present invention is to provide an analysis system, an analysis method, and a program that can quantitatively reflect customer opinions in a design.

The analysis system according to one aspect of the present invention uses a trained generator that has learned the characteristics of the target design to generate a variation design having the characteristics of the target design, and uses the characteristics of the generated variation design in a multidimensional manner. a variation design generation unit that generates information associated with the variation design as feature information as a vector value of each feature vector in a feature vector space; an opinion quantification unit that quantifies the opinion by converting it into a vector value of each feature vector in , and causes the learned generator to generate a design in which the opinion is reflected based on the quantified opinion. and an opinion reflection design generation section.

An analysis method according to one aspect of the present invention is an analysis method performed by a computer, in which a variation design generation unit uses a trained generator that has learned the characteristics of the target design to create a variation design having the characteristics of the target design. and generates information in which features in the generated variation design are associated with the variation design as vector values of each feature vector in a multidimensional feature vector space as feature information, and the opinion quantification unit The opinion is quantified by converting the opinion to a vector value of each feature vector in the multidimensional feature vector space using the feature information, and the opinion reflection design generation unit is configured to , causing the learned generator to generate a design that reflects the opinion.

A program according to one aspect of the present invention causes a computer to generate a variation design having the characteristics of the target design using a trained generator that has learned the characteristics of the target design, and sets the characteristics of the generated variation design to: Information associated with the variation design is generated as feature information as a vector value of each feature vector in the multidimensional feature vector space, and opinions regarding the variation design are expressed using the feature information for each feature in the multidimensional feature vector space. The program quantifies the opinion by converting it into a vector value, and causes the learned generator to generate a design in which the opinion is reflected based on the quantified opinion.

According to the present invention, customer opinions can be quantitatively reflected in the design.

1 is a block diagram showing a configuration example of an analysis system 1 according to an embodiment. FIG. 3 is a diagram showing an example of a variation design according to the embodiment. FIG. 3 is a diagram for explaining characteristic information 35 according to the embodiment. FIG. 3 is a diagram for explaining characteristic information 35 according to the embodiment. FIG. 3 is a diagram for explaining processing performed by the opinion quantification unit 32 according to the embodiment. FIG. 3 is a diagram for explaining processing performed by the opinion quantification unit 32 according to the embodiment. FIG. 3 is a diagram for explaining processing performed by the opinion quantification unit 32 according to the embodiment. FIG. 3 is a diagram for explaining processing performed by the opinion quantification unit 32 according to the embodiment. It is a sequence diagram showing the flow of processing performed by the analysis system 1 according to the embodiment.

Embodiments of the present invention will be described below with reference to the drawings.

<Configuration of analysis system 1>
FIG. 1 is a diagram illustrating an overview of an analysis system 1 according to an embodiment. The analysis system 1 includes, for example, a learning device 10, a generation device 20, and an analysis device 30.

In the following, an example will be described in which the analysis system 1 is constituted by three devices: a learning device 10, a generation device 20, and an analysis device 30, but the present invention is not limited to this. The analysis system 1 may be realized by one device having the functions of the learning device 10, the generation device 20, and the analysis device 30, or the functions of two of the learning device 10, the generation device 20, and the analysis device 30. may be realized by one device.

In the analysis system 1, a variation design having characteristics of the target design is generated using a machine learning method (see variation design VD1Aa etc. in FIG. 5). The target design is the design on which customer opinions are analyzed. In the following description, a case where the target design is a three-dimensional pattern that is a three-dimensional shape generated on the surface of building materials such as panels used for building exteriors, floors, ceilings, etc. will be exemplified and explained. However, the target design is not limited to this, and any three-dimensional shape can be used as the target design.

In the example of FIG. 1, in order to make it easier to understand the difference between the types of designs, an example is shown in which designs in which the numbers "0, 1, 2, . . . , 9" are drawn are used as target designs. An example is shown in which handwritten-like numbers "0, 1, 2, ..., 9" are generated as a variation design of the numbers "0, 1, 2, ..., 9" as the target design. Then, as a design that reflected the customer's opinions, a variation design in which the target design for the number "9" was drawn in a handwritten style was changed to another variation design in which the target design for the number "4" was drawn in a handwritten style. An example is shown.

In the analysis system 1, when generating a variation design, characteristic information of the variation design is generated. Feature information is information in which attributes and features of a design are expressed numerically. For example, the vector value of each feature vector in a multidimensional feature vector space called a latent vector or the like can be used as the feature information. By generating feature information, a design can be quantified as a latent vector rather than verbalized, and differences in each design can be expressed quantitatively as differences in latent vectors.

In the analysis system 1, the customer's opinion regarding the variation design is quantified using the feature information. For example, the customer's preferences are converted into numerical values as vector values of each feature vector in a multidimensional feature vector space using feature information corresponding to a design in which the customer had a good impression. This quantifies customer opinions, including preferences and impressions that are difficult to accurately verbalize. By doing this, you can use quantified customer opinions to accurately understand the design that customers would prefer, and change the design to reflect customer opinions correctly. This makes it possible to generate new designs.

Learning device 10 is a computer. As the learning device 10, for example, a server device, a PC (Personal Computer), a cloud, a tablet terminal, etc. can be applied. The learning device 10 uses a machine learning method to generate a trained model that has learned the characteristics of a three-dimensional shape that is a target design (hereinafter referred to as a target shape). Further, the learning device 10 generates material information indicating the characteristics of the material included in the target shape.

