CN117315194B - Triangular mesh representation learning method for large aircraft appearance - Google Patents
Triangular mesh representation learning method for large aircraft appearance Download PDFInfo
- Publication number
- CN117315194B CN117315194B CN202311263784.2A CN202311263784A CN117315194B CN 117315194 B CN117315194 B CN 117315194B CN 202311263784 A CN202311263784 A CN 202311263784A CN 117315194 B CN117315194 B CN 117315194B
- Authority
- CN
- China
- Prior art keywords
- graph
- dual
- node
- feature
- original image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 29
- 230000009977 dual effect Effects 0.000 claims abstract description 60
- 239000013598 vector Substances 0.000 claims abstract description 35
- 238000010586 diagram Methods 0.000 claims abstract description 21
- 230000011218 segmentation Effects 0.000 claims abstract description 10
- 238000011176 pooling Methods 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 19
- 238000013507 mapping Methods 0.000 claims description 12
- 238000010606 normalization Methods 0.000 claims description 12
- 238000013528 artificial neural network Methods 0.000 claims description 10
- 230000009466 transformation Effects 0.000 claims description 10
- 238000005295 random walk Methods 0.000 claims description 9
- 230000002776 aggregation Effects 0.000 claims description 8
- 238000004220 aggregation Methods 0.000 claims description 8
- 238000012512 characterization method Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 238000005096 rolling process Methods 0.000 claims description 2
- 230000004931 aggregating effect Effects 0.000 description 2
- 239000012141 concentrate Substances 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000002860 competitive effect Effects 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/20—Finite element generation, e.g. wire-frame surface description, tesselation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/15—Vehicle, aircraft or watercraft design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G06N3/0455—Auto-encoder networks; Encoder-decoder networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Multimedia (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Medical Informatics (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Computer Hardware Design (AREA)
- Databases & Information Systems (AREA)
- Biodiversity & Conservation Biology (AREA)
- Computer Graphics (AREA)
- Pure & Applied Mathematics (AREA)
- Aviation & Aerospace Engineering (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Computational Mathematics (AREA)
- Automation & Control Theory (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a triangle mesh representation learning method for a large aircraft appearance, which comprises the following steps: s1: establishing an original image and a dual image on an original image based on triangular mesh data corresponding to the appearance of the large-scale aircraft; s2: RWPE position coding is applied to the original image, so that node feature vector embedding is obtained; s3: embedding the node feature vector into an input diagram transform module, and inputting the dual diagram into a diagram convolution module to obtain an original diagram feature diagram and a dual diagram feature diagram; s4: and (3) maximally pooling the original image feature map and the dual image feature map, and adopting MLP and softmax to realize the classification or segmentation of the triangular meshes of the large-scale airplane appearance. The invention can combine the advantages of the convolution network and the transducer architecture, furthest reserve local details and global dependency relationships, extract effective features with rich semantic information, fully exert the advantages of grid representation, and do not need artificial priori knowledge.
Description
Technical Field
The invention belongs to the technical field of aircraft appearance triangular mesh representation, and particularly relates to a large aircraft appearance-oriented triangular mesh representation learning method.
Background
In the design and manufacturing of large aircraft, it is important to accurately describe and analyze the shape of the aircraft. Triangular meshes are a commonly used graphical representation method that divides the surface of an object into small triangles and uses vertices and connected edges to describe the shape of the object. The complexity of the topology and geometry information contained in the triangle mesh representation increases the difficulty of characterization learning. Due to the non-uniformity and high dimension of the triangular mesh, the traditional machine learning method is difficult to directly apply to the characterization learning of the triangular mesh.
Deep learning is a branch of machine learning, and is applicable to learning and characterizing triangular meshes by learning and characterizing complex data patterns through a multi-layer neural network. At present, deep learning can complete downstream tasks such as classification, segmentation, shape generation, reconstruction and the like by learning the internal features of the triangular mesh of the airplane, but artificial priori knowledge is generally required, effective features with rich semantic information are difficult to extract, and the advantages of mesh representation are fully exerted.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a triangular grid representation learning method for the appearance of a large airplane, which can combine the advantages of a convolution network and a Transformer framework, respectively concentrate on local information and global information, fuse the two to be effectively combined, furthest reserve local details and global dependency, extract effective characteristics with rich semantic information, fully exert the advantages of grid representation, and do not need artificial priori knowledge.
