Design support device and design support method

Abstract

To provide a design support device and a design support method that can optimize designs, including design aesthetic.SOLUTION: There is provided a design support device for a shaped object that generates a shape by using a shape generation algorithm in an optimization calculation processing unit, the device further comprising a deep learning calculation processing unit for performing deep learning of features of a design desired to be cited. The optimization calculation processing unit inputs the shape generated by the shape generation algorithm into a trained network that has deep learned the features of the design desired to be cited, and optimizes a similarity between a feature amount of the shape and a feature amount of the design desired to be cited by using an optimization algorithm.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a design support apparatus and a design support method for supporting the design of a model. [Background technology]

[0002] The design of buildings, parts of buildings (fittings, interior and exterior), furniture, and other objects must meet the required functions and constraints (structural strength, production costs, etc.) in terms of functionality, safety, durability, solar radiation acquisition, etc. In order to meet these functions and constraints, designs are carried out using shape generation algorithms with set parameters.

[0003] In designs using conventional shape generation algorithms, the number of parameter patterns is enormous, making it difficult to search for optimal parameters that also satisfy functions and constraints, so Patent Document 1 proposes a design support device that automatically optimizes parameters for frame structures used for self-weight support or earthquake resistance of buildings and large structures, for example. In detail, the design support device in Patent Document 1 uses an optimization algorithm as an optimization method for calculating the optimal combination of the number, arrangement, and cross-sectional dimensions of steel members, which are parameters in the frame structure, to determine the adaptability of input parameters, automatically optimize the parameters, and perform optimal design of the frame structure. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2001-134628 A (pages 4 to 6, Figure 2) Summary of the Invention [Problem to be solved by the invention]

[0005] The design support device in Patent Document 1 can automatically optimize parameters using an optimization algorithm for quantitative evaluation axes such as the structural strength of a framework structure and production costs. However, it does not take into account design, which requires a qualitative evaluation. As a result, the design needs to be reviewed every time the design is changed, making it practically difficult to select a suitable design from among many designs.

[0006] The present invention has been made in light of these problems, and has an object to provide a design support device and a design support method that are capable of optimizing design, including aesthetic design. [Means for solving the problem]

[0007] In order to solve the above problems, the design support device of the present invention comprises: A design support device for a shaped object that generates a shape by a shape generation algorithm in an optimization calculation processing unit, Further comprising a deep learning processing unit that performs deep learning of the features of the design to be cited; The optimization calculation processing unit inputs the shape generated by the shape generation algorithm into a trained network that has undergone deep learning of the features of the design to be cited, and optimizes the similarity between the features of the shape and the features of the design to be cited using an optimization algorithm. According to this feature, the features of the design to be referenced and the shape generated by the shape generation algorithm are input into a trained network that has undergone deep learning, the design in that shape is expressed as computable features, and an optimization algorithm is used to optimize the similarity between the features of the design in that shape and the features of the design to be referenced, thereby providing support for the design of a structure having an optimized shape, including design, while satisfying specified evaluation criteria such as structural strength.

[0008] The deep learning calculation processing unit is characterized in that it performs deep learning using image data excluding color information. According to this feature, by using image data that has had color information removed for deep learning, the influence of color is suppressed, making it easier to express the features of the design.

[0009] The image data is characterized in that it is depth image data. According to this feature, by using the depth image data, rendering becomes easy and the geometric features of the design can be easily expressed.

[0010] The shape is characterized in that it is converted into image data excluding color information and input to the trained network. According to this feature, by converting the shape to be optimized into image data that excludes color information, the influence of color is suppressed, making it easier to express the features of the design.

[0011] The image data is characterized in that it is depth image data. According to this feature, by using the depth image data, rendering becomes easy and the geometric features of the design can be easily expressed.

[0012] The shape is characterized in that it is a shape generated based on a parameter t of shape=f(t) of an arbitrary parametric 3D model. According to this feature, by simply changing the value of the parameter t in shape=f(t), it is possible to generate a shape that optimizes the similarity to the features of the design to be cited.

[0013] The optimization calculation processing unit performs optimization until the similarity converges to a predetermined range. According to this feature, it is possible to obtain a plurality of shapes that are optimized for the design to be cited within a specified range of similarity, making it easy to optimize the design of the object.