The generation device 20 is a computer. As the generation device 20, for example, a server device, a PC, a cloud, a tablet terminal, etc. can be applied. Information on the learned model generated by the learning device 10 and material information are input to the generation device 20 . The generation device 20 generates a variation design using the learned model, and synthesizes material information with the generated variation design to generate a variation design that has a textured feel. Further, when generating a variation design using the learned model, the generation device 20 generates feature information in the variation design, and outputs the generated feature information to the analysis device 30.

Analyzer 30 is a computer. As the analysis device 30, for example, a server device, a PC, a cloud, a tablet terminal, etc. can be applied. The feature information generated by the generation device 20 is input to the analysis device 30 . Further, the analysis device 30 receives input of customer opinions regarding the variation designs generated by the generation device 20. The customer's opinion here is, for example, ``This design looks good,'' or ``That design is the best (I like),'' etc., and details such as what parts of the design are good and how. assumes unspoken opinions.

The analysis device 30 analyzes customer opinions. For example, the analysis device 30 identifies a design that seems to suit the customer's taste based on the customer's opinion. The analysis device 30 uses the identified feature information of the design to extract the vector value of each feature vector in the multidimensional feature vector space of a design that is likely to match the customer's taste.

For example, the analysis device 30 generates a design that accurately reflects the customer's opinions by changing a design requested by the customer to include design features that are likely to match the customer's preferences.

In this case, for example, the analysis device 30 outputs the quantified customer's opinion to the generation device 20 and requests that the customer's opinion be correctly reflected in the generation of a design. That is, the analysis device 30 outputs to the generation device 20 the quantified customer opinion, that is, the vector value of each feature vector in a design that seems to match the customer's taste. Thereby, the generation device 20 generates a design that has features corresponding to the vector values of each feature vector notified from the analysis device 30, that is, a design that seems to match the customer's taste.

<Configuration of learning device 10>
As shown in FIG. 1, the learning device 10 includes, for example, an acquisition section 11, a material processing section 12, a variation design processing section 13, and a z information color information correspondence table 15. The acquisition unit 11 acquires three-dimensional information of a three-dimensional pattern in the target shape. The three-dimensional information of the three-dimensional pattern is, for example, three-dimensional information expressed by a three-axis xyz coordinate system, and height information in the z-axis direction is associated with the position coordinates of the three-dimensional pattern on the xy plane. It is information. The acquisition unit 11 outputs the acquired three-dimensional information to the material processing unit 12 and variation design processing unit 13.

The material processing unit 12 generates material information. The material information is information indicating the characteristics of the material in the target shape. The material processing unit 12 acquires three-dimensional information from the acquisition unit 11 and extracts high frequency components from the acquired three-dimensional information. For example, the material processing unit 12 extracts a target shape whose height change characteristic along a predetermined direction is equal to or higher than a specific frequency as a high frequency component.

The material processing unit 12 generates material information using the extracted high frequency components. For example, the material processing unit 12 determines a parameter indicating the feel of the material. Parameters indicating the texture include, for example, height, length, and pitch. The height is information indicating the height in the z-axis direction. Length is the length that the height continues. The pitch is the frequency with which regions whose height continues for the length appear.

For example, the material processing unit 12 generates a plurality of combinations of height, length, pitch, etc. as candidates for parameters indicating the feel of the material, and determines a combination that gives the human eye a feeling of the material. You can also do this. In this case, for example, a combination that gives a feel of the material is determined by making a prototype shape using each combination, asking multiple users to look at the prototype shape, and asking whether or not they feel the feel of the material. . Alternatively, the material processing unit 12 may determine parameters indicating the texture of the material using a statistical method. In this case, the material processing unit 12 uses, for example, the average height, length, pitch, etc. in a frequency band that includes many components among high frequency components as parameters indicating the texture of the material.

The material processing unit 12 generates, as material information, a three-dimensional shape formed using parameters indicating the texture of the material, such as the determined height, length, and pitch.

The material processing unit 12 converts three-dimensional material information into two-dimensional image information using the z information color information correspondence table 15. This is because the generation device 20 uses image synthesis technology in the process of synthesizing the material information with the variation design.

The z information/color information correspondence table 15 is a table in which height information (z information) is associated with color information. For example, the z information color information correspondence table 15 has each item of z, R, G, and B. z indicates height information in the three-dimensional shape. R, G, and B each indicate an RGB value as color information in the two-dimensional image. R, G, and B correspond to Red, Green, and Blue, respectively.

Note that although the case where RGB is used as an example of color information is illustrated here, the color information may be information indicating at least a color, and is not limited to RGB values. For example, information indicating colors in the CMKY system may be used as the color information.

The material processing unit 12 extracts height information from three-dimensional material information. Based on the extracted height information, the material processing unit 12 refers to the z information color information correspondence table 15 and acquires color information (RGB values) corresponding to the height information (z coordinate value). The material processing unit 12 converts a z coordinate value of the xyz coordinates shown as three-dimensional information into color information (RGB values) corresponding to the z coordinate value. Thereby, the material processing unit 12 generates image information of an image in which the features of the material indicated as three-dimensional material information are converted into image features in a two-dimensional image.

The material processing unit 12 transmits the generated image information to the generation device 20 as image information of material image information corresponding to three-dimensional material information.

The variation design processing section 13 includes, for example, an image processing application section 130 (not shown), a learning section 131 (not shown), and an evaluation section 132 (not shown).