In order to achieve the technical purpose, the invention adopts the following technical scheme:
A triangle mesh characterization learning method facing to the appearance of a large aircraft comprises the following steps:
S1: establishing an original image and a dual image on an original image based on triangular mesh data corresponding to the appearance of the large-scale aircraft;
S2: RWPE position coding is applied to the original image, so that node feature vector embedding is obtained;
S3: embedding the node feature vector into an input diagram transform module, and inputting the dual diagram into a diagram convolution module to obtain an original diagram feature diagram and a dual diagram feature diagram;
S4: and (3) maximally pooling the original image feature map and the dual image feature map, and adopting MLP and softmax to realize the classification or segmentation of the triangular meshes of the large-scale airplane appearance.
In order to optimize the technical scheme, the specific measures adopted further comprise:
in the above S1, each face of the mesh is used as a node, and an edge is connected between two adjacent faces, and each edge of the mesh is used as a node, and an edge connected between edges sharing the same vertex is used to connect two nodes.
The step S2 includes:
Step S21, selecting a node p on an original image, setting a k step of random walk, calculating probability of returning to the node p under the condition of k times of random walk, and obtaining a k-dimensional probability vector, wherein the initial position code of the network is obtained by embedding RWPE position code into a d-dimensional vector;
And S22, repeating the step S21 until all the position code vectors corresponding to all the nodes on the original image are completely calculated, and embedding RWPE position codes of all the nodes into the d-dimensional vector through linear mapping to obtain the initial position code of the node.
And S23, embedding the original characteristics of the nodes into d-dimensional vectors through linear mapping, and adding the d-dimensional vectors with initial position codes of the corresponding nodes to obtain node characteristic vector embedding.
The step S3 includes:
Step S31, the node characteristics and the dual graph characteristics are projected to hidden layer characteristics of d dimension through a linear mapping;
S32, projecting the position codes of the nodes to hidden layer features of d dimension through linear projection embedding, and adding the hidden layer features to the input node features;
Step S33, calculating the implicit attention score of each node by using a neural network parameter KQV matrix, multiplying the implicit attention score by a dual graph feature matrix to fuse side information, transmitting the output to a feedforward network FFN, and separating the feedforward network FFN by a residual connection and a normalization layer;
step S34, performing feature aggregation on dual graph features by using a main neighborhood aggregation graph neural network PNA, and eliminating linearity and normalization by using a ReLU activation function layer and a BatchNorm layer;
step S35, repeating the step S34 until the number of layers of the graph convolution network is set, splicing the characteristics of all the graph convolution layers together, and carrying out normalization operation on the characteristics of each batch by using BatchNorm;
step S36, performing linear transformation on the dual graph feature graph obtained in the step S34, and using the dual graph feature graph as an input of the step S33; performing linear transformation on the fused dual graph characteristics obtained in the step S33, and taking the dual graph characteristics as input of the step S34;
And S37, repeating the steps S31 to S36 until the number of the set convolution-conversion modules is reached, and outputting an original image characteristic diagram and a dual image characteristic diagram.
The step S33 includes:
step S331, calculating an attention score, when node i focuses on node j, calculating an intermediate attention score before softmax Then side information is injected for side (i, j) and the calculated/>, is improvedObtaining output/>, based on layer update equation of edgeAnd/>
Step S332, willAnd/>Is passed to the feed forward network and separated by the residual connection and normalization layer.
The layer update equation for the edge described above is as follows:
Wherein the method comprises the steps of
Wherein, the Q k,l,Kk,l is that,Representing a matrix of learnable parameters,/>Representing a dual graph feature matrix,/> Represents a linear layer parameter, k=1 to H, represents the number of attention headers, and ii represents the connection.
