A product form generation design method based on deep learning and parametric modeling

Abstract

The present invention discloses a product form generation design method based on deep learning and parametric modeling, which relates to the technical field of styling design and solves the technical problem of rapid three-dimensional model generation for complex products facing user intentions. The key points of its technical solution are to integrate the professional domain knowledge in design science into the generative design algorithm. This method mines the user's perceptual awareness of the style image of product styling elements, uses the degree of conformity between the styling and the image as the design guidance to help designers accurately grasp the design feature elements and user needs, reduce design risks. At the same time, a generative product design system is established through deep learning and parametric modeling technologies to assist designers in quickly generating design schemes in the conceptual design stage, efficiently and extensively exploring the design space, enabling rapid exploration of product families, accelerating the product iteration speed, and shortening the R & D cycle.

Description

Technical Field

[0001] This application relates to the technical field of product styling design, and particularly to a product form generation design method based on deep learning and parametric modeling. Background Art

[0002] The design process of complex product styling is relatively long. Taking automobiles as an example, the styling features of automobiles focus on the description of the overall style of automobiles, the changes in form and surface, the expression of details, and the family characteristics that continue for multiple generations. Relevant scholars have studied the styling features of automobiles from aspects such as styling elements, feature extraction, style recognition, and brand styling genes. However, most research results are mainly applied to traditional automobile design means, such as hand-drawing automobile sketches. After the designer determines the sketch plan, it is handed over to a CAS (Computer Aided Styling) digital model engineer to convert it into a three-dimensional model. During the process of establishing the three-dimensional model, the CAS engineer needs to communicate with the designer repeatedly and adjust the positional relationship of curves and key points many times, resulting in low efficiency. Therefore, in the wave of computational design and parametric design, how to apply the research results related to automobile styling features to computer-aided design and realize the intelligence of automobile conceptual design has great research value.

[0003] In the early generative design research, generative design was usually defined as algorithmically modifying the initial form according to set rules to generate a series of different design schemes. There are many application cases in architecture because the styling features of geometric stacking of building objects are very suitable for three-dimensional form generation based on voxels, and there are few applications in product design. There is still a lack of a good implementation method for generating product design schemes. The two most commonly used methods are evolutionary computation and parametric modeling. With the development of computer-aided design, related technologies such as parametric design, shape grammar, cellular automata, genetic algorithms, fractal loops, and deep learning have gradually been applied to the field of generative design. Generative design is no longer limited to the generative process of evolutionary algorithms. There are a large number of cases showing that other computer-aided design methods have also achieved good results in the field of generative design.

[0004] Generally, generative design usually goes through the following steps:

[0005] The first step: Define the design object and the solution goal.

[0006] The second step: Select an appropriate generative design technical route.

[0007] The third step: Construct a generative design system.

[0008] The fourth step: Optimize and adjust the algorithms and parameters of the generative design system.

[0009] The fifth step: Select the final design scheme.

[0010] The current product generative design methods have the following key problems: The product generative design method based on evolutionary computation mainly studies the relationship between product feature lines and perceptual images, and combines genetic algorithms to achieve the "optimal solution of perceptual image vocabulary" of product feature lines at the two-dimensional level. It is limited to a single feature line and the two-dimensional level, and has very limited guiding effect on the actual complex three-dimensional modeling design. The product generative design method based on deep learning takes the two-dimensional side view as the research object. Since deep learning requires a large amount of parallel computing, the image data samples usually collected need to be processed into low-resolution sizes. This generative design method has problems such as huge demand for data samples, difficulty in collecting data sets, poor quality of generated images, and uncontrollable generation results; The product generative design based on parametric modeling is also limited to the two-dimensional product feature line level, and few people conduct research on related intelligent auxiliary design methods for product feature lines at the three-dimensional level of complex products. Summary of the Invention

[0011] This application provides a product form generative design method based on deep learning and parametric modeling, and its technical purpose is to achieve the generation of complex three-dimensional products at the three-dimensional level.