In this embodiment, an existing learning model developed for images (hereinafter referred to as image AI) is used as a learning model for learning the pattern characteristics of the target shape. By using image AI, there is no need to newly develop a learning model for learning three-dimensional features, and it is possible to suppress an increase in development costs.

In this embodiment, in order to cause the image AI to learn the pattern features, an image is generated in which the pattern features in the target shape are converted into image features in the two-dimensional image.

Specifically, the image processing application unit 130 acquires three-dimensional information from the acquisition unit 11 and extracts low frequency components from the acquired three-dimensional information. For example, the image processing application unit 130 extracts a target shape whose height change characteristic along a predetermined direction is less than a specific frequency as a low frequency component.

The image processing application unit 130 generates a target image by converting the three-dimensional information consisting of the extracted low frequency components into two-dimensional image information using the z information color information correspondence table 15. The method by which the image processing application unit 130 converts three-dimensional information into image information is the same as the method by which the material processing unit 12 converts three-dimensional information into image information, and therefore the description thereof will be omitted.

The learning unit 131 causes the learning model (image AI) to learn the features that the target image has. Thereby, the learning unit 131 generates a trained image generator (trained model) that generates a variation image based on the characteristics of the target image.

The learning unit 131 generates training data for learning by the learning model. The learning unit 131 generates a large amount of training data by performing image processing on the target image. The image processing here is, for example, any one of processing such as enlargement, reduction, deformation, cutting out, noise removal by filter operation, or color adjustment, or a combination thereof. By learning a learning model using a large amount of training data, it is possible to improve the accuracy and speed at which images are generated by a trained image generator that is generated through learning.

The learning unit 131 uses, for example, a generative adversarial network called GAN (Generative Adversarial Network) as a method for causing the learning model to learn. A GAN includes, for example, a generator that generates an image and a discriminator that identifies the image. In GAN, a network is constructed in such a way that a generator generates a fake image, and a classifier distinguishes between the fake image and the real image. By constructing such a network, the classifier learns to improve the accuracy of classification in an attempt to see through the fake images generated by the generator. On the other hand, the generator learns to improve generation accuracy so that it can generate fake images that cannot be detected by the classifier. Therefore, by the interaction between the generator and the classifier, the fake image generated by the generator can be gradually trained to resemble a real image.

By using GAN to train a learning model, there is no need to label training data. Therefore, it becomes possible to have the learning model learn the features with less effort.

For example, when multiple types of three-dimensional information are obtained for a design expansion target, the learning unit 131 uses a trained image generator (learning Build a GAN that generates an image generator). For example, the learning department 131 uses information published in Reference 1 ([searched on June 14, 2020], Internet <URL: https://arxiv.org/pdf/1812.04948.pdf>.) A trained image generator (learned image generator).

Alternatively, when only a single piece of three-dimensional information is obtained for the design expansion target, the learning unit 131 uses a target image corresponding to the three-dimensional information to generate a trained image generator (trained image generator). Build a GAN that generates For example, the learning department 131 uses information published in Reference 2 ([searched on June 14, 2020], Internet <URL: https://arxiv.org/pdf/1905.01164.pdf>.) A trained image generator is generated by learning the features of the target image using a GAN that is composed of an image generator and an image classifier that were originally implemented.

The learning unit 131 outputs information indicating the generated trained image generator to the evaluation unit 132.

The evaluation unit 132 evaluates the learned image generator generated by the learning unit 131. For example, the evaluation unit 132 prepares in advance an image that is not used in the training data as an image for evaluation. The evaluation unit 132 causes the trained image generator generated by the learning unit 131 to generate an image. The evaluation unit 132 derives the accuracy of the image generated by the trained image generator. For example, the evaluation unit 132 derives the degree of similarity between the image generated by the trained image generator and the target image as the accuracy of the image. The evaluation unit 132 also derives the speed at which the trained image generator generates images.

The evaluation unit 132 comprehensively evaluates the trained image generator based on the accuracy of the image generated by the trained image generator and the speed at which the trained image generator generates images. For example, the evaluation unit 132 evaluates that the trained image generator is appropriate when the trained image generator generates an image with acceptable accuracy at an acceptable speed. On the other hand, if the trained image generator fails to generate an image with acceptable accuracy and fails to generate an image at an acceptable speed, the evaluation unit 132 determines that the trained image generator is not appropriate. evaluate.

If the evaluation unit 132 evaluates that the learned image generator is appropriate, it transmits information about the learned image generator to the generation device 20. On the other hand, if the evaluation unit 132 evaluates that the trained image generator is not appropriate, it causes the learning unit 131 to perform learning again.

The evaluation unit 132 repeatedly performs learning while changing the training data until the trained image generator is evaluated as appropriate by the evaluation unit 132. The learning unit 131 changes the training data by, for example, enlarging, reducing, deforming, cutting out, removing noise by filter operation, adjusting color, etc., or by a combination thereof, and generates the changed training data. Execute learning again using . Alternatively, the learning unit 131 changes the conditions for generating the target image by changing the conversion degree for converting height information into color information, and performs learning again using training data generated from the changed target image. Execute. The learning unit 131 may change the degree of conversion as well as the training data.

The z information color information correspondence table 15 is stored in a storage unit (not shown) included in the learning device 10. The storage unit included in the learning device 10 includes, for example, a HDD (Hard Disk Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), a RAM (Random Access read/write Memory), a ROM (Read Only Memory), etc. media, or a combination of these. The storage unit included in the learning device 10 stores programs for executing various processes of the learning device 10 and temporary data used when performing the various processes.