The step S332 specifically includes the following steps:
Wherein the method comprises the steps of Representing a weight matrix,/>And/>Represents the intermediate representation of, relative,
Wherein the method comprises the steps ofRepresenting a weight matrix,/>And/>Representing an intermediate representation.
The step S36 includes: adjusting feature mapping dimension by using a linear layer, and recombining edge features; when the dual-graph feature map obtained in the step S34 is subjected to linear transformation and then is used as the input of the step S33, the dual-graph feature map needs to be copied to be used as a feature of a reverse side; when the fused dual graph features obtained in the step S33 are subjected to linear transformation and then are used as the input of the step S34, the features of the edges and the reverse edges are aggregated; batchNorm and LeakyReLU are used to activate and regularize features.
The S4 includes:
step S41, calculating the final aggregated feature representation by using average pooling according to the original image feature map and the dual image feature map obtained in the step S3;
Step S42, selecting a corresponding module according to the downstream task: the classifying task uses MLP and softmax to calculate the class probability, and the dividing task uses MLP to calculate the class corresponding to each triangular mesh surface.
The invention has the following beneficial effects:
According to the invention, the triangular Mesh (Mesh) is characterized and learned by utilizing geometric knowledge and a transducer in the graph neural network, namely, a new network structure MeshGeoFormer is provided, and the geometric information in the Mesh is utilized to the greatest extent by establishing an original graph and a dual graph on the Mesh, so that the three-dimensional data is characterized, learned and processed, and further, the classification and segmentation task of the large-scale aircraft appearance data is realized; establishing an original image on original data, processing a feature image of the original image by using an image transducer, and carrying out global feature aggregation on the original image; establishing a dual graph on the original data, and carrying out local feature aggregation and geometric information extraction on the dual graph by using a graph neural network; meshGeoFormer adopts a double structure, can combine the advantages of a convolutional network and a Transformer architecture, respectively concentrate on local information and global information, namely, the original image and the dual image are processed at the same time and have specific information interaction, then the original image and the dual image are effectively combined, local details and global dependency relations are reserved to the greatest extent, effective characteristics with abundant semantic information can be extracted, the advantages of grid representation are fully exerted, and no artificial priori knowledge is needed; and finally, aggregating the extracted features by using corresponding modules according to the downstream tasks to obtain a final classification segmentation result. The proposed method achieves competitive experimental results in the representative tasks of mesh classification and mesh segmentation.
Drawings
FIG. 1 is a flow chart of a large aircraft outline oriented triangular mesh characterization learning method of the invention.
Fig. 2 is a general schematic of a neural network according to an embodiment of the present invention.
FIG. 3 is a schematic diagram of a portion of a graph roll-up network in accordance with an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Although the steps of the present invention are arranged by reference numerals, the order of the steps is not limited, and the relative order of the steps may be adjusted unless the order of the steps is explicitly stated or the execution of a step requires other steps as a basis. It is to be understood that the term "and/or" as used herein relates to and encompasses any and all possible combinations of one or more of the associated listed items.