[0012] The above technical purpose of this application is achieved through the following technical solutions:

[0013] A product form generative design method based on deep learning and parametric modeling, including:

[0014] S1: Extract the two-dimensional product feature line image with the greatest relevance to the product shape, conduct user multi-dimensional psychological intention research on this two-dimensional product feature line image to obtain user psychological intention data, and convert the user psychological intention data into category labels in each dimension through clustering, then obtain a two-dimensional product feature line image with category labels. The two-dimensional product feature line in the two-dimensional product feature line image is the intention-driven feature line;

[0015] S2: Input the two-dimensional product feature line image with category labels into a deep learning network to generate a set of two-dimensional feature lines, and construct the projection relationship between any two-dimensional feature line in the set of two-dimensional feature lines and each view of the product to obtain the three-dimensional feature curve of the product; Among them, the two-dimensional feature line used to construct the three-dimensional feature curve is the driving feature line;

[0016] S3: Associate other two-dimensional feature lines in the set of two-dimensional feature lines with the driving feature line to parametrically reconstruct the three-dimensional feature curve, and obtain a three-dimensional feature curve parameter model corresponding to the driving feature line;

[0017] S4: Drive the three-dimensional feature curve parameter model through the intention-driven feature line to generate a morphological transformation, thereby generating a new product form.

[0018] Further, in step S1, the user psychological intention data is converted into category labels under each dimension by clustering, including:

[0019] The user conducts a Likert scale scoring on the style intention of the product sample image, and performs cluster analysis on the user psychological intention data according to the Likert scale scoring to obtain the category labels of each product sample image based on the compliance degree score of each style intention. The category label with the highest compliance degree score is the style intention category label of the product sample image, and thus a two-dimensional product feature line image with category labels is obtained;

[0020] Among them, the style intention is defined by perceptual vocabulary.

[0021] Further, the deep learning network is a VAE-GAN network, and the VAE-GAN network includes an encoder, a generator, and a discriminator.

[0022] Further, in step S3, other two-dimensional feature lines in the two-dimensional feature line set are associated with the driving feature line, including: the driving feature line includes a positioning reference point, and the control points in the two-dimensional feature line of the product establish a parameter association through the relative position relationship with the positioning reference point on the projection view;

[0023] Among them, the positioning reference point is obtained according to the empirical value in the design convention.

[0024] Further, in step S3, the two-dimensional feature lines in the two-dimensional feature line set are all third-order Bezier curves, and each two-dimensional feature line includes 4 control points, and the 4 control points are represented as where j represents the curve number; and respectively represent the starting point and the ending point of the third-order Bezier curve, and represent the two middle control points of the third-order Bezier curve, then the third-order Bezier curve is represented as:

[0025]

[0026] where t represents a real number in the interval [0,1].

[0027] Further, and The positional relationship of includes:

[0028] and None of them lies on the cubic Bezier curve;

[0029] Or when the cubic Bezier curve is a straight line, and both lie on the cubic Bezier curve.

[0030] The beneficial effects of this application are as follows: The product form generative design method based on perceptual image and parametric modeling proposed in this application integrates professional domain knowledge in design science into the generative design algorithm. This method explores the user's perception of the style image of product shape elements, uses the degree of conformity between the shape and the image as the design guidance, helps designers accurately grasp the design feature elements and user needs, reduces design risks, and at the same time establishes a generative design system through deep learning and parametric modeling technologies, assisting designers to quickly generate design schemes in the conceptual design stage, efficiently and extensively explore the design space, enabling rapid exploration of product families, accelerating the product iteration speed, and shortening the R & D cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flowchart of the method described in this application;

[0032] Figure 2 is a structural diagram of the VAE - GAN network;

[0033] Figure 3 is a schematic diagram of a 3 - order 4 - point NURBS curve;

[0034] Figure 4 is a schematic diagram of the feature lines and numbers extracted from each view in a specific embodiment;

[0035] Figure 5 is a schematic diagram of the key positioning reference points in a specific embodiment;

[0036] Figure 6 is a schematic diagram of the body length transformation and height transformation in a specific embodiment;

[0037] Figure 7 is a schematic diagram of the morphological change of the associated feature lines driven by the position movement of the front windshield curve in a specific embodiment;

[0038] Figure 8 is a flowchart of generating a car form driven by the side - view top - type line in a specific embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The technical solution of this application will be described in detail below with reference to the drawings.