<Configuration of generation device 20>
The generation device 20 includes, for example, an acquisition section 21, an image generation section 22, a three-dimensional pattern generation section 23, and a color information z information correspondence table 24. The acquisition unit 21 acquires information indicating the trained image generator generated by the learning device 10 and image information of the material image. The acquisition unit 21 outputs the acquired information to the image generation unit 22.

The image generation section 22 includes, for example, a material image generation section 220 (not shown), a variation image generation section 221 (not shown), and an image composition section 222 (not shown).

The material image generation unit 220 generates a material image. The material image generation unit 220 generates a material image using the image information of the material image notified from the learning device 10. The material image generation section 220 outputs image information of the generated material image to the image composition section 222.

The variation image generation unit 221 generates variation images. The variation image is an image that is generated by a trained image generator and has features of low frequency components. The variation image generation unit 221 acquires information indicating the learned image generator from the acquisition unit 21, and generates the learned image generator based on the acquired information. The variation image generation unit 221 generates a variation image by inputting arbitrary image generation parameters to the generated trained image generator. The image generation unit 22 outputs image information of the generated variation image to the image synthesis unit 222.

The image composition unit 222 generates a composite image by combining the variation image with the material image. Any method using existing image processing technology may be adopted as a method for the acquisition unit 11 to combine the two images. The image composition section 222 outputs image information of the generated composite image to the three-dimensional pattern generation section 23.

The three-dimensional pattern generation unit 23 converts the image features in the composite image into three-dimensional features based on the image information of the composite image. Thereby, the three-dimensional pattern generation unit 23 generates three-dimensional information of a three-dimensional pattern (variation shape) having three-dimensional features corresponding to the image features in the composite image.

The three-dimensional pattern generation unit 23 uses the color information z information correspondence table 24 to convert the image features in the composite image into three-dimensional features in the variation shape.

The color information z information correspondence table 24 is a table in which color information is associated with height information (z information). For example, the color information z information correspondence table 24 has each item of R, G, B, and z. Similar to the R, G, and B items in the z information color information correspondence table 15, R, G, and B indicate RGB values as color information in the two-dimensional image. Similarly to the item z in the z information/color information correspondence table 15, height information in the three-dimensional shape is indicated in z.

The three-dimensional pattern generation unit 23 extracts color information from the image information acquired from the image synthesis unit 222, for example. For example, the three-dimensional pattern generation unit 23 extracts the pixel value (for example, RGB value) of each pixel in the image information as color information. The three-dimensional pattern generation unit 23 refers to the color information z information correspondence table 24 based on the extracted color information, and obtains height information (z coordinate value) corresponding to the color information (RGB values). The three-dimensional pattern generation unit 23 converts the color information (RGB values) of each pixel in the composite image into height information (z coordinate value) corresponding to the color information. Thereby, the three-dimensional pattern generation unit 23 converts the image features in the composite image into three-dimensional features in the variation shape. That is, the three-dimensional pattern generation unit 23 generates three-dimensional information of a variation shape that has three-dimensional features in the target shape and has a texture.

The three-dimensional pattern generation unit 23 outputs the generated three-dimensional information. For example, the three-dimensional pattern generation unit 23 displays the three-dimensional shape of the variation shape by outputting three-dimensional information to the display. Alternatively, the three-dimensional pattern generation unit 23 generates a three-dimensional shape of a variation shape by outputting three-dimensional information to a 3D printer.

The color information z information correspondence table 24 is stored in a storage unit (not shown) included in the generation device 20. The storage unit included in the generation device 20 is configured by, for example, a storage medium such as an HDD, flash memory, EEPROM, RAM, or ROM, or a combination thereof. The storage unit included in the generation device 20 stores programs for executing various processes of the generation device 20 and temporary data used when performing the various processes.

FIG. 2 is a diagram showing an example of a variation design. As shown in Figure 2, in the variation design, for each of the numbers "0, 1, 2, ..., 9" in the target design, handwritten numbers "0, 1, 2, ..., 9" are added. generated.

Here, in the embodiment described above, the degree to which material information is added in the variation shape may be adjusted.

When generating a composite image, the material image generation unit 220 generates a plurality of composite images in which the ratio of combining material images is changed. For example, a composite image A is a combination of a variation image and a material image at a ratio of 10%, a composite image B is a combination of a variation image and a material image at a ratio of 15%, and a composite image is a combination of a variation image and a material image at a ratio of 15%. A composite image D is generated by combining a material image at a ratio of 25% with a composite image C and a variation image, which are composited at a ratio of 20%.

The three-dimensional pattern generation unit 23 generates variation shapes A to D corresponding to the composite images A to D, respectively. Digital mockups A to D are generated based on the three-dimensional information of each of the variation shapes A to D. Display digital mockups A to D on a display and select a digital mockup that has a textured feel. Furthermore, if necessary, a mock-up of the actual item may be generated using a 3D printer or the like, and the appearance may be observed by placing it side by side with the original, or the appearance of reflected light when exposed to sunlight may be compared. It is desirable to select an amount that gives a good overall impression (the degree to which material information is added) without damaging the image of the original.

Further, the three-dimensional pattern generation unit 23 may adjust the correspondence between height information and color information in the color information z information correspondence table 24 according to the type of shape or the like.