As shown in fig. 1-3, the triangular mesh representation learning method for the appearance of the large aircraft is used for extracting key features and structures aiming at the appearance of the large aircraft and improving the performance and generalization capability of a model, and comprises the following steps:
S1: establishing an original image and a dual image on an original image based on triangular mesh data corresponding to the appearance of the large-scale aircraft;
The method comprises the steps of establishing an original image and a dual image on an original image on the premise of keeping consistent receptive fields for triangular grid data corresponding to the appearance of an airplane; the original graph takes each face of the grid as a node, edges are connected between two adjacent faces, each edge of the grid is taken as a node, and edges which are connected between edges sharing the same vertex are connected with the two nodes;
S2: RWPE position coding is applied to the original image, so that node feature vector embedding is obtained;
The RWPE position coding is applied to the step, low-complexity random walk is carried out on an original image, the access probability of the node to the node is considered, the unique node representation under the condition that each node has unique k-hop topological neighborhood is obtained, the node representation is embedded into d-dimensional vectors, and node feature vector embedding is obtained;
s3: the original features are input to the network as: embedding the node feature vector into an input graph transducer module; inputting the dual graph into a graph rolling module to obtain an original graph characteristic graph and a dual graph characteristic graph;
the node characteristic vector is embedded into the dual characteristic map, and the dual characteristic map is respectively passed through a map transform module and a map convolution module, wherein the map convolution module mainly comprises three GCN layers, and the input and output results are added by using the concept of residual connection. The graph Transformer module embeds node feature vectors into hidden layer features of the d dimension through linear mapping, adds the hidden layer features with position codes, calculates attention scores through interaction with dual feature graphs representing side information, and finally transmits output to a feedforward network;
s4: the feature map of the original map and the feature map of the dual map are pooled maximally, and the classification or segmentation of the triangular mesh of the appearance of the large-scale airplane is realized by adopting MLP and softmax;
The method comprises the steps of maximizing the original pool chart and the dual chart feature representation, using two MLPs with the same structure as a classifier, calculating the value of softmax of an output result to obtain probability distribution, and selecting the class with the highest probability as the classification or segmentation result.
In an embodiment, step S1 includes: each surface of the original grid M is regarded as a node, if two surfaces corresponding to two nodes are adjacent in the M, an edge is arranged between the two nodes, an original image is built, and the input characteristics are that: face center shop coordinates, vectors from three center points to the corner points of the triangular patch, and unit normal vectors; correspondingly, when each edge e epsilon M is regarded as a node, and a vertex is shared by the grid edges corresponding to the two nodes and M (or on the same plane), an edge connecting the two nodes is arranged, a dual graph is built, and the input characteristics are that: the dihedral angle between faces a and B, the ratio between the shared edge of a and B and the height of the two faces relative to the shared edge (edge height ratio), and the interior angle of the two faces, as illustrated in the left half of fig. 2.
In an embodiment, step S2 includes:
Step S21, selecting a node p on an original image, setting a k step of random walk, calculating probability of returning to the node p under the condition of k times of random walk, and obtaining a k-dimensional probability vector, wherein the initial position code of the network is obtained by embedding RWPE position code into a d-dimensional vector;
And S22, repeating the step S21 until all the position code vectors corresponding to all the nodes on the original image are completely calculated, and embedding RWPE position codes of all the nodes into the d-dimensional vector through linear mapping to obtain the initial position code of the node.
And S23, embedding the original characteristics of the nodes into d-dimensional vectors through linear mapping, and adding the d-dimensional vectors with initial position codes of the corresponding nodes to obtain node characteristic vector embedding.
Further, a node p is selected from the original graph, and RWPE is defined as k steps of random walk:
Where rw=ad -1 is a random walk operator.
The RWPE adopts a low-complexity random walking matrix using method, only the access probability of the node i to the RWPE is considered, and for a sufficiently large k, the RWPE provides unique node representation under the condition that each node has unique k-hop topology neighborhood.
The initial position coding of the network is obtained by embedding the laplace position coding or RWPE into a d-dimensional vector:
In an embodiment, step S3 includes:
Step S31, the node characteristics and the dual graph characteristics (equivalent to edge characteristics) are projected to the hidden layer characteristics of the d dimension through a linear mapping;
S32, projecting the position codes of the nodes to hidden layer features of d dimension through linear projection embedding, and adding the hidden layer features to the input node features;
Step S33, calculating the implicit attention score of each node by using a neural network parameter KQV matrix, multiplying the implicit attention score by a dual graph feature matrix to fuse side information, transmitting the output to a feedforward network FFN, and separating the feedforward network FFN by a residual connection and a normalization layer;
Step S34, performing feature aggregation on dual graph features by using a main neighborhood aggregation graph neural network (PNA), and then eliminating linear and normalization operations by using a ReLU activation function layer and a BatchNorm layer, as shown in FIG. 3;
step S35, repeating the step S34 until the number of layers of the graph convolution network is set, splicing the characteristics of all the graph convolution layers together, and carrying out normalization operation on the characteristics of each batch by using BatchNorm;
Step S36, performing linear transformation on the dual graph feature graph obtained in the step S34, and using the dual graph feature graph as an input of the step S33; and (3) performing linear transformation on the fused dual graph characteristics obtained in the step (S33) as an input of the step (S34).