[0014] The design support device is characterized in that it displays the plurality of shapes together with their optimization scores, and prompts the user to select one. According to this feature, it is possible to select a shape having design features similar to the design to be cited while using an optimization score based on the design features of the shape to be optimized as an index, thereby making it easier to optimize the design of the object.

[0015] The design support method of the present invention comprises the steps of: A design support method for supporting the design of a model, comprising the steps of: an optimization stage for optimizing the shape generated by the shape generation algorithm using an optimization algorithm; a deep learning stage for deep learning of the features of the design to be cited; The optimization stage includes a step of inputting the shape into a trained network that has deep-learned the features of the design to be cited in the deep learning stage, and a step of optimizing the similarity between the features of the shape and the features of the design to be cited by the optimization algorithm. According to this feature, the features of the design to be referenced and the shape generated by the shape generation algorithm are input into a trained network that has undergone deep learning, the design in that shape is expressed as computable features, and an optimization algorithm is used to optimize the similarity between the features of the design in that shape and the features of the design to be referenced, thereby providing support for the design of a structure having an optimized shape, including design, while satisfying specified evaluation criteria such as structural strength. [Brief description of the drawings]

[0016] [Figure 1] 1 is a schematic diagram showing a design optimization configuration in a design support apparatus and a design support method according to an embodiment of the present invention; [Diagram 2] 13(a) to 13(f) are diagrams showing a plurality of optimization target shapes optimized for a design to be cited (a sunshine filtering image) displayed on a selection screen of a display in the design support device of the embodiment, and their optimization scores (loss). [Diagram 3]3A and 3B are diagrams showing a plurality of optimization target shapes optimized for a design (a sunshine filtering image) to be cited, which are displayed on a selection screen of the display of the design support device of the embodiment, and their optimization scores (loss). In FIG. 3, the Voronoi boundary is generated as a frame. [Figure 4] FIG. 1A is a diagram showing a specific sun-filtering image used as a design to be cited, and FIG. 1B is a diagram showing an example of an optimization target shape optimized to approximate the specific sun-filtering image shown in FIG. 1A by a design support device of an embodiment. [Diagram 5] FIG. 1 is a schematic diagram showing an optimization configuration when optimizing to approximate a particular design to be cited. [Figure 6] FIG. 11 is a diagram showing how light like "sunlight filtering through the trees" is created by a structure having an optimization target shape that is optimized using a polygon-based geometric algorithm as a shape generation algorithm in the design support system of the embodiment. [Figure 7] FIG. 11 is a diagram showing how light like "sunlight filtering through the trees" is created by a structure having an optimization target shape that is optimized using a triangle-based geometric algorithm as a shape generation algorithm in the design support system of the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] Hereinafter, an embodiment of the present invention will be described. However, the present invention can be embodied in many different forms, and is not limited to the following embodiments and examples.

[0018] After much trial and error research, the inventors discovered that by expressing the qualitative and difficult-to-handle feature of the "design" of any citation object (artificial object, natural object) as a calculable feature quantity such as a vector through deep learning, it becomes possible to newly treat it as an evaluation axis for optimization in an optimization algorithm. Based on this knowledge, in an optimization algorithm for optimizing the shape of a structure such as a building, a part of a building (fittings, interior, exterior), or furniture generated by a shape formation algorithm, in addition to various evaluation axes related to the functions and constraints of the structure that have been conventionally treatable as evaluation axes, the design of the structure that is optimized while satisfying structural strength, etc., including design, can be used as an evaluation axis for optimization.

[0019] First, an outline of a design optimization configuration in the design support apparatus and design support method of the present embodiment will be described with reference to FIG.

[0020] As shown in Fig. 1, the design optimization configuration in this embodiment includes an optimization stage that optimizes a shape (input (hereinafter referred to as "optimization target shape")) generated by an arbitrary parametric 3D model (shape generation algorithm) using an optimization algorithm, and a deep learning stage that performs deep learning of the features of the design to be cited. The optimization stage also includes a step of inputting the optimization target shape into a trained network that has deep-learned the features of the design to be cited in the deep learning stage, and a step of optimizing the similarity between the design feature (current feature) of the optimization target shape and the feature (target feature) of the design to be cited using the optimization algorithm. This will be explained in detail below.

