CN116258818A

CN116258818A - Design system of three-dimensional bionic product

Info

Publication number: CN116258818A
Application number: CN202310191211.7A
Authority: CN
Inventors: 梁天一; 毛溪
Original assignee: East China Normal University
Current assignee: East China Normal University
Priority date: 2021-07-30
Filing date: 2021-07-30
Publication date: 2023-06-13
Also published as: CN114445586B; CN114445586A

Abstract

The invention discloses a three-dimensional bionic design method based on a generated countermeasure network, which comprises the following steps: step 1: building a training set; step 2: establishing a depth generation model, and training the depth generation model through the training set; the depth generation model comprises an implicit automatic encoder and a hidden vector generation model; the implicit automatic encoder comprises an encoder 3DCNN and a Decoder IM-Decoder; the hidden vector generation model adopts an countermeasure generation network; step 3: and generating a three-dimensional bionic product by using the depth generation model. The invention also discloses a system for realizing the method, which comprises a data preprocessing module, a depth generation model, a sample vector repository and a post-processing module.

Description

Design system of three-dimensional bionic product

Technical Field

The invention belongs to the technical field of machine learning and intelligence, and relates to a design system of a three-dimensional bionic product.

Background

The bionic Design (Bionics Design) is to simulate the characteristics of the structure, function, shape, color, etc. of the nature and biological system, and perform creative Design. The bionic design starts from psychological demands of people, takes a personal bionic idea of a designer as a host, creates a design product full of biological interests and natural forms, and in turn serves the human beings, thereby completing the circulation from the human beings to the human beings in a natural state. Depending on the manner of application of bionics, the bionic design is mainly classified into morphological bionics, functional bionics, structural bionics, color bionics, texture bionics, and the like of the product. Morphological bionics are widely used in modern product modeling designs, such as household articles, small household appliances, furniture, vehicles, and the like. Today designers have been directing product model innovation towards the field of biomimetic design.

In order to generate more excellent creative designs, the academia and industry have been trying to explore the laws of the creative designs of the biomimetic shapes by means of computer technology for many years. Before the depth generation model is applied to the bionic design of the product form, the academic world establishes a modeling database and extraction rules through evolution design, uses the evolution algorithm to define fitness and a product characteristic model, and generates diversified design schemes on the product contour through extraction of contour characteristic elements of the bionic object. Recent studies have attempted to automatically synthesize the modeling contours of biomimetic products using deep learning techniques. Particularly, the Simiao Yu and the like apply the depth generation models such as the generation countermeasure network and the like to the bionic design of the appearance of the product for the first time, and the depth generation models are designed to form the bionic design of the appearance outline of the product by performing mathematical definition on the bionic design problem from the angle of deep learning.

Depth generation models combined from encoders and Generation Antagonism Networks (GAN) have shown unique advantages and great potential in the areas of planar image generation, restoration, and style migration. In terms of three-dimensional design, wu et al, university of Massa, 2016, proposed a specific method for applying the generation of an antagonistic network to three-dimensional shape generation, and the trained model could reconstruct a photograph of a chair into a three-dimensional model of the chair by restoration. The concept of hidden vector space in a three-dimensional shape representation is first proposed using 3D data models based on voxelization. The application of self-encoders and intermediate hidden generation antagonism networks (latex gan) by Panos et al completes the embedding and representation of three-dimensional point cloud data. A hidden vector space is trained directly based on a three-dimensional data set by using a self-encoder, then a series of three-dimensional generation tasks are completed by using bijection between an European space and a three-dimensional shape, and the diversity of the generated result is ensured by the existence of an intermediate LatentGAN. The effect of 3D model fitting of depth generation model latex gan is optimized by the implicit method by chen et al, so that the effect of three-dimensional model generated based on interpolation is more continuous and smooth, three-dimensional data of voxels are reused, and voxel data are converted into more widely applied three-dimensional grid data by marchang cube algorithm. However, the depth generation model has not been applied in the design process of the bionic product morphology.

Disclosure of Invention

The invention provides a three-dimensional bionic design system based on a generation countermeasure network, which comprises:

the data preprocessing module is used for inputting a three-dimensional shape data set represented by a bionic object and a design target, expanding the data set through data enhancement based on the design target and performance, and obtaining a symbol distance field of a three-dimensional model required by model training through a mesh2voxel and an SDF sampler;

a depth generation model comprising an implicit automatic encoder, a generation countermeasure network/interpolation operator/creative product solver; the implicit automatic encoder comprises an encoder 3DCNN and a Decoder IM-Decoder;

the sample vector storage library is used for storing hidden vector point sets, and hidden vectors in the storage library obtain a three-dimensional bionic product through an implicit decoder;

and the post-processing module is used for carrying out post-processing on the three-dimensional bionic product to obtain a three-dimensional bionic design.