And S37, repeating the steps S31 to S36 until the number of the set convolution-conversion modules is reached, and outputting an original image characteristic diagram and a dual image characteristic diagram.
In an embodiment, step S33 includes:
step S331, calculating an attention score, when node i focuses on node j, calculating an intermediate attention score before softmax Then inject the available side information for side (i, j) and increase the implicit attention score that has been calculatedFor edges, a layer update equation is defined as follows:
Wherein the method comprises the steps of
Wherein, the Q k,l,Kk,l is that,Representing a matrix of learnable parameters,/>Representing a dual graph feature matrix,/>Represents a linear layer parameter, k=1 to H, represents the number of attention headers, and ii represents the connection.
Step S332, for numerical stability, the output after indexing the entries inside softmax is limited to a value between-5 and +5. Then outputAnd/>Is transferred to the feed forward network and separated by the residual connection and normalization layer as follows:
Wherein the method comprises the steps of Representing a weight matrix,/>And/>Represents the intermediate representation of, relative,
Wherein the method comprises the steps ofRepresenting a weight matrix,/>And/>Representing an intermediate representation.
In an embodiment, step S36 includes:
Step S361, continuously coupling the local features and the global representation in an interactive manner using the FCU. First, a linear layer is used to adjust the feature map dimension:
eij=Aeij+a
Wherein the method comprises the steps of And reorganizes the edge features.
Step S362, the original image m= { V, E, F } is an undirected image, and the edge E and the reverse edge E' are stored during calculation, so that the dual image feature image is copied into a duplicate image feature image as a reverse edge when the feature image is transferred from the image convolution (i.e. the output of step 34) into the transducer (i.e. the input of step 33);
Step S363, aggregating the edge and reverse edge features when the edge features are transmitted from the transducer as dual graph features into the graph convolution, using simple average aggregated edge and reverse edge features:
where e i is an edge feature and e i' is a reverse edge feature;
step S364, uses BatchNorm and LeakyReLU to activate and regularize features.
In an embodiment, step S4 includes:
step S41, calculating the final aggregated feature representation by using average pooling according to the original image and the dual image feature image calculated in the step S37;
And S42, selecting a corresponding module according to the downstream task, calculating the class probability by using the MLP and the softmax by the classification task, and calculating the class corresponding to each triangular mesh surface by using the MLP by the segmentation task.
It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference sign in a claim should not be construed as limiting the claim concerned.
Furthermore, it should be understood that although the present disclosure describes embodiments, not every embodiment is provided with a separate embodiment, and that this description is provided for clarity only, and that the disclosure is not limited to the embodiments described in detail below, and that the embodiments described in the examples may be combined as appropriate to form other embodiments that will be apparent to those skilled in the art.