[0021] In the optimization stage of this embodiment, any parametric 3D model can change the shape to be optimized, which is generated based on the parameter t in shape=f(t). The shape to be optimized is input to the trained network, and the shape, i.e., the design features, can be expressed as computable features.

[0022] In the optimization stage, optimization is performed by an optimization algorithm (evolutionary algorithm) until the similarity between the design features of the optimization target shape and the design features of the design to be cited that have been deep-learned in advance in the deep learning stage converges to a predetermined range. Specifically, using Categorical Cross Entropy (CCE), which is an objective function, a parameter t is optimized so that the class probability for each class based on the design features of the optimization target shape is maximized for the class that has the design to be cited.

[0023] By optimizing the parameter t using the optimization algorithm, the shape to be optimized, which is generated based on the parameter t of shape=f(t) in the parametric 3D model, is changed, and the optimization score calculated by the objective function becomes lower, i.e., the similarity between the design features of the optimized shape and the design to be cited becomes higher.

[0024] Here, the deep learning stage will be described. In the deep learning stage, a trained network is formed by having a deep learning network learn the shape features of the design to be cited. In this embodiment, a set of learning images (input) and labels (labels) are input to the deep learning network. As the learning images to be input to the deep learning network, in addition to the image of the design to be cited, it is preferable to use multiple types of images of natural objects and artificial objects having a shape similar to the design to be cited. The label corresponds to a correct answer value indicating what image the learning image input to the deep learning network is, and is used for classification as a class (class).

[0025] In the deep learning stage of this embodiment, the learning image input to the deep learning network is preferably image data converted based on specific elements in the image from the viewpoint of making it easier to evaluate (express) the features of the design. Also, for the optimization target shape input to the trained network, it is preferable to use image data converted based on specific elements from the same viewpoint. Note that the specific elements may be appropriately selected according to the features of the design to be cited, and examples of image data converted based on specific elements include depth image data, luminance image data, color image data, and three-dimensional point cloud data.

[0026] In addition, from the viewpoint of improving the accuracy of optimization, it is preferable that the training image input to the deep learning network in the deep learning stage and the optimization target shape input to the trained network are image data converted based on the same element as a specific element.

[0027] In addition, when evaluating the geometric features of a design, it is preferable to use image data that has been converted based on specific elements and that does not include color information, such as depth image data, brightness image data, etc.

[0028] In addition, the training images input to the deep learning network and the optimization target shapes input to the trained network do not necessarily have to be image data converted based on specific elements, and normal image data (e.g., photo data, 3D-CG data, 3D CAD model data, etc.) may be input.

[0029] The design support device of this embodiment may be configured to use a display means such as a display to display multiple optimization target shapes generated in the optimization stage along with their optimization scores, as shown in Fig. 2, for example, to encourage selection. Note that in Fig. 2, the optimization score displayed is a loss (see Fig. 1) that indicates how similar the trained network has determined that the design features of the optimization target shape are to the design features to be cited, but this is not limiting, and an optimization score calculated by another objective function that can be incorporated in the optimization stage, for example, may also be displayed. EXAMPLES

[0030] Here, a design support device according to the embodiment described above was actually manufactured and used to support the design of a shaped object.

[0031] The design support device of the embodiment is mainly composed of an optimization calculation processing unit that generates the shape of a molded object using a shape generation algorithm, a deep learning calculation processing unit that performs deep learning of the features of the design to be cited, and a display that can display the optimization target shape generated by the shape generation algorithm together with the optimization score.

[0032] In this embodiment, the design support device performs design support for creating a structure having a shape capable of producing light like "sunlight filtering through the trees" as a design to be cited, such as a curtain wall or louver in architecture, or a screen, partition, or lighting fixture in interior design. Note that a shape capable of producing light like "sunlight filtering through the trees" can be rephrased as the shape of branches and leaves that creates a comfortable light leak.

[0033] In the deep learning stage of this embodiment, in addition to images of "sunlight filtering through the trees" (multiple types of images including that shown in Figure 4(a)), images of "stone walls," "sliding doors," "dining tables," "monasteries," "libraries," and "fences" were used as learning images.

[0034] In this embodiment, the learning image input to the deep learning network and the optimization target shape input to the trained network are converted into depth image data. By converting normal image data into image data from which color information has been removed, such as depth image data converted based on depth, rendering becomes easier and designs having shape characteristics other than color, such as a "shape that can produce light like sunlight filtering through trees," can be easily handled.