For example, when converting RGB values in a composite image into height information, the maximum height of the actual mockup W is 7 [mm], the actual mockup X is 10 [mm], and the maximum height is 12 [mm]. ] and an actual mockup Z with a thickness of 15 [mm] are generated.

In this way, multiple actual mockups W to Z are generated with different degrees of conversion of RGB values in the composite image to height information, and the appearance can be observed by placing them side by side with the original, or by checking the reflected light when exposed to sunlight. We compared the appearance of It is desirable to select an amount (degree of conversion of RGB values to height information) that gives a good impression as a whole without damaging the image of the original.

Furthermore, when generating a variation image, the variation image generation unit 221 of the generation device 20 generates feature information of the generated variation image. For example, the variation image generation unit 221 obtains parameters used by the trained image generator to generate variation images, that is, vector values set for each feature vector in the multidimensional feature vector space. The generation device 20 generates feature information by associating the vector value of each acquired feature vector with a variation image. Furthermore, the image generation unit 22 of the generation device 20 updates the feature information by associating the material information synthesized with the variation image with the synthesized image. Then, the three-dimensional pattern generation unit 23 of the generation device 20 updates the feature information by associating the feature information associated with the composite image with the variation shape. The generation device 20 outputs the feature information to the analysis device 30. The analysis device 30 stores the feature information acquired from the generation device 20 as feature information 35.

FIG. 3 (FIGS. 3A and 3B) is a diagram showing an example of the feature information 35. Figure 3 shows a schematic diagram in which variation designs are plotted in a multidimensional space formed by feature vectors according to their feature information, and the plotted multidimensional space is projected onto a two-dimensional area to visualize it. ing.

FIG. 3A shows the distribution in a multidimensional space of variation designs divided into patterns for each type of design. FIG. 3B shows the distribution in the multidimensional space of each variation design indicated by numbers according to the type of design.

As shown in the example of FIG. 3, when variation designs are plotted as feature information 35 in a multidimensional space formed by feature vectors, it can be seen that each design type has a somewhat coherent distribution. Depending on the variation design, designs of one type and designs of another type may be located close to each other in multidimensional space, and even designs of different types may have similar characteristics. I understand.

<Configuration of analyzer 30>
Returning to FIG. 1, the analysis device 30 includes, for example, an acquisition section 31, an opinion quantification section 32, an opinion reflection design generation section 33, an opinion reflection design output section 34, and characteristic information 35.

The acquisition unit 31 acquires the feature information generated by the generation device 20. The acquisition unit 21 stores the acquired feature information as feature information 35. The acquisition unit 31 also acquires customer opinions. The acquisition unit 31 acquires, for example, the customer's impression of each variation design, for example, the result of selecting “good” or “bad” as the customer's opinion. The acquisition unit 31 outputs the acquired customer opinions to the opinion quantification unit 32 .

The opinion quantification unit 32 quantifies customer opinions. The opinion quantification unit 32 acquires customer opinions from the acquisition unit 31, and identifies variation designs for which "good" has been selected based on the acquired customer opinions. The opinion quantification unit 32 refers to the feature information 35 based on the specified variation design and extracts feature information associated with the variation design.

The opinion quantification unit 32 uses the vector value of each feature vector in the extracted feature information as a quantified customer opinion. The opinion quantification unit 32 outputs the quantified customer opinion, that is, the vector value of each feature vector preferred by the customer, to the opinion reflection design generation unit 33.

FIG. 4 (FIG. 4A to FIG. 4C) is a diagram for explaining the processing performed by the opinion quantification unit 32. FIG. 4 shows a diagram in which variation designs are sorted in order of similar features according to their feature information.

As shown in FIG. 4A, when sorting in order of similar features, the numbers "0, 1, 2, . . . , 9" will not be in the order. For example, when viewed in the vertical direction, the design of the number "7" approaches the design of the number "9". It can also be seen that the design of the number "9" approaches the design of the number "1". Furthermore, when viewed in the horizontal direction, the design of the number "7" approaches the design of the number "9". Also, the design of the number "9" approaches the design of the number "4". Furthermore, it can be seen that the design of the number "4" approaches the design of the number "6".

FIG. 4B shows a range E1 of features that the customer prefers, which is specified by the opinion quantification unit 32. For example, when there are multiple variation designs for which "good" is selected by the customer, the opinion quantification unit 32 determines whether "good" is selected by the customer in the multidimensional feature vector space based on the feature information corresponding to each variation design. plots the position of each of the selected variation designs. The opinion quantification unit 32 sets the area including the plotted position as the range E1 of features preferred by the customer.

In the example shown in this figure, the opinion quantification unit 32 has identified a design area with the number "9" that is closer to the number "4" than the number "1" or the number "7" as a design preferred by the customer. I understand. In this case, for example, the opinion quantification unit 32 can identify a design that is close to the number "9" but closer to the number "4" as a design that the customer prefers.

FIG. 4C shows a direction from the characteristic range E2 to the range E3 as the direction in which the design is changed to the one preferred by the customer. For example, as shown in FIG. 4B, when the customer prefers a design that is close to the number "9" but closer to the number "4", the opinion quantification unit 32 determines that the design is close to the number "9" and is closer to the number "4". ” The value placed on the design is a quantified customer opinion. In this case, for example, the opinion quantification unit 32 calculates the vector value of each feature vector in the design with the number "9" shown in the range E2 and the vector value of each feature vector in the design with the number "4" shown in the range E3. Let the difference vector indicating the difference be the quantified customer opinion. In this way, the opinion quantification unit 32 may use the difference vector as the quantified opinion of the customer, or may use the vector value itself as the quantified opinion of the customer.