Claims (5)
1. The triangle mesh characterization learning method for the large aircraft appearance is characterized by comprising the following steps of:
S1: establishing an original image and a dual image on an original image based on triangular mesh data corresponding to the appearance of the large-scale aircraft;
S2: RWPE position coding is applied to the original image, so that node feature vector embedding is obtained;
s3: the original features are input to the network as: embedding the node feature vector into an input graph transducer module; inputting the dual graph into a graph rolling module to obtain an original graph characteristic graph and a dual graph characteristic graph;
s4: the feature map of the original map and the feature map of the dual map are pooled maximally, and the classification or segmentation of the triangular mesh of the appearance of the large-scale airplane is realized by adopting MLP and softmax;
The step S3 comprises the following steps:
Step S31, the node characteristics and the dual graph characteristics are projected to hidden layer characteristics of d dimension through a linear mapping;
S32, projecting the position codes of the nodes to hidden layer features of d dimension through linear projection embedding, and adding the hidden layer features to the input node features;
Step S33, calculating the implicit attention score of each node by using a neural network parameter KQV matrix, multiplying the implicit attention score by a dual graph feature matrix to fuse side information, transmitting the output to a feedforward network FFN, and separating the feedforward network FFN by a residual connection and a normalization layer;
step S34, performing feature aggregation on dual graph features by using a main neighborhood aggregation graph neural network PNA, and eliminating linearity and normalization by using a ReLU activation function layer and a BatchNorm layer;
step S35, repeating the step S34 until the number of layers of the graph convolution network is set, splicing the characteristics of all the graph convolution layers together, and carrying out normalization operation on the characteristics of each batch by using BatchNorm;
step S36, performing linear transformation on the dual graph feature graph obtained in the step S34, and using the dual graph feature graph as an input of the step S33; performing linear transformation on the fused dual graph characteristics obtained in the step S33, and taking the dual graph characteristics as input of the step S34;
step S37, repeating the steps S31 to S36 until the number of the set convolution-conversion modules is reached, and outputting an original image characteristic diagram and a dual image characteristic diagram;
The step S33 includes:
step S331, calculating an attention score, when node i focuses on node j, calculating an intermediate attention score before softmax Then side information is injected for side (i, j) and the calculated/>, is improvedObtaining output/>, based on layer update equation of edgeAnd/>
Step S332, willAnd/>Transmitting the data to a feedforward network, and separating the data through residual connection and a normalization layer;
the layer update equation for the edge is as follows:
Wherein the method comprises the steps of
Wherein the method comprises the steps ofRepresenting a matrix of learnable parameters,/>Representing a dual graph feature matrix,/>Represents a linear layer parameter, k=1 to H, represents the number of attention headers, and ii represents a connection;
The step S332 is specifically as follows:
Wherein the method comprises the steps of Representing a weight matrix,/>And/>Represents the intermediate representation of, relative,
Wherein the method comprises the steps ofRepresenting a weight matrix,/>And/>Representing an intermediate representation.
2. The large-aircraft-appearance-oriented triangular mesh representation learning method according to claim 1, wherein the original image is formed by taking each face of a mesh as a node, edge connection exists between two adjacent faces, each edge of the mesh is taken as a node, and two nodes are connected by edges connected between edges sharing the same vertex.
3. The large aircraft outline-oriented triangular mesh representation learning method according to claim 1, wherein S2 comprises:
Step S21, selecting a node p on an original image, setting a k step of random walk, and calculating probability of returning to the node p under the condition of k times of random walk to obtain a k-dimensional probability vector;
step S22, repeating the step S21 until all the position code vectors corresponding to all the nodes on the original image are calculated, embedding RWPE position codes of all the nodes into the d-dimensional vector through linear mapping, and obtaining the initial position codes of the nodes;
And S23, embedding the original characteristics of the nodes into d-dimensional vectors through linear mapping, and adding the d-dimensional vectors with initial position codes of the corresponding nodes to obtain node characteristic vector embedding.
4. The method for learning triangular mesh representation for large aircraft shapes according to claim 1, wherein the step S36 includes: adjusting feature mapping dimension by using a linear layer, and recombining edge features; when the dual-graph feature map obtained in the step S34 is subjected to linear transformation and then is used as the input of the step S33, the dual-graph feature map needs to be copied to be used as a feature of a reverse side; when the fused dual graph features obtained in the step S33 are subjected to linear transformation and then are used as the input of the step S34, the features of the edges and the reverse edges are aggregated; batchNorm and LeakyReLU are used to activate and regularize features.