[0040] For the convenience of public understanding, the technical solution of this application will be described below taking an automotive product as an example, specifically including:

[0041] (1) Research on the Image in the Context of Automotive Generative Design:

[0042] Product style image can be defined by perceptual words. Multiple automotive samples are collected, and several of the most representative perceptual words are selected using the expert evaluation method, such as "sporty", "elegant", "grand", "steady", and "modern". Users are asked to rate the collected automotive samples on the selected perceptual words using a Likert scale. The user attitude data is systematically cluster - analyzed to obtain the compliance category labels of each sample image based on the evaluation scores of each style image word, such as three labels: "very sporty", "relatively sporty", and "not sporty", which lays the foundation for generating the side - view top - type line that conforms to the user's psychological intention in the next stage using a deep - learning - based generative design method.

[0043] (2) Generation of Two - Dimensional Product Feature Lines Based on Deep Learning:

[0044] This application selects the variational auto - encoder - based generative adversarial network VAE - GAN to generate image samples of the side - view top - type line, and its structure is as Figure 2 shown. First, fully - connected neural networks for the encoder Encoder, generator Generator, and discriminator Discriminator are established, and their network parameters θ Enc 、θ Gen 、θ Dis are initialized. The image data set X of the side - view top - type line is loaded. The encoder Encoder first receives an m - sized minibatch of the original data sample set X = {x (1) ,...,x (m)}, compresses and encodes the original image to obtain the feature - encoded latent vector Z = {z (1) ,...,z (m)} based on the true distribution. The mean vector set μ = {μ (1) ,...,μ (m)} and the variance vector set σ = {σ (1) ,...,σ (m)} of the latent vector Z are calculated. Let Z′ = μ+exp(σ)·ε, ε~N(0, I), to obtain the combined latent vector set Z′ = {z′ (1) ,...,z′ (m)}.

[0045] The generator Generator receives Z′ as input and generates a new image sample set Y = {y (1) ,...,y (m)}, finally, the generated image sample Y and the original sample X are input into the discriminator Discriminator. The discriminator outputs the probability value of judging the image as a real image, and then, through stochastic gradient descent, backpropagation and iterative update are performed on the discriminator Discriminator and the generator Generator respectively. The update formula of the discriminator is shown in Equation (1), and the update formula of the generator is shown in Equation (2). When the number of training epochs reaches the set number of Epochs, stop training and save the trained VAE-GAN model parameters to the specified directory.

[0046]

[0047]

[0048] (3) Construction of the 3D feature curve parameter model of the vehicle:

[0049] To establish an interrelated 3D feature curve parameter model of the vehicle, it is first necessary to extract and quantitatively express the 2D body feature lines. In industrial modeling software, NURBS curves are usually used to depict the modeling feature lines. The formula of the NURBS curve is shown in Equation (3). The NURBS curve is defined by four elements - control points P i , weight values w i , knot vector u, and order p; among them, the control points include P0, P1, P2,..., P n , a total of n + 1; the weight values include w0, w1, w2,..., w n , a total of n + 1; the knot vector includes u0, u1,..., u m , a total of m + 1; the quantitative relationship among m, n, and p satisfies m = n + p + 1.

[0050]

[0051] NURBS can flexibly represent various curves and can solve problems such as spline curves being unable to draw ellipses, circles, and hyperbolas. However, its function definition is relatively complex and has many parameters, making it troublesome to extract certain parameter values of the NURBS curve and not conducive to subsequent establishment of the vehicle 3D curve parameter model. Therefore, this application uses a 3rd-order 4-point NURBS curve with fixed weight values, as Figure 3 shown. At this time, it is simplified to a Bezier curve, and its expression is shown in Equation (4), where t represents a real number in the interval [0, 1], used to determine the position and density of the output discrete points; P represents the control point coordinates. Through this representation method, it is ensured that each 2D feature line has the same topological structure and geometric expression, all being 3rd-order Bezier curves, and the four control points can be expressed as where j represents the curve number, and respectively represent the starting point and the ending point of the curve, and represent two control points in the middle of the curve, and are not necessarily on the curve, and there are various situations for their positions: (1) and are both not on the curve; (2) In the special case where the curve is a straight line, and are both on the curve. The simplified curve description method is convenient for quickly extracting coordinate parameter variables, which is beneficial to improving the efficiency of establishing the three-dimensional feature curve parameter model of the car subsequently.