[0035] In the optimization calculation processing unit of this embodiment, the shape generation algorithm uses a geometric algorithm that generates a three-layer Voronoi panel shape. In detail, the shape generation algorithm generates a shape using the Voronoi density for each layer, the ratio of polygons to be opened, and the pattern as parameters.

[0036] As shown in Fig. 2, the selection screen of the display of the design support device displays multiple optimization target shapes generated by a shape generation algorithm along with their optimization scores (loss). In this embodiment, a multi-class cross entropy, which is a loss function, is used to calculate the loss in the deep learning stage, and the closer the predicted value is to the correct answer, the closer the loss value is to 0, and the farther the predicted value is from the correct answer, the larger the loss value is. In other words, the smaller the loss value (closer to 0), which is the optimization score, the more likely it is that the trained network has mistaken the depth image of the optimization target shape for the depth image of "sunlight filtering through the trees."

[0037] 2(a)-(c), the optimization target shapes with loss values ​​smaller than 2 have many openings (white areas) and many complex overlaps and gaps between layers, whereas the optimization target shapes with loss values ​​larger than 2, as shown in FIG2(d)-(f), have few openings and few complex overlaps and gaps between layers. Note that in this embodiment, "openings" refer to empty areas in each layer, and "gaps" refer to areas where the openings in each layer overlap.

[0038] In addition, for the optimization target shapes shown in Figs. 2(a) to (f), the results of an analysis of relative impressions (N=14) conducted through a questionnaire using the paired comparison method are shown in Table 1.

[0039] [Table 1]

[0040] As shown in Table 1, the optimized shapes in Figures 2(a) to (c) with small loss values ​​received more votes, i.e., people tended to recognize the design as being similar to "sunlight filtering through the trees." Note that the optimized shape in Figure 2(b) (loss 0.008) received the most votes, while the optimized shape in Figure 2(a) (loss 0.001) and the optimized shape in Figure 2(c) (loss 1.817) received similar numbers of votes, indicating that designs with excellent optimization scores (loss values ​​close to 0) are not necessarily recognized by people as being similar to "sunlight filtering through the trees."

[0041] In addition, there is a large difference in the number of votes between the optimization target shape in Fig. 2(c) (loss 1.817) and the optimization target shape in Fig. 2(d) (loss 2.178), which have similar loss values. This is presumably due to the fact that the optimization target shape in Fig. 2(d) has fewer openings and holes.

[0042] Next, a geometric algorithm that uses the Voronoi boundary as a frame is used as a shape generation algorithm.

[0043] As shown in Figure 3, by using a geometric algorithm that uses the Voronoi boundary as a frame as a shape generation algorithm, it is possible to prevent floating polygons from occurring, as shown in the optimization target shape in Figure 2(b), for example. This makes it possible to optimize the design of the object while ensuring that manufacturing constraints are met using the geometric algorithm, and provides support for the design of the object, including its design.

[0044] It is also possible to perform optimization so as to approach a specific "sun filtering through the trees" depth image (see FIG. 4(a)) like the optimization target shape shown in FIG. 4(b). This is possible by performing optimization using the cosine similarity with the feature of a specific "sun filtering through the trees" depth image as the objective function, as shown in FIG. 5. In detail, in the optimization stage, the optimization algorithm optimizes the parameter t, so that the vector end point (see cross mark) of the design feature (current feature) of the optimization target shape in the feature space can be brought closer to the representative vector end point R of the feature of the design to be cited.

[0045] Furthermore, as shown in Fig. 6, by combining multiple shapes to be optimized, it is possible to support the design of a structure having a shape that can produce light like "sunlight filtering through the trees," and it is possible to explore the balance with the evaluation axis of other functions such as solar radiation acquisition. It goes without saying that it is also possible to explore the balance with the evaluation axis of other functions such as solar radiation acquisition using a single shape to be optimized.

[0046] In Figures 2 to 6, the shapes formed by Voronoi are mainly polygons such as quadrangles, pentagons, and hexagons. However, even if a triangle-based geometric algorithm is used as the shape generation algorithm, it is possible to assist in the design of a structure having a shape that can create a "sunlight filtering through the trees" effect within the range that the shape generation algorithm can generate, as shown in Figure 7, enabling a wide range of designs.

[0047] In addition, although an example has been described in which the boundary portion of a geometric shape based on polygons and triangles formed by Voronoi is used as a frame to ensure manufacturing constraints, optimization may be performed so as to satisfy the evaluation axis of structural strength based on the configuration of the frame generated by a geometric algorithm. Also, the portion that primarily ensures structural strength may be something other than the frame.

[0048] As explained above, by inputting the features of the design to be cited of the shape (optimized shape) generated by the shape generation algorithm into a trained network that has undergone deep learning, expressing the shape as a feature of the design, and optimizing the similarity to the feature of the design to be cited by the optimization algorithm, it is possible to provide design support for a structure that is optimized, including designability, while satisfying predetermined evaluation axes such as structural strength. In other words, by adding evaluation axes based on the design expressed as features that can be calculated by deep learning to quantitative evaluation axes such as structural strength, allowable stress, solar radiation acquisition, and production costs that could be handled by the shape generation algorithm, it is possible to provide design support for a structure having a shape optimized to the features of any design to be cited within the range of shapes that can be generated by the shape generation algorithm.

[0049] Although the embodiments of the present invention have been described above with reference to the drawings, the specific configuration is not limited to these embodiments, and the present invention also includes modifications and additions that do not deviate from the gist of the present invention.

[0050] For example, in the above embodiment, an example of design support for creating a structure having a shape capable of producing light like "sunshine filtering through the trees" as a design to be cited is described, but the design to be cited may be selected from any man-made or natural object. In addition, by using an image converted based on a specific element suitable for the characteristics of the selected design to be cited, it is possible to easily express the characteristics of the design.

[0051] In addition, in the above embodiment, a configuration has been described in which the multi-class cross entropy, which is a loss function, is used to calculate the loss in the deep learning stage. However, this is not limiting, and any loss function may be used.

[0052] In the above embodiment, the multi-class cross entropy or cosine similarity is used to calculate the similarity in the optimization stage, but the present invention is not limited to this and any objective function may be used. In addition, in the optimization stage, the objective function may be treated as either a maximization problem or a minimization problem. [Industrial Applicability]

[0053] The present invention expresses the qualitative characteristics of any cited design as a feature that can be calculated using deep learning, and can be used as an evaluation axis for an optimization algorithm. It can provide support for the design of objects that are optimized not only for functionality, safety, durability, etc., but also for design within the range of shapes that can be generated by a shape generation algorithm. Therefore, by applying a wide range of designs optimized for any cited design to objects, the present invention can be applied to the development of new products and has industrial applicability.

Claims

1. A design support device for a shaped object that generates a shape by a shape generation algorithm in an optimization calculation processing unit, Further comprising a deep learning processing unit that performs deep learning of the features of the design to be cited; The optimization calculation processing unit inputs the shape generated by the shape generation algorithm into a trained network that has undergone deep learning of features of the design to be cited, and optimizes the similarity between the features of the shape and the features of the design to be cited by an optimization algorithm.

2. The design support device according to claim 1 , wherein the deep learning calculation processing unit performs deep learning using image data excluding color information.

3. 3. The design support device according to claim 2, wherein the image data is depth image data.

4. 2. The design support device according to claim 1, wherein the shape is converted into image data excluding color information and input to the trained network.

5. 5. The design support device according to claim 4, wherein the image data is depth image data.

6. The apparatus of claim 1 , wherein the shape is generated based on a parameter t of shape=f(t) of an arbitrary parametric 3D model.

7. 7. The design support apparatus according to claim 1, wherein the optimization processing unit performs optimization until the similarity converges to a predetermined range.

8. The design support device according to claim 7 , wherein the design support device displays a plurality of the shapes together with their optimization scores, and prompts the user to select one of the shapes.

9. A design support method for supporting the design of a model, comprising the steps of: an optimization stage for optimizing the shape generated by the shape generation algorithm using an optimization algorithm; a deep learning stage for deep learning of the features of the design to be cited; The design support method, wherein the optimization stage includes a step of inputting the shape to a trained network that has deep-learned features of the design to be cited in the deep learning stage, and a step of optimizing a similarity between features of the shape and features of the design to be cited by the optimization algorithm.

Images