In the three-dimensional bionic design system based on the generation countermeasure network, the implicit automatic encoder is used for fitting a symbol distance field of a three-dimensional model and performing training in a first stage; in order to obtain a product which meets the design requirements and has high fidelity, a proper hidden vector point set needs to be searched in a hidden space; the hidden vector point set is obtained through Gaussian sampling after an countermeasure network is generated through training, or is obtained through interpolation of hidden vector pairs obtained through an encoder through a three-dimensional model of a bionic object and a design target in a given training set, or is obtained through a creative product solver; and training the model to be trained in a second stage after the coding and decoding network training in one stage is completed to obtain a callable hidden vector generation model.

3DCNN encoder

Layer(s)	Convolution kernel size	Stride length	Activation function	Output shape
					Input voxels				(64,64,64,1)
conv3d	(4,4,4)	(2,2,2)	BNLReLU	(32,32,32,32)
					conv3d	(4,4,4)	(2,2,2)	BNLReLU	(16,16,16,64)
conv3d	(4,4,4)	(2,2,2)	BNLReLU	(8,8,8,128)
					conv3d	(4,4,4)	(2,2,2)	BNLReLU	(4,4,4,256)
conv3d	(4,4,4)		Sigmoid	(1,1,1,128)

Implicit decoder

Generating an countermeasure network

A generator

Layer(s)	Activation function	Output shape
			Input hidden vector		(128)
Full connection	LReLU	(2048)
			Full connection	LReLU	(2048)
Full connection	Sigmoid	(128)

Distinguishing device

Layer(s)	Activation function	Output shape
			Input hidden vector		(128)
Full connection	LReLU	(2048)
			Full connection	LReLU	(2048)
Full connection		(1)

According to the three-dimensional bionic design method based on the depth generation model, the depth generation model can be independently one of a generation countermeasure network, a creative product solver and an interpolation arithmetic unit, or can be simultaneously combined by a plurality of depth generation models.

The invention provides a three-dimensional bionic design method based on a generated countermeasure network, which comprises the following steps:

step 1: building a training set;

step 2: establishing a depth generation model, and training the depth generation model through the training set;

step 3: and generating a three-dimensional bionic product by using the depth generation model.

Step 4: and carrying out post-treatment on the three-dimensional bionic product in cooperation with a designer to obtain a three-dimensional bionic design.

The step 1 of the invention comprises the following steps: the method for constructing the identification conversion path from the 3D file format to the voxel model format specifically comprises the following steps:

step 1.1: the obj file of the three-dimensional grid is voxelized by binvox and converted into a binvox file;

step 1.2: the voxelized file is filled by flooding to obtain point value pairs, which are equivalent to discrete sampling of the isosurface;

step 1.3: learning a mapping of three-dimensional voxels to 128-dimensional hidden vectors with an encoder in an implicit auto-encoder, and a decoder learning a mapping of 128-dimensional hidden vectors plus three-dimensional voxel positions to a symbol distance field for the voxel; the symbol distance field obtains three-dimensional grid data through a marking cube algorithm; the symbolic distance field is thresholded to obtain voxel data.

The step 1 of the invention further comprises the following steps: and a training set enhancement step, which comprises the steps of enhancing the total amount of the training set from 113+6000 to more than 20k based on design target data enhancement and performance data enhancement.

The design goal based data enhancement includes: design-oriented data enhancement focuses on transferring design knowledge into the preparation of a dataset to meet design requirements. For example, the factors of direction are also taken into account during the design process by the designer. After evaluating existing biomimetic products and brain storms, the design of generative in multiple directions is considered to enable the invention to explore more possibilities. Considering the symmetry of a target product and the trade-off between diversity and efficiency, the invention selects to apply seven-direction rotation transformation to the bionic object data set to realize diversified mixed results.

The performance data based enhancements include: performance-oriented data enhancement aims to improve the performance of the depth generation model. The generator is affected by unbalanced training data, especially only hundreds of animal datasets. Non-rigid deformations, such as tensile transformations, are effective methods to alleviate this problem. The present invention stretches the training dataset in three orthogonal axes, ranging in extent from 0.25 to 2.

The step 1 of the invention further comprises the following steps: the resolution of the generated voxel file is improved from 16 to 32 to 64.

It comprises the following steps: the model generates a sign distance field (sdf), allowing voxels of arbitrary resolution to be generated by controlling the sdf sampling density. The degree of detail of the network fit depends on the voxel file resolution size used.

The training adopts progressive training, firstly training 100 rounds on a 16-3 resolution voxel file, then training 200 rounds on a 32-3 resolution voxel file, and finally training 1000 rounds on a 64-3 resolution voxel file. The training may be terminated during the training process to control the generation of the three-dimensional model of multi-scale detail.

In the invention, the depth generation model comprises an implicit automatic encoder and a hidden vector generation model; the implicit automatic encoder comprises an encoder 3DCNN and a Decoder IM-Decoder; the hidden vector generation model employs an countermeasure generation network.

Compared with a traditional Decoder based on a convolutional neural network, the IM-Decoder can enable the hidden vector space to be more continuous with the generated three-dimensional voxels, so that a more continuous interpolation effect is obtained.

The loss function of the IM-Decoder model depends on the sampling mode, where w _p The weight of the sampling mode is weighted and depends on the sparseness degree of the sampling points. The loss function is the weighted average squared error between the true value label and the predicted label for each point. Let S be the point set of the target shape, the loss function L (θ) be:

where θ is the network weight, f _θ For the mapping of the network(s),

the true value sampled at p-point for the actual symbol distance field.

In the implicit generation of challenge-networks (ens) framework, both the generator and the arbiter comprise three fully connected layers, with the first two layers of activation functions being the leak-relu and the last layer being the sigmoid function. The network is used for fitting the training set through the antagonism training to obtain the hidden vector point set through the encoder, and the final generator can output hidden vectors conforming to the training point set distribution in the training set hidden space through inputting standard normal distribution.

The generating countermeasure network comprises a generator and a discriminator, and the generating countermeasure network is used for fitting the distribution of the embedded points of the training set in the hidden space.

The invention establishes a generating countermeasure network, which comprises the following steps:

step 2.1: training an automatic encoder to learn a low-dimensional representation;

step 2.2: training a generation countermeasure network comprising a generator and a discriminator in a hidden vector space between an encoder and a decoder of the automatic encoder;

step 2.3: after the training set preparation is completed, two-stage training is performed on the automatic encoder and the countermeasure generation network respectively.

The two-stage training means: the first stage adopts progressive training to train an automatic encoder, so that the generated model has more details; and in the second stage, the antagonism generating network in the hidden vector space is trained by the generator and the discriminator, so that the result generated by the generator is more realistic and diverse.

After the automatic encoder and the countermeasure generation network are trained, an embedded mapping AE is obtained _S→Z An encoder comprising an automatic encoder, and generating a mapping G _Z→S It includes a generator that generates a countermeasure network and a decoder that is a self-encoder.

In the invention, the hidden vector generation model can also adopt an creative product solver. The creative product solver includes: the heuristic target function comprises a heuristic sampling algorithm and a hidden space optimization algorithm; the heuristic sampling algorithm samples the hidden space; and training a generating model through a weighted retraining method, reducing reasoning to complete black box optimization, and obtaining the hidden vector generator.

The invention establishes a creative product solver comprising the following steps:

step 2.2: deploying a creative product solver in a hidden vector space between an encoder and a decoder of the automatic encoder;

step 2.3: after the training set preparation is completed, two-stage training is respectively carried out on the automatic encoder and the creative generation solver.

The two-stage training means: the first stage adopts progressive training to train the self-encoder, so that the generated model has more details; and in the second stage, training the creative product solver, firstly sampling the hidden space by a heuristic sampling method, then calculating to obtain the fraction of sampling points by a heuristic objective function, and using a high-fraction sampling point set for training the kernel density estimation of the non-parameter estimation. While a high-scoring set of sample points serves as a set for subsequent evaluation.

Considering that the calculation cost of the previous step is large, the high-resolution sampling points are used as a part of the training set, and the generation model of the training document Sample-Efficient Optimization in the Latent Space of Deep Generative Models viaWeighted Retraining can finally complete black box optimization with less reasoning due to the effectiveness of the weighted retraining method, so as to obtain the hidden vector generator. The hidden vector generator may automatically input 32-dimensional training output that obeys a standard multivariate gaussian distribution to sample points with higher scores in the heuristic objective function, which may generate creative products that are considered to be of high quality after passing through the hidden Decoder IM-Decoder.