[0052]

[0053] Taking a car (2016 Audi A6) as an example, this application extracts the two-dimensional body shape feature lines from 4 views. The feature lines and numbers extracted from each view are as Figure 4 shown. A total of 44 feature lines are extracted. By extracting the two-dimensional feature lines of the body shape (the set of two-dimensional feature lines in step S2) and establishing the correct projection relationship with each view, the three-dimensional curve network of the body shape can be described in three-dimensional space. The feature lines of the vehicle front face are jointly constructed by the top view and the left view, including the front edges of the windshield, hood, intake grille and other surfaces; the feature lines at the front corner of the vehicle are projected from the left view and the front view, including curves such as the A-pillar, headlights and fender edges; the top view and the front view project the feature lines of the side window, door, waistline, etc.; the right view and the side view establish the feature lines at the rear fender, rear roof and C-pillar of the vehicle; the top view and the right view jointly determine the feature curves of the rear window, trunk, taillights, etc. in three-dimensional space.

[0054] After extracting the two-dimensional feature lines of the body, it is necessary to build a three-dimensional feature curve parameter model with the help of a parametric modeling platform. In this study, the parametric programming plug-in Grasshopper based on the mainstream three-dimensional design software Rhinoceros 7.0 for industrial design is selected to build the three-dimensional feature curve parameter model of the car. This application explores a method for generating the three-dimensional shape change of the car driven by the change of the side view top line. Therefore, when reconstructing the parameters of the three-dimensional feature lines of the car, it is necessary to associate other feature lines with the side view top line and change the car shape according to the relative position of the side view top line. First, it is necessary to define the key positioning reference points, such as Figure 5As shown, the side view top line is represented by a thick solid line. A and B represent two points, which are the front and rear endpoints respectively. The contour line of the lower edge of the vehicle body is represented by a dashed line. During the process of establishing the parameter correlation model, points A and B are used as the positioning reference points. The control points of the feature line on the front side of the vehicle body establish parameter correlation with point A according to the relative vector relationship, and the control points of its feature line on the rear side of the vehicle body establish parameter correlation with point B according to the relative vector relationship. Among them, the control points of the lower edge line of the vehicle body (the dashed part) always keep the Z-axis coordinate unchanged and only change the coordinate values in the X and Y directions, that is, control the height of the lower edge of the vehicle body to remain unchanged, serving as the bottom reference for the vehicle body height. When the height of the side view top line changes, it drives the change of the vehicle body height. When the positions of the positioning reference points A and B change, it drives the change of the overall vehicle length. As Figure 6 shown.

[0055] In addition to simply driving the overall vehicle dimensions through the key position transformation of the reference points, the more important significance of establishing the three-dimensional feature curve parameter model lies in changing the curve shapes and scales of each feature line. Now, take the feature curve near the A-pillar of the vehicle body as an example for illustration. As Figure 7 shown. Figure 7 In (a), it shows the initial position and shape of the feature line. Among them, curve A is the side edge line of the engine hood, curve B is the roof contour line near the vehicle side window, curve C is the rear edge line of the engine hood, that is, the common boundary curve between the engine hood and the front windshield, and curve D is the upper edge line of the front windshield. Since they are all 3rd-order 4-point Bezier curves, the control points of curves A, B, C, and D are When curves C and D move to the positions in Figure 7 (b), curves C' and D' are obtained respectively. Curve A is scaled three-dimensionally with unequal ratios in a two-point positioning manner to obtain curve A' positioned according to the endpoints of C'. Since curve B has established parameter correlation with curve A, curve B changes to curve B' according to the position change of curve A'. As Figure 7 shown in (c). Finally, B' is scaled and transformed in a two-point positioning manner to obtain B''. As Figure 7 shown in (d).

[0056] Based on the above curve transformation idea, after adjusting other feature lines in the same steps, the three-dimensional feature curve parameter model of the whole vehicle can be obtained. When the shape and position of the side view top line change, the three-dimensional feature curves of the whole vehicle can change accordingly.

[0057] (4) Generate the vehicle form based on the side view top line:

[0058] After parametrically reconstructing the 3D feature curves of the car in the parametric modeling platform Grasshopper, a parametric model of the car's 3D feature curves based on the shape and scale of the side view top line is obtained. Combining with the side view top line generation results of deep learning (i.e., the set of 2D feature lines in step S2), taking the side view top line image generation results of each style image vocabulary (i.e., the intention-driven feature lines in step S1) as the input variables of the parametric model of the car's 3D feature curves, a large number of car shapes with different forms and conforming to the user's style image psychological cognition can be generated. In addition, since the overall car shape is mainly driven by the side view top line, other feature lines of the car still retain features highly similar to the initial state, making the generated results maintain the consistency of the design language with the original samples, which is conducive to designers quickly exploring car shape design schemes that conform to the brand's family design features in the conceptual design stage.