Further, the opinion quantification unit 32 uses the material information in the extracted feature information as a quantified customer opinion. The opinion quantification unit 32 outputs the quantified customer opinion, that is, the material information of the customer's preferred material, to the opinion reflection design generation unit 33.

FIG. 5 is a diagram for explaining the processing performed by the opinion quantification unit 32. In FIG. 5, combinations of unevenness and material texture are shown in a matrix for each of patterns 1 to 5 of the variation design. The unevenness here can be selected from levels A to E depending on the degree of conversion of RGB values in the color information z information correspondence table 24 to height information. Further, the texture here corresponds to material information, and each level from level a to level e can be selected depending on the type of material information.

For example, when there are multiple variation designs for which "good" has been selected by the customer, the opinion quantification unit 32 determines that the texture and texture that are common to many of the variation designs are the texture and texture that the customer prefers. Identify. In this way, the opinion quantification unit 32 may use both the degree of conversion of the RGB values indicating the unevenness into height information and the material information as the quantified customer opinion, or only the unevenness or the material information. Only the information may be quantified customer opinions.

By combining the unevenness and material texture that the customer has identified in this manner with the variation design that the customer prefers, it is possible to generate a design that has the unevenness and material texture that the customer prefers.
FIG. 5 shows an example in which a variation design VD1Aa is generated by combining the design corresponding to pattern 1 with the unevenness corresponding to level A and the material texture corresponding to level a.
An example is shown in which a variation design VD1Ab is generated by combining a design corresponding to pattern 1 with a texture corresponding to level A and a texture corresponding to level b.
An example is shown in which variation design VD2Ba is generated by combining the design corresponding to pattern 2 with the unevenness corresponding to level B and the material texture corresponding to level a.
An example is shown in which variation design VD3Cd is generated by combining the design corresponding to pattern 3 with the unevenness corresponding to level C and the material texture corresponding to level d.
An example is shown in which variation design VD4Ec is generated by combining the design corresponding to pattern 4 with the unevenness corresponding to level E and the material texture corresponding to level c.

Returning to FIG. 1, the opinion-reflecting design generation unit 33 generates a design that reflects the customer's opinions. For example, the opinion reflection design generation unit 33 outputs the customer's opinion quantified by the opinion quantification unit 32 to the generation device 20, and causes the generation device 20 to generate a design preferred by the customer.

The generation device 20 acquires quantified customer opinions from the opinion reflection design generation section 33. For example, the generation device 20 converts parameters set when generating a variation image corresponding to a design originally proposed to a customer into vector values of each feature vector preferred by the customer based on the customer's quantified opinion. Change the parameters to get closer. Then, the generation device 20 causes the learned image generator to generate a variation image using the changed parameters. Further, the generation device 20 generates a material image corresponding to material information preferred by the customer. The generation device 20 generates a composite image by combining the thus generated variation image preferred by the customer with a material image preferred by the customer. The generation device 20 generates variations preferred by the customer by converting the composite image into a three-dimensional shape. The generation device 20 outputs the three-dimensional information of the generated variation design to the analysis device 30 as a design reflecting the customer's opinion.

The opinion reflection design output unit 34 outputs the design generated by the generation device 20 and reflecting the customer's opinion. For example, the opinion-reflecting design output unit 34 displays a three-dimensional shape of a design that reflects the customer's opinions by outputting three-dimensional information to a display. Alternatively, the opinion-reflecting design output unit 34 outputs three-dimensional information to a 3D printer to generate a three-dimensional shape of a design that reflects the customer's opinion.

The characteristic information 35 is information indicating the characteristics of the variation design. The three-dimensional pattern generation section 23 is stored in a storage section (not shown) included in the analysis device 30. The storage unit included in the analysis device 30 is configured by, for example, a storage medium such as an HDD, flash memory, EEPROM, RAM, or ROM, or a combination thereof. The storage unit included in the analyzer 30 stores programs for executing various processes of the analyzer 30 and temporary data used when performing the various processes.

<Flow of processing performed by analysis system 1>
FIG. 6 is a sequence diagram showing the flow of processing performed by the analysis system 1 according to the embodiment.

First, the learning device 10 acquires three-dimensional information of the target design (step S10). The learning device 10 generates material information (step S11). For example, the learning device 10 extracts high frequency components from the target shape and generates material information using the extracted high frequency components. The learning device 10 generates material information by, for example, statistically processing the characteristics of high frequency components and setting material parameters such as height, length, pitch, etc. that indicate the feel of the material. Furthermore, the learning device 10 generates image information of a material image on which the material information is reflected). The learning device 10 generates image information of the material image by converting height information in three-dimensional information indicating material information into color information in the material image.

Further, the learning device 10 generates a trained image generator (trained model) (step S12). For example, the learning device 10 extracts a low frequency component from the target shape and generates image information of the target image using the extracted low frequency component. The learning device 10 generates image information of the target image by converting height information in three-dimensional information consisting of low frequency components into color information in the target image. The learning device 10 uses machine learning using GAN to generate an image generator that has learned image features in a target image. The learning device 10 evaluates the performance of the generated image generator. The learning device 10 determines whether the performance of the image generator is appropriate.