5. The large aircraft outline-oriented triangular mesh representation learning method according to claim 1, wherein S4 comprises:
step S41, calculating the final aggregated feature representation by using average pooling according to the original image feature map and the dual image feature map obtained in the step S3;
Step S42, selecting a corresponding module according to the downstream task: the classifying task uses MLP and softmax to calculate the class probability, and the dividing task uses MLP to calculate the class corresponding to each triangular mesh surface.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311263784.2A CN117315194B (en) | 2023-09-27 | 2023-09-27 | Triangular mesh representation learning method for large aircraft appearance |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311263784.2A CN117315194B (en) | 2023-09-27 | 2023-09-27 | Triangular mesh representation learning method for large aircraft appearance |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117315194A CN117315194A (en) | 2023-12-29 |
CN117315194B true CN117315194B (en) | 2024-05-28 |
Family
ID=89259824
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202311263784.2A Active CN117315194B (en) | 2023-09-27 | 2023-09-27 | Triangular mesh representation learning method for large aircraft appearance |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117315194B (en) |
Citations (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6366279B1 (en) * | 1998-12-29 | 2002-04-02 | Intel Corporation | Triangle mesh compression |
CN105303614A (en) * | 2015-10-12 | 2016-02-03 | 中国人民解放军空军航空大学 | Improved QEM three-dimensional model simplifying method |
CN109726509A (en) * | 2019-01-21 | 2019-05-07 | 南京航空航天大学 | A kind of part geometry feature representation model and construction method towards aircraft assembly |
CN110555285A (en) * | 2019-09-12 | 2019-12-10 | 北京工商大学 | full-automatic generation method for complex aircraft surface structure grid |
CN112200733A (en) * | 2020-09-09 | 2021-01-08 | 浙江大学 | Grid denoising method based on graph convolution network |
CN112348963A (en) * | 2020-12-01 | 2021-02-09 | 南京北冶机电设备有限公司 | Efficient FDTD grid subdivision method for unstructured complex target |
KR102272101B1 (en) * | 2021-03-24 | 2021-07-01 | 인천대학교 산학협력단 | Graphic processing apparatus capable of generating a mesh structure through learning of a pointer network for delaunay triangulation and the operating method thereof |
CN113593033A (en) * | 2021-06-03 | 2021-11-02 | 清华大学 | Three-dimensional model feature extraction method based on grid subdivision structure |
CN114419275A (en) * | 2021-12-13 | 2022-04-29 | 深圳先进技术研究院 | Method for denoising triangular mesh based on dual-graph neural network |
CN114741507A (en) * | 2022-03-25 | 2022-07-12 | 西北大学 | Method for establishing and classifying quotation network classification model of graph convolution network based on Transformer |
CN114972687A (en) * | 2022-07-21 | 2022-08-30 | 中汽研(天津)汽车工程研究院有限公司 | Mesh adjusting method based on elimination of triangular mesh pairs |
CN115079116A (en) * | 2022-04-14 | 2022-09-20 | 杭州电子科技大学 | Radar target identification method based on Transformer and time convolution network |
CN115374542A (en) * | 2022-10-24 | 2022-11-22 | 中汽研(天津)汽车工程研究院有限公司 | Mesh adjusting method, device and storage medium based on triangular mesh position |
CN116681895A (en) * | 2023-06-15 | 2023-09-01 | 南京航空航天大学 | Method, system, equipment and medium for segmenting airplane grid model component |
-
2023
- 2023-09-27 CN CN202311263784.2A patent/CN117315194B/en active Active
Patent Citations (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6366279B1 (en) * | 1998-12-29 | 2002-04-02 | Intel Corporation | Triangle mesh compression |
CN105303614A (en) * | 2015-10-12 | 2016-02-03 | 中国人民解放军空军航空大学 | Improved QEM three-dimensional model simplifying method |
CN109726509A (en) * | 2019-01-21 | 2019-05-07 | 南京航空航天大学 | A kind of part geometry feature representation model and construction method towards aircraft assembly |
CN110555285A (en) * | 2019-09-12 | 2019-12-10 | 北京工商大学 | full-automatic generation method for complex aircraft surface structure grid |
CN112200733A (en) * | 2020-09-09 | 2021-01-08 | 浙江大学 | Grid denoising method based on graph convolution network |