[0059] The process of generating the car shape driven by the side view top line is as Figure 8 shown. First, store the side view top line results based on deep learning in folders according to different style image vocabularies respectively. Then, read the generated side view top line sample images in the parametric modeling platform Grasshopper, and read the sample images as Mesh format within the set rectangular area. Since the generated pattern samples are binary images, according to the color values of the grid vertices, a color value of 0 represents the black area of the image, and the grid vertices with color values not equal to 0 represent the area of the side view top line. After screening out the vertices with non-zero color values, the side view top line of the generated image can be extracted as a NURBS curve. Finally, replace the initial side view top line sample with the extracted side view top line within the standard rectangular range, and the parametric model of the car's 3D feature curves will generate a large number of form design results with different forms according to the shape and scale changes of the side view top line.

[0060] The above is an exemplary embodiment of this application, and the protection scope of this application is defined by the claims and their equivalents.

Claims

1. A product form generation design method based on deep learning and parametric modeling, characterized in that, Including: S1: Extract the two-dimensional product feature line image with the highest correlation with the product shape. Conduct a user multi-dimensional psychological intention study on this two-dimensional product feature line image to obtain user psychological intention data. Convert the user psychological intention data into category labels in each dimension through clustering, and then obtain a two-dimensional product feature line image with category labels. The two-dimensional product feature line in the two-dimensional product feature line image is the intention-driven feature line; S2: Input the two-dimensional product feature line image with category labels into a deep learning network to generate a set of two-dimensional feature lines. Construct the projection relationship between any two-dimensional feature line in the set of two-dimensional feature lines and each view of the product to obtain the three-dimensional feature curve of the product; among them, the two-dimensional feature line used to construct the three-dimensional feature curve is the driving feature line; S3: Associate the other two-dimensional feature lines in the set of two-dimensional feature lines with the driving feature line to parametrically reconstruct the three-dimensional feature curve, and obtain a three-dimensional feature curve parameter model corresponding to the driving feature line; S4: Drive the three-dimensional feature curve parameter model to generate a morphological transformation through the intention-driven feature line, thereby generating a new product form.

2. The product form generation design method according to claim 1, characterized in that In step S1, converting the user psychological intention data into category labels in each dimension through clustering includes: The user conducts a Likert scale score on the style intention of the product sample image. Cluster analysis is performed on the user psychological intention data according to the Likert scale score to obtain the category labels of each product sample image based on the compliance degree score of each style intention. The category label with the highest compliance degree score is the style intention category label of the product sample image, and then a two-dimensional product feature line image with category labels is obtained; Among them, the style intention is defined by perceptual vocabulary.

3. The product form generation design method according to claim 1, characterized in that, The deep learning network is a VAE-GAN network, and the VAE-GAN network includes an encoder, a generator, and a discriminator.

4. The product form generation design method according to claim 1, wherein, In step S3, associating the other two-dimensional feature lines in the set of two-dimensional feature lines with the driving feature line includes: The driving feature line includes a positioning reference point, and the control points in the two-dimensional feature line of the product establish a parameter association through the relative position relationship with the positioning reference point on the projection view; Among them, the positioning reference point is obtained according to the empirical value in the design convention.

5. The product form generation design method according to claim 1, characterized in that, In step S3, the two-dimensional feature lines in the two-dimensional feature line set are all third-order Bezier curves, and each of the two-dimensional feature lines includes 4 control points, and the 4 control points are expressed as where j represents the curve number; and respectively represent the starting point and the ending point of the third-order Bezier curve, and represent the two middle control points of the third-order Bezier curve, then the third-order Bezier curve is expressed as: Among them, t represents a real number in the interval [0,1].

6. The product form generation design method according to claim 5, characterized in that and The positional relationships include: and are not on the cubic Bezier curve; When the cubic Bezier curve is a straight line, and both lie on the cubic Bezier curve.

Images