The mixed heuristic objective function f (x) consists of a connected branch count function and a binary classifier for distinguishing design targets from bionic objects. The connected branch count function makes the objective function f (x) black box, i.e., unknown to the general closed form or differential information.

Where delta is for discretization of continuityThreshold value of symbol distance field, N _r Is the resolution of the voxel. f (f) _θ (z, p) is a network prediction value at the point p of the symbol distance field obtained by decoding z, and represents the whole integer.

Embedded representations of a given bionic object and a design target, respectively

And->

M and N are the numbers of bionic objects and design target training set samples. Zd and zb are just named symbols, k being the sequence number taken from 1 to either M or N, ijk being essentially all used.

Training each perception classifier f _c (z)＝argmax _i P (y=i|z). The loss function is:

wherein:

is a hidden vector, Y represents a binary random variable, ideally f _c (z) ∈ {0,1}, the actual network values that are differentiable are continuous f _c (z)∈[0,1]. The object of the present invention is to encourage the production as a trade-off between biomimetic object and design target class. When the classifier uncertainty reaches a maximum, the persistence () should reach an optimal value, so there is defined as follows:

Percept(z)＝H(Z)＝-f _c (z)log f _c (z)-(1-f _c (z))log(1-f _c (z))

the hybrid heuristic objective function is ultimately expressed as:

S(z)＝Percept(z)+α ₁ N _cg (z)

wherein alpha is ₁ Is a coefficient.

A hidden space optimization algorithm (LSO) maps a set of hidden vector points to a high value region of the objective function S (z). Wherein the target functionThe number f is evaluated as few times as possible, and finally the evaluation point sequence is obtained

Algorithms are currently used (Austin Tripp, erik Daxberger, and Jose Miguel HernEz-Lobato 2020.Sampleefficient optimization in the latent space ofdeep generative models via weighted retrieval. Advances in Neural Information Processing Systems (2020)). The method finally obtains a hidden vector generator, and outputs the hidden vector in the high value area of the objective function S (z) through simple prior distribution sampling.

In order to evaluate the effect of the hidden space optimization algorithm, a heuristic sampling method and kernel density estimation are adopted to form a true value point set on the basis of Bayesian optimization.

The heuristic sampling method is a prime interval sampler. Firstly, the range of the high-quality sample hidden vector point set is reduced to be a convex hull of the hidden vector of the training data set. The selection of the middle points from the interpolation is an acceptable creative exploration method, and is suitable for the instant evaluation of the convex hull S. If the invention is applied to each pair

And->

Z in (b) ₁ And z ₂ Interpolation sampling is performed N times, and the middle m points are selected in each pair, and the sample set +.>

The following is shown:

wherein N represents the number of interpolation samples; lambda takes a value from 1 to m.

The amount of samples obtained in this way is very large, N _{interpolation} ＝m×|D _b ||D _d | a. The invention relates to a method for producing a fibre-reinforced plastic composite. In order to reduce the sampling amount, the invention discovers that sampling at intervals of a plurality of intervals can maintain a sample set

Algebraic structure of (a). Sample set->

Is a cyclic group of order m, which is sampled>

The subset of (c) is still a cyclic group of order m, while methods similar to uniform sampling etc. result in a subset that converges to this structure probabilistically.

The final sample set was estimated by nuclear density, with a bandwidth of 0.1. The samples may then be sampled by kernel density estimation to obtain the creative product.

In the invention, the hidden vector generation model can also adopt an interpolation arithmetic unit. The interpolation arithmetic unit has the structure that: given a bionic object hidden vector x1 and a design target hidden vector x2, the interpolated vector is expressed as x=γx ₁ +(1-γ)x ₂ Wherein gamma is E [0,1 ]]And obtaining a corresponding interpolation hidden vector by uniformly sampling gamma, and finally obtaining a corresponding product by a decoder. The implicit field obtained by mapping the interpolation hidden vector through the generator is used as a creative product, and the triangular mesh file is obtained through a marchngcube algorithm.

The method for establishing the depth generation model comprises the following steps:

step 2.2: training an interpolation operator in a hidden vector space between an encoder and a decoder of the automatic encoder;

step 2.3: after the training set preparation is completed, the automatic encoder is trained. The automatic encoder is trained by progressive training, so that the generated model has more details.