CN112348963A (en) * | 2020-12-01 | 2021-02-09 | 南京北冶机电设备有限公司 | Efficient FDTD grid subdivision method for unstructured complex target |
KR102272101B1 (en) * | 2021-03-24 | 2021-07-01 | 인천대학교 산학협력단 | Graphic processing apparatus capable of generating a mesh structure through learning of a pointer network for delaunay triangulation and the operating method thereof |
CN113593033A (en) * | 2021-06-03 | 2021-11-02 | 清华大学 | Three-dimensional model feature extraction method based on grid subdivision structure |
CN114419275A (en) * | 2021-12-13 | 2022-04-29 | 深圳先进技术研究院 | Method for denoising triangular mesh based on dual-graph neural network |
CN114741507A (en) * | 2022-03-25 | 2022-07-12 | 西北大学 | Method for establishing and classifying quotation network classification model of graph convolution network based on Transformer |
CN115079116A (en) * | 2022-04-14 | 2022-09-20 | 杭州电子科技大学 | Radar target identification method based on Transformer and time convolution network |
CN114972687A (en) * | 2022-07-21 | 2022-08-30 | 中汽研(天津)汽车工程研究院有限公司 | Mesh adjusting method based on elimination of triangular mesh pairs |
CN115374542A (en) * | 2022-10-24 | 2022-11-22 | 中汽研(天津)汽车工程研究院有限公司 | Mesh adjusting method, device and storage medium based on triangular mesh position |
CN116681895A (en) * | 2023-06-15 | 2023-09-01 | 南京航空航天大学 | Method, system, equipment and medium for segmenting airplane grid model component |
Non-Patent Citations (2)
Title |
---|
一种基于对偶图的三维表面模型编码方法;王永波;盛业华;闾国年;;中国矿业大学学报;20100315(02);第259-264页 * |
飞机外形件三角网格模型光滑B样条曲面重建;王宏涛;孙秀慧;周来水;刘胜兰;安鲁陵;;南京航空航天大学学报;20070615(03);第323-328页 * |
Also Published As
Publication number | Publication date |
---|---|
CN117315194A (en) | 2023-12-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Shaham et al. | Singan: Learning a generative model from a single natural image | |
CN108038906B (en) | Three-dimensional quadrilateral mesh model reconstruction method based on image | |
CN112257597B (en) | Semantic segmentation method for point cloud data | |
CN105930382A (en) | Method for searching for 3D model with 2D pictures | |
CN112200266B (en) | Network training method and device based on graph structure data and node classification method | |
Zhang et al. | Point cloud completion via skeleton-detail transformer | |
CN112862949B (en) | Object 3D shape reconstruction method based on multiple views | |
CN112686816A (en) | Image completion method based on content attention mechanism and mask code prior | |
CN112967178A (en) | Image conversion method, device, equipment and storage medium | |
CN113269848A (en) | Remote sensing image reconstruction system based on attention mechanism | |
CN116187200A (en) | Lightweight three-dimensional CAD model classification and retrieval method based on graph convolution network | |
CN117315194B (en) | Triangular mesh representation learning method for large aircraft appearance | |
CN101930452A (en) | Intelligent computation-based multi-granularity image retrieval method | |
CN117115404A (en) | Method, device, computer equipment and storage medium for three-dimensional virtual scene adjustment | |
CN112861862A (en) | Unsupervised segmentation method for terracotta warriors point cloud based on combination of region growing and deep learning | |
CN114565774B (en) | 3D (three-dimensional) graph volume integration class method based on local geometry and global structure joint learning | |
CN116129118A (en) | Urban scene laser LiDAR point cloud semantic segmentation method based on graph convolution | |
CN113610711B (en) | Single-image-guided three-dimensional surface reconstruction method and device | |
Luanyuan et al. | MGNet: Learning Correspondences via Multiple Graphs | |
Wang et al. | A geometry feature aggregation method for point cloud classification and segmentation | |
Li et al. | SAPCGAN: Self-attention based generative adversarial network for point clouds | |
CN113688715A (en) | Facial expression recognition method and system | |
Li et al. | Efficient and accurate object detection for 3D point clouds in intelligent visual internet of things | |
Zhou et al. | Sparsemae: Sparse training meets masked autoencoders | |
Zhao et al. | The Advisable Technology of Key-Point Detection and Expression Recognition for an Intelligent Class System |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |