CAD Model Classification Method Based on Feature-Level Graph Descriptors and Graph Convolutional Networks

Abstract

The present invention discloses a CAD model classification method based on feature-level graph descriptors and graph convolutional networks. The present invention extracts the dependencies between the features of a CAD model, converts them into a feature dependency graph, further transforms the array features, refines the structure, and generates feature-level graph descriptors to characterize three-dimensional CAD models. The model classification method based on feature-level graph descriptors and graph convolutional networks proposed by the present invention introduces a skip connection network and a hierarchical pooling mechanism based on edge contraction to construct a classification model of the graph convolutional network. The pooling mechanism based on edge contraction focuses on local key structures by gradually aggregating node information in the graph, and the skip connection network aggregates multiple intermediate graph representations to strengthen graph representation learning, so as to improve the classification effect. The present invention realizes accurate classification of CAD models, and the classification accuracy is higher than that of other existing three-dimensional model classification methods with better performance, and promotes the application of structured description and graph convolutional networks in the CAD model classification problem.

Description

Technical Field

[0001] The present invention belongs to the technical fields of CAD model reuse and deep learning, and particularly relates to a graph descriptor based on the feature structure of a CAD model, and a model classification method based on a graph convolutional network. Background Art

[0002] In recent years, with the development of computer graphics and 3D reconstruction technologies, 3D models have been widely used in fields such as CAD, VR / AR, etc. In the process of product design and development, approximately 80% of products are directly obtained by using existing models or making minor modifications to existing models. Therefore, the effective reuse of existing 3D CAD models has become the key to part standardization, optimization management, cost savings in development, and product quality improvement. Traditional model classification methods mainly include methods based on visual similarity and methods based on semantic and functional descriptions. The former refers to classifying by extracting the shape features of 3D models and analyzing the similarity between models; the latter refers to constructing corresponding descriptors based on the engineering semantics of the model feature structure or the functional description of the model design to distinguish different categories of CAD models. Models of the same category of CAD models have similarity in appearance, which is also a direct basis for distinguishing CAD models. At the same time, CAD models generally have specific design, manufacturing, or application context environments in the engineering field. Therefore, various semantic information such as manufacturing and processing features, attributes, and functions in 3D CAD models can be used as important bases for CAD model classification.

[0003] However, due to the simplification and imprecision of feature extraction, it is difficult to improve the classification accuracy of some solutions. When facing large-scale data samples, the similarity comparison method and the complexity of descriptors consume a large amount of computing resources and storage resources, limiting the improvement of the method's accuracy. With the development of deep learning, accurate and effective classification and recognition of CAD models based on deep learning technology have become one of the hot and difficult issues in current domestic and international research. With the rapid development of deep learning in the field of image processing, researchers have begun to use multi-view representations of 3D models and obtain high-order descriptors through convolutional neural networks for classification. At the same time, there are also studies on the original 3D models, using voxels or point clouds to represent the models and constructing corresponding feature extraction networks to extract high-order global features of the 3D models. Thus, the classification problem of CAD models is transformed into the classification problem of corresponding descriptors due to different 3D representations. Since the commonly studied multi-view, voxel, point cloud, etc. descriptions are based on visual shapes and cannot reflect the unique structural semantics and functional descriptions of CAD models, the classification accuracy of previous methods will be affected when facing models with complex features and detailed differences. Therefore, constructing descriptors with engineering semantics to represent CAD models is of great significance. In many studies on such 3D representations, the face-edge structure of CAD models or the functional descriptions of models are extracted to characterize the models. However, these descriptors have problems such as single characterization information and low description accuracy, making it difficult to effectively express the complex information of CAD models and unable to extract the key feature information therein to improve the model classification accuracy.

[0004] In addition, conventional convolutional neural networks are only oriented to data with raster structures and cannot effectively process the structural information in descriptors. CAD models often have clear structural divisions, and the detailed differences between some models exist between corresponding structures. Therefore, in order to reasonably process the structured information and use the feature differences between structures to distinguish model categories, graph convolutional networks are used in the CAD model classification method, aiming to effectively utilize the connections between structures through graph convolutional networks, strengthen the extraction of local key features, and improve the accuracy of model classification. Summary of the Invention

[0005] The present invention aims at the deficiencies of the prior art and proposes a method for extracting feature-level graph descriptors based on the feature structure of CAD models, as well as a model classification method based on feature-level graph descriptors and graph convolutional networks.

[0006] The feature-level graph descriptor extraction method based on the feature structure of the CAD model proposed by the present invention is oriented to the feature-based CAD model file, extracts the dependency relationship between the features of the CAD model, converts it into a feature dependency graph, and further transforms the array features and refines the structure to generate a feature-level graph descriptor to characterize the 3D CAD model. At the same time, the model classification method based on the feature-level graph descriptor and the graph convolutional network proposed by the present invention introduces a skip connection network and a hierarchical pooling mechanism based on edge contraction to construct a classification model of the graph convolutional network. The pooling mechanism based on edge contraction focuses on the local key structure by gradually aggregating the node information in the graph, and the skip connection network aggregates multiple intermediate graph representations to strengthen the graph representation learning, so as to improve the classification effect. The present invention can accurately classify CAD models, and the classification accuracy is higher than that of other existing 3D model classification methods with better performance. At the same time, it further promotes the application of structured description and graph convolutional network in the CAD model classification problem.