The interpolation arithmetic unit can directly obtain the closed solution through the training set data without training.

The invention also provides application of the method in three-dimensional representation, which specifically comprises the following steps:

step 1, learning three-dimensional shape embedding by an encoder, wherein the embedding refers to conversion into elements in a low-dimensional space;

step 2, searching in the low-dimensional space according to a specific rule to find a series of elements in the low-dimensional space fused with semantic features of different three-dimensional shapes; the specific rule comprises interpolation, generation of an antagonism network generation hidden vector and a creative product solver;

and 3, utilizing an implicit encoder and decoder to learn mapping, and remapping elements in the low-dimensional space back to the three-dimensional shape to obtain the three-dimensional shape of the object.

In the invention, the design in the three-dimensional bionic design refers to a design obtained by optimizing a product. Both "product" and "composition" correspond to the english term synhetic, specifically the product produced by the generator.

A generator. A generator is a neural network model for generating tasks. In the model training phase, a hidden representation space Z needs to be learned for the space X in which the multimedia content in a given training set is located, and the Z space is a low-dimensional euclidean space. In the model reasoning phase, the invention has x=f _θ (z), i.e. given a hidden spatial sampling point z, the generator can generate data of a corresponding modality.

The implicit self-encoder comprises a 3DCNN encoder and an implicit decoder which are used for fitting a symbol distance field of a three-dimensional model, reconstructing the symbol distance field, and representing the symbol distance field as a three-dimensional shape data set after the training of the first stage is completed. The loss function is the reconstruction error of the implicit field sample points. The specific network structure is shown in the table.

The middle hidden layer of the automatic encoder is a low-dimensional European space, and semantics in the Europe space are extracted through a deep learning method. The specific method comprises the following steps: let the low dimensional space be R ⁿ Is defined as a subset of the hidden vector space Z, a sample of a given design target domain

Sample of bionic object domain->

Assuming that the distribution of the design target domain D is p (D), the distribution of the bionic target domain B is p (B). First step training encoder AE _S→Z To learn from samples a subset Z of the distributions p (d) and p (b) embedded in a low-dimensional hidden vector space Z _B And Z is _D In, assume that their distribution is p (z _b ) And p (z) _d ). Second, according to a specific rule, the distribution p (z _b ) And p (z) _d ) Searching through the middle zone of the creative generation domain to obtain the potential representation space Z of the creative generation domain _I Assuming a distribution of p (z _i ) At p (z _i ) Sampling to obtain a series of samples->

Third, using the trained generator G for generating model _Z→S Sample->

Mapping into a three-dimensional shape domain, thereby obtaining a series of three-dimensional shapes with bionic creative heuristic values.

The invention adopts an interpolation method in the hidden space to synthesize a novel bionic design form. Related work for representational learning has shown that three-dimensional shapes can be transformed by operations on hidden vectors. The model trained by the invention can also generate two hidden vectors z by two different objects through the encoder ₁ And z ₂ Then the interpolated hidden vector z' =γz ₁ +(1-γ)z ₂ (0<γ<1) The generator is input to produce a three-dimensional shape. These three-dimensional shapes exhibit smooth transitions between objects as gamma increases.

An antagonism network is generated for mapping N (0, 1) normal distribution samples to new sample points fitting the training set hidden vector distribution.

Gaussian sampling is used to sample the sample points directly in hidden space N (0, 1). Tested effects are less than generating an antagonism network.

And selecting proper bionic objects and design targets from the voxel models in the input data set. By encoder AE _S→z Obtaining hidden vectors, interpolating between every two hidden vectors to obtain a large number of intermediate products, then carrying out intervention selection by a designer, and trying different angle combinations in the subsequent generation process to obtain a richer creative heuristic model.

Gaussian distribution in the hidden space of the latex-GAN model is also a method for generating novel bionic design morphology. And after the sampling is finished, a designer intervenes, a bionic object and a design target with ideal model learning effect are selected, and preparation is performed for generating a bionic product by subsequently calling a model. In order to make the generated results and designers more acceptable, the invention adopts the expanded data set and direction adjustment to generate three-dimensional shapes and morphological details with rich and various more morphologies.

Drawings

FIG. 1 is a schematic illustration of the process of the present invention.

FIG. 2 is a schematic diagram of a three-dimensional bionic design method based on a depth generation model according to the present invention.

FIG. 3 is a flow chart of a two-stage training method of the present invention.

Fig. 4 is a schematic diagram of the data representing method and the conversion mode between the four three-dimensional shapes according to the present invention.

FIG. 5 is a diagram of a model call overview of the present invention.

FIG. 6 is a hypothetical graph of a creative generation problem targeting a bionic design in an embodiment of the present invention.

Fig. 7 is a diagram of a data format of a result generated by an embodiment of the present invention.

FIG. 8 is a schematic diagram of the three-dimensional form and design heuristic points and the development direction of the design scheme of the typical product according to the embodiment of the invention.

FIG. 9 is a diagram of a design performed in part under the product heuristic of an embodiment of the present invention.

FIG. 10 is a diagram showing a comparison of a partial product form and an international excellent bionic child home chair design form according to an embodiment of the present invention.

FIG. 11 is a diagram showing a comparison of the design of the example of the present invention with the design of an international excellent simulated child home chair.

FIG. 12 is a schematic diagram of a partial creative product solver generating a result score according to an embodiment of the present invention.

FIG. 13 is a schematic diagram of the inventive product solver.

Detailed Description

The present invention will be described in further detail with reference to the following specific examples and drawings. The procedures, conditions, experimental methods, etc. for carrying out the present invention are common knowledge and common knowledge in the art, except for the following specific references, and the present invention is not particularly limited.

Example 1

By means of the potential self-encoder, only the distribution of training data can be fitted, and the fusion structure between the products cannot be necessarily obtained. Sometimes there is some degree of design abstraction to reconstruct a structure alone. In this embodiment, a three-dimensional bionic design is obtained by generating an countermeasure network, and the specific process is as follows:

generating the challenge network includes generating a challenge network for fitting a distribution of embedded points of the training set in the hidden space, and determining a challenge network.

In the training phase, the generator inputs 128-dimensional gaussian noise following the normal distribution of N (0, 1), and outputs a generated 128-dimensional vector. The discriminator inputs the generated vector and the embedded point set of the training set in the hidden space, and classifies whether the vector category belongs to the training set or the generator. The two parties perform countermeasure training, and after the final training is finished, the generator can input Gaussian noise and output hidden vectors which are close to the distribution of the embedded points of the training set in the hidden space.

In the call phase, 128-dimensional vectors sampled from the normal distribution of N (0, 1) are mapped to new sample points fitting the training set hidden vector distribution.

A generator

Distinguishing device

As shown in fig. 1, the potential self-encoder can only fit the distribution of training data, and a fusion structure between products is not necessarily obtained. Sometimes there is some degree of design abstraction to reconstruct a structure alone.

Example 2

In this embodiment, a three-dimensional bionic design is obtained by adopting an interpolation arithmetic unit, and the specific process is as follows:

according to the embodiment of the invention, a child with the age of 3-6 years is selected as a design object, a bionic home chair is used as a design object, a dog familiar to people in life is used as a bionic object, and an urban family room is used as an application scene. The designer participating in the design comprises 1 red dot product design prize acquirer, one product designer who takes the role of the child education institution, one product designer who has the child aged 5 years in one house, and the computer engineer is a team composed of a study student and a teacher of the university computer school of the east China.

Based on the work designer and the computer engineer, a bionic design resource library is built together, an open source model library of a chair is adopted in training, and a simple and decoration-free backrest chair is used for preparing training of a generation model; a bionic modeling library is constructed by selecting the models of dogs with obvious characteristics and consistent postures.

The three-dimensional deep learning model is trained for multiple times according to the set interpolation calculation, each product consists of three views and model files, and more than 1.5 ten thousand products are formed for calling, evaluating and screening after 9 times of training and calling. Designers and computer engineering make continuous corrections and advances from the following 5 aspects, so that the subsequent products have higher bionic design morphological value.

The first stage is a three-dimensional deep learning product calling and optimizing measure stage, as shown in the following table:

TABLE 1 three-dimensional deep learning product invocation and optimization measures

The deep learning product comprises a plane contour and a three-dimensional form file, the bionic form value evaluation of the product is divided into two aspects of bionic form creative value evaluation and model generation quality evaluation, a designer selects the product with higher bionic design value relatively quickly according to whether the product reflects important characteristics of an original bionic object or not, and the product has a standard morphological structure, a bionic object structure, the correspondence of a product structure and the like; the algorithm engineer evaluates the product from the aspect of the completion effect of the deep learning, finds out abnormal phenomena and reasons such as torsion, fracture, default and the like in the product, and observes and evaluates the influence of each interpolation, parameter adjustment and data set on the generation result. After the adjustment and screening of the multiple generation processes, approximately 5 types of products with high bionic morphology values are obtained, and the products can be further optimized into design schemes. Both the three-dimensional morphology and the planar profile of the product can be very inspired by the designer.

The second stage is to further optimize the 3D product with high bionic morphology value into a design scheme. The designer's effort at this stage is divided into two parts: 1, selecting a product with higher bionic design morphological value, and performing morphological optimization treatment to form a design scheme; 2 under the inspired of single or multiple products, the designer directly creates a secondary creative to form a new design scheme. The sensitivity of the designer to the bionic morphology, morphology processing skills and experience of the product design play a key role in the process, and the computer engineer needs to fully understand the selection criteria of the designer and assist the designer in completing the screening work and performing the conversion of the product format and model.

The designer combines the product structure, the functional design and the like to adjust and optimize the product model, so as to form a plurality of design schemes of the child home chairs.

In order to evaluate the value of the depth generation model in the aspect of three-dimensional bionic morphology creative generation, the invention compares the product morphology screened by a designer in the work with the morphology of a well-known bionic product design, as shown in fig. 10.

The comparison shows that the problems of the surface roughness, the partial form integrity and the like of the product are removed, the form of the product is very similar to the form designed by most of the existing bionic products, and the product accords with the bionic modeling rules of abstract bionic, intention bionic and the like; the product is also relatively close to the bionic object in aspects of gesture dynamics, animal typical characteristics and the like. The product has rich forms, and basically covers the change types of the existing creative forms of the bionic design.

In the embodiment of the invention, three product designers in total fuse own creatives to form 10 design schemes and a plurality of design creative sketches in a short time after observing the characteristics of the products, as shown in fig. 11. From the aspect of scheme distribution, a plurality of products are inspired by a designer, and a design scheme can be formed after morphological optimization; from the viewpoint of creative inspiring points of the scheme, a designer can obtain creative inspiring points from a single product, and can fuse creative points in more than 2 products to form a new creative form. This process demonstrates that the product is a creative heuristic that can be given to the designer and assists the designer in forming the design solution, as opposed to the design of an international designer, which is superior in morphological processing skills. The designer also gets a heuristic from the outline of the product, creating a design solution.

Example 3

As shown in fig. 12, from left to right, the creative product solver generates a result score from high to low from top to bottom, and it can be seen that the high-score products can often fuse morphological features of the bionic object and the design target. The products in the intermediate stage can only fit the distribution of training data, and a fused structure between the products is not necessarily obtained. Products at low stages often exhibit substantial breakage, affecting product functionality.

In this embodiment, a creative product solver is adopted to obtain a three-dimensional bionic design, and the specific process is as follows:

firstly, sampling the hidden space by a heuristic sampling method, then calculating to obtain the fraction of sampling points by a heuristic objective function, and using a high-fraction sampling point set for training the kernel density estimation of non-parameter estimation. While a high-scoring set of sample points serves as a set for subsequent evaluation.

Example 4

According to the three-dimensional bionic generation method based on the depth generation model, the depth generation model can be independently one of a generation countermeasure network, a creative generation solver and an interpolation arithmetic unit, or can be a combination of a plurality of depth generation models. When a plurality of models are adopted at the same time, the accuracy of the three-dimensional bionic design is improved.

The protection of the present invention is not limited to the above embodiments. Variations and advantages that would occur to one skilled in the art are included within the invention without departing from the spirit and scope of the inventive concept, and the scope of the invention is defined by the appended claims.

Claims

1. A design system of three-dimensional bionic resultant is characterized in that,

the system adopts a three-dimensional bionic design method based on a generated countermeasure network, and the method comprises the following steps:

step 1: building a training set;

step 2: establishing a depth generation model, and training the depth generation model through the training set; the depth generation model comprises an implicit automatic encoder and a hidden vector generation model; the implicit automatic encoder comprises an encoder 3DCNN and a Decoder IM-Decoder; the hidden vector generation model adopts an countermeasure generation network;

step 3: generating a three-dimensional bionic product by using the depth generation model;

the system comprises:

a depth generation model comprising an implicit automatic encoder, generating an antagonism network; the implicit automatic encoder comprises an encoder 3DCNN and a Decoder IM-Decoder;

the sample vector storage library is used for storing the hidden vector point set, and hidden vectors in the storage library are used for obtaining the three-dimensional bionic product through the implicit decoder.