If the performance of the image generator is not appropriate, the learning device 10 generates the image generator again by changing the conditions for generating the target image and learning the characteristics of the target image generated under the changed conditions. . The learning device 10 changes the training data by, for example, enlarging, reducing, deforming, cutting out, removing noise by filter operation, adjusting colors, etc., or by a combination thereof, and generates the changed training data. The image generator is generated again by executing learning using .

If the performance of the image generator is appropriate, the learning device 10 sets the image generator as a learned image generator, and generates information indicating the learned image generator and image information of the material image generated in step S11. to the device 20.

The generation device 20 receives information indicating a trained image generator (trained model) and image information of the material image from the learning device 10 (step S13). The generation device 20 generates a variation design (step S14). First, the generation device 20 generates a material image using the image information of the material image received from the learning device 10. Next, the generation device 20 generates a learned image generator using the information indicating the learned image generator received from the learning device 10. Next, the generation device 20 generates a variation image using the trained image generator.

The generation device 20 generates a composite image by combining the variation image with the material image. The generation device 20 generates a variation shape by converting color information in the composite image into height information in the variation shape.

The generation device 20 outputs three-dimensional information of the generated variation shape (step S15). Thereby, the variation design generated by the generation device 20 is presented to the customer. Further, the generation device 20 outputs characteristic information of each of the generated variation designs to the analysis device 30. The analysis device 30 acquires the feature information notified from the generation device 20 (step S16).

On the other hand, the analysis device 30 acquires the customer's opinion regarding the variation design (step S17). The analysis device 30 quantifies the obtained customer opinions (step S18). For example, the analysis device 30 plots the position in the multidimensional feature vector space of each variation design for which "good" is selected by the customer, and sets the region including the plotted position as the range of features preferred by the customer. The analysis device 30 quantifies the vector value of each representative feature vector within the range of features preferred by the customer, or the difference vector for changing an existing design to a design within the range of features preferred by the customer. Based on customer opinions.

In addition to the above-mentioned vector values or difference vectors (hereinafter referred to as vector values, etc.), the analysis device 30 also uses quantified customer opinions on the unevenness and texture that the customer prefers, or in addition to the vector values and the like. It may be done as follows. For example, the analysis device 30 determines the unevenness and texture that are common to many of the variation designs for which "good" has been selected by the customer as the unevenness and texture that the customer prefers. The analysis device 30 converts the customer's preferred unevenness and material texture into a quantified customer opinion.

The analysis device 30 generates a design that reflects customer opinions. For example, the analysis device 30 requests the generation device 20 to generate a design that reflects the customer's opinion by outputting the quantified customer's opinion to the generation device 20 (step S19).

The generation device 20 generates a design based on the quantified customer opinion notified from the analysis device 30 (step S20). The generation device 20 generates a design that reflects the customer's opinions by, for example, generating a design having vector values of representative feature vectors within a range of features preferred by the customer. Alternatively, the generation device 20 generates a design that reflects the customer's opinion by changing the existing design according to a difference vector for changing the range of features preferred by the customer. Further, the generating device 20 may generate a design that reflects the customer's opinion by combining the generated design with the texture and texture that the customer prefers. The generation device 20 outputs the generated design to the analysis device 30 as a design that reflects the customer's opinion.

The analysis device 30 acquires the design that reflects the customer's opinion, notified from the generation device 20, and outputs the acquired design (step S21). As a result, a design that reflects the customer's opinions is presented to the customer.

In addition, although the above-mentioned embodiment illustrated and demonstrated the case where a three-dimensional shape (variation shape) is produced|generated via an image (a material image and a variation image), it is not limited to this. It is not necessary to use a vector value format for each feature vector as described above, as long as the design can be generated in such a way that the details of the features can be understood. For example, a functional model may be used in which the characteristics of variation shapes generated by any method are quantified. In this case, for example, it is conceivable to use a functional model that digitizes, based on input three-dimensional information, the characteristics of a three-dimensional shape formed by the three-dimensional information. In this way, the design features may be quantified.

Further, in the above-described embodiment, a case where a three-dimensional shape (variation shape) is generated as a variation design has been described as an example, but the present invention is not limited to this. A two-dimensional image may be used as the variation design. The analysis system 1 of this embodiment can be applied even when a two-dimensional image is used as a variation design.

As described above, the analysis system 1 of the embodiment includes the generation device 20 (variation design generation section), the opinion quantification section 32, and the opinion reflection design generation section 33. The generation device 20 generates variation designs. A variation design is a design that has characteristics of the target design. The generation device 20 generates a variation design using a trained model (trained generator) that has learned the characteristics of the target design. The generation device 20 also generates feature information. Feature information is information indicating features in a variation design. Feature information is information in which features in a variation design are associated with the variation design as vector values of each feature vector in a multidimensional feature vector space. The opinion quantification unit 32 quantifies customer opinions (opinions regarding variation designs). The opinion quantification unit 32 quantifies the customer's opinion by converting the customer's opinion into a vector value of each feature vector in the multidimensional feature vector space using the feature information associated with the variation design. . The opinion-reflecting design generation unit 33 generates a design that reflects the customer's opinions. The opinion-reflecting design generation unit 33 causes the trained model to generate a design in which the customer's opinion is reflected, based on the quantified opinion.

Thereby, in the analysis system 1 of the embodiment, the breakdown of design features preferred by the customer can be grasped as the vector value of each feature vector in the multidimensional feature vector space. In other words, customers' opinions can be quantified by converting the designs that customers prefer into vector values for each feature vector, and based on the quantified opinions, new designs can be generated or existing designs can be changed. You can do it. Therefore, customer opinions can be quantitatively reflected in the design.

Moreover, the analysis system 1 of the embodiment further includes a material processing section 12. The material processing unit 12 generates material information indicating the material feel of the target design. The generation device 20 generates, as a variation design, a design in which material information is combined with the design generated using the learned model. Furthermore, the generation device 20 generates, as feature information, information that associates the material information used for synthesis with the variation design. The opinion quantification unit 32 quantifies the customer's opinion by converting the customer's opinion into material information that is synthesized into a variation design using feature information. The opinion-reflecting design generation unit 33 generates a design in which material information reflecting customer opinions is synthesized. The opinion reflection design generation unit 33 generates a design in which the material information in the quantified customer opinion is synthesized with the design generated by the learned model based on the quantified customer opinion, in which the customer's opinion is reflected. The generated design is generated by the generation device 20. As a result, the analysis system 1 of the embodiment can quantify the customer's opinion as material information that the customer prefers, and can generate a design that reflects the texture of the material that the customer prefers.

Further, in the analysis system 1 of the embodiment, the target design has a three-dimensional shape. The analysis system 1 further includes a variation design processing section 13 (learning section). The variation design processing unit 13 generates a trained image generator that has learned the characteristics of a target image generated by converting three-dimensional information in the target design into image information in a two-dimensional image. The generation device 20 generates variation images using a trained image generator. The generation device 20 generates three-dimensional information of a variation design by converting image information in the generated variation image into three-dimensional information in a three-dimensional shape. Thereby, in the analysis system 1 of the embodiment, it becomes possible to generate a variation design having a three-dimensional shape using an existing image AI (an existing learning model developed for images).

All or part of the analysis system 1, learning device 10, generation device 20, and analysis device 30 in the embodiments described above may be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed. Note that the "computer system" herein includes hardware such as an OS and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into computer systems. Furthermore, a "computer-readable recording medium" refers to a storage medium that dynamically stores a program for a short period of time, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. It may also include a device that retains a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client in that case. Further, the above-mentioned program may be one for realizing a part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system. It may also be realized using a programmable logic device such as an FPGA.

Although the embodiments of the present invention have been described above in detail with reference to the drawings, the specific configuration is not limited to these embodiments, and includes designs within the scope of the gist of the present invention.

DESCRIPTION OF SYMBOLS 1... Analysis system, 10... Learning device, 12... Material processing section, 13... Variation design processing section (learning section), 20... Generation device (variation design generation section), 30... Analysis device, 31... Acquisition section, 32... Opinion quantification unit, 33...Opinion reflection design generation unit, 34...Opinion reflection design output unit, 35...Characteristic information

Claims (5)

Using a trained generator that has learned the characteristics of the target design, a variation design having the characteristics of the target design is generated, and the features in the generated variation design are used as vector values of each feature vector in a multidimensional feature vector space. a variation design generation unit that generates information associated with the variation design as feature information;
an opinion quantification unit that quantifies the opinion regarding the variation design by converting the opinion to a vector value of each feature vector in the multidimensional feature vector space using the feature information;
an opinion reflection design generation unit that causes the learned generator to generate a design in which the opinion is reflected, based on the quantified opinion;
An analysis system equipped with further comprising a material processing unit that generates material information indicating the material feel of the target design,
The variation design generation unit generates, as the variation design, a design in which the material information is synthesized with the design generated using the learned generator, and generates information associating the material information used in the synthesis with the variation design. is generated as the feature information,
The opinion quantification unit quantifies the opinion by converting the opinion into the material information that is synthesized into the variation design using the feature information,
The opinion reflection design generation unit generates a design in which the material information in the quantified opinion is synthesized with a design generated by the learned generator based on the quantified opinion, in which the opinion is reflected. causing the variation design generation unit to generate the design as a
The analysis system according to claim 1. The target design has a three-dimensional shape,
further comprising a learning unit that generates a trained image generator that has learned the characteristics of a target image generated by converting three-dimensional information in the target design into image information in a two-dimensional image,
The variation design generation unit generates a variation image using the learned image generator, and converts image information in the generated variation image into three-dimensional information in a three-dimensional shape, thereby creating a three-dimensional image of the variation design. generate information,
The analysis system according to claim 1. An analysis method performed by a computer,
A variation design generation unit generates a variation design having the characteristics of the target design using a trained generator that has learned the characteristics of the target design, and converts the characteristics of the generated variation design into each of the features in the multidimensional feature vector space. Generating information associated with the variation design as a vector value of a feature vector as feature information,
an opinion quantification unit quantifying the opinion regarding the variation design by converting the opinion to a vector value of each feature vector in the multidimensional feature vector space using the feature information;
an opinion reflection design generation unit causes the learned generator to generate a design in which the opinion is reflected, based on the quantified opinion;
Analysis method. to the computer,
Using a trained generator that has learned the characteristics of the target design, a variation design having the characteristics of the target design is generated, and the features in the generated variation design are used as vector values of each feature vector in a multidimensional feature vector space. generating information associated with the variation design as feature information;
Quantifying the opinion regarding the variation design by converting the opinion regarding the variation design into a vector value of each feature vector in the multidimensional feature vector space using the feature information;
Based on the quantified opinion, causing the learned generator to generate a design in which the opinion is reflected;
program.

Images