2. The system for designing a three-dimensional biomimetic product of claim 1, further comprising: and the post-processing module is used for carrying out post-processing on the three-dimensional bionic product to obtain a three-dimensional bionic design.

3. The system for designing a three-dimensional biomimetic product according to claim 2, wherein the implicit automatic encoder is used for fitting a symbol distance field of a three-dimensional model for a first stage of training; in order to obtain a product which meets the design requirements and has high fidelity, a proper hidden vector point set needs to be searched in a hidden space; the hidden vector point set is obtained through Gaussian sampling after an countermeasure network is generated through training; and training the model to be trained in a second stage after the one-stage coding and decoding network training is completed to obtain a callable hidden vector generation model.

4. The system for designing a three-dimensional biomimetic product according to claim 1, wherein the method further comprises step 4: and carrying out post-treatment on the three-dimensional bionic product in cooperation with a designer to obtain a three-dimensional bionic design.

5. The system for designing a three-dimensional biomimetic product according to claim 1, wherein the step 1 comprises: the method for constructing the identification conversion path from the 3D file format to the voxel model format specifically comprises the following steps:

step 1.3: learning a mapping of three-dimensional voxels to 128-dimensional hidden vectors with a 3DCNN encoder in an implicit auto encoder, and a decoder learning a mapping of 128-dimensional hidden vectors plus three-dimensional voxel positions to a symbol distance field for the voxel; the symbol distance field obtains three-dimensional grid data through a marking cube algorithm; the symbol distance field obtains voxel data through thresholding;

and/or the number of the groups of groups,

the step 1 further includes: a training set enhancement step including design target data based enhancement and performance data based enhancement;

the design goal based data enhancement includes: applying seven-direction rotation transformation to the bionic object data set to realize diversified mixed results;

the performance data based enhancements include: stretching the training dataset in three orthogonal axes to a degree of from 0.25 to 2;

and/or the number of the groups of groups,

the step 1 further includes: enhancing the resolution of the generated voxel file, comprising: the model generates a symbol distance field, and voxels with any resolution are allowed to be generated by controlling the sampling density of the symbol distance field;

training is carried out progressively, wherein 100 rounds of training are firstly carried out on a 16-3 resolution voxel file, 200 rounds of training are carried out on a 32-3 resolution voxel file, and 1000 rounds of training are carried out on a 64-3 resolution voxel file;

the generation of a three-dimensional model of multi-scale detail is controlled by terminating training in a training process.

6. The system for designing three-dimensional biomimetic artifacts according to claim 1, wherein the generating countermeasure network comprises a generator and a discriminator, the generating countermeasure network being used to fit the distribution of embedded points of the training set in the hidden space;

and/or the number of the groups of groups,

the depth generation model is built by the following steps:

7. The system for designing a three-dimensional biomimetic product according to claim 6, wherein the two-stage training means: the first stage adopts progressive training to train an automatic encoder, so that the generated model has more details; and in the second stage, the antagonism generating network in the hidden vector space is trained by the generator and the discriminator, so that the result generated by the generator is more realistic and diverse.

8. The system for designing three-dimensional biomimetic products according to claim 6, wherein the embedded map AE is obtained after the automatic encoder and the countermeasure generation network are trained _S→Z An encoder comprising an automatic encoder, and generating a mapping G _Z→S Which includes a generator of the countermeasure generation network and a decoder of the self-encoder.

9. The system for designing a three-dimensional biomimetic product according to claim 1, wherein the loss function of the IM-Decoder model depends on the sampling mode; wherein the loss function is a weighted average squared error between the true value label and the predictive label for each point; let S be the point set of the target shape, the loss function L (θ) be:

wherein θ is a network weight, f _θ For the mapping of the network(s),

sampling true values of the actual symbol distance field at the p point; w (w) _p The weight of the sampling mode is weighted and depends on the sparseness degree of the sampling points.

10. The system for designing three-dimensional biomimetic products according to claim 1, wherein the generating countermeasure network comprises a generator and a discriminator, both of which comprise three fully connected layers, wherein the first two activating functions are leakage-relu, and the last layer is a sigmoid function; the generating countermeasure network is used for fitting the training set through countermeasure training to obtain the hidden vector point set through the encoder, and finally the generator outputs hidden vectors conforming to the training point set distribution in the training set hidden space through inputting standard normal distribution.