[0007] The method of the present invention specifically includes the following steps:

[0008] Step 1: Obtain a feature-oriented model file as the basic data set;

[0009] Step 2: Convert the.sldprt file of the CAD model in the data set into a TU-format feature-level graph descriptor. The specific operations are as follows:

[0010] Step 2-1: Read the feature tree of the CAD model in the.sldprt model file, extract the features of the specified feature interfaces, where the feature interfaces include ExtrudeFeatureData, LoftFeatureData, SweepFeatureData, RevolveFeatureData, LinearPatternFeatureData, CircularPatternFeatureData, IMirrorPatternFeatureData, ISimpleFilletFeatureData, IVariableFilletFeatureData, IChamferFeatureData, WizardHoleFeatureData, SimpleHoleFeatureData, ShellFeatureData, ThreadFeatureData, and then extract the parent-child dependency relationship in the feature attributes. For the extracted features, construct a feature dependency graph G with features as nodes and the parent-child dependency relationship of features as edges fea , as the basis of the feature-level graph descriptor;

[0011] Step 2-2: Design the attribute vector of the feature-level graph descriptor node, read the dependency relationship and interface parameters of the feature, and define which attributes the descriptor node has accordingly.

[0012] Read the sub-dependency relationships of the features in the feature dependency graph G fea Among them, read the first sketch in sequence as the sketch for generating the feature. If there is no sketch in the sub-dependency relationship, set the corresponding feature sketch to none. Extract the various parameters of the sketch line segment according to the predefined attributes, and analyze the features of the line segment to obtain the attribute vector of the sketch line segment. Combine the attribute vectors of the sketches in this way. If there is no sketch, set the corresponding dimension to E.

[0013] Preprocess the model dataset. Obtain all features according to the method of capturing features in Step 2-1. Read all the parameters of the feature corresponding to the feature interface and convert them into quantifiable numerical values. Extract the numerical values of all parameters for the dataset, construct the numerical distribution for the same parameters. If the distribution of the same numerical values for the same parameter is higher than 90% of the overall, it is determined that the parameter does not have geometric expressiveness and is screened out. At the same time, perform weighted combination on the parameters of the same type, x d = w1x1 + w2x2 + … + w n x n where x n represents the nth parameter, and w n represents the weighting coefficient of the nth parameter; to compress the dimension of the parameters, finally design the attribute vector for each feature according to the feature type divided by the feature interface. For the attribute vectors with inconsistent dimensions, assign the value E to the undefined dimensions.

[0014] Step 2-3: Perform the "array relationship conversion" on the array features in the feature dependency graph G fea Among them, the array feature is generated by array replication of other features. Therefore, the array feature contains the information of multiple features and itself represents multiple structures of the CAD model. Expand the array feature into multiple source features of its array replication and connect them to G fea to form the feature-level graph descriptor G fea-level ;

[0015] Step 3: Combine the skip connection network and the hierarchical pooling mechanism based on edge contraction to construct a graph convolutional classification network for the CAD model to realize the classification of the CAD model. The specific operations are as follows:

[0016] The CAD model classification network uses the GraphSAGE graph convolutional network as the backbone network, and incorporates a skip connection network and an edge contraction pooling mechanism. For CAD models of constructive geometry, this network is denoted as CSG-GraphNet. CSG-GraphNet consists of a total of 5 convolutional parts, conv1 to conv5, and hierarchical pooling parts pool1 and pool2 are added after its conv2 and conv4 convolutional parts; a residual unit JK-Net of the skip connection network is added to aggregate the network layer outputs of conv1 to conv5 and pool1 to pool2 to strengthen the learning of key features, enrich the structural information, and enhance the representation of global features.

[0017] The GraphSAGE graph convolutional network realizes node message passing based on the spatial domain, and proposes a neighbor sampling mechanism to enable the graph convolutional network to adapt to graphs of different sizes. When performing convolution, neighbor nodes are sampled, and the number of sampled nodes is controlled by setting the sampling magnification Si. For example, when S1 = 3 and S2 = 2, 3 nodes are sampled within the neighborhood range with a hop count of 1 outward from the central node. On this basis, in the neighborhood with a hop count of 2, each sampled node continues to sample 2 nodes outward, thereby controlling the sampling range and the number of nodes.

[0018] Step 3-1: A hierarchical pooling module based on edge contraction is added to the classification model. In order to effectively capture key features and protect the structural information from being lost as much as possible during the pooling process, the edge contraction pooling follows the edge relationship between nodes to achieve pooling.

[0019] Step 3-2: A skip connection network module is added to the classification model to aggregate intermediate graph representations at different levels, enrich the graph representation information at each level, and further improve the classification accuracy.

[0020] The operation of the skip connection network module is as follows: First, taking the intermediate graph representation output by each network layer as input, respectively passing through a global average readout layer to read the graph structure as a canonical feature matrix. Then, the feature matrices output by all readout layers are concatenated and merged into a multi-level feature matrix to better represent the hierarchical changes of the graph.

[0021] The key details of the feature-level graph descriptor are emphasized through the hierarchical pooling module based on edge contraction and the skip connection network module. At the same time, the hierarchical changes of the descriptor are enriched, the learning of the global graph embedding is strengthened, and the effect of enhancing the CAD model classification accuracy is achieved.

[0022] Preferably, the operation of obtaining the feature-oriented model file as the basic data set is as follows:

[0023] Obtain model files from public model websites and partial model libraries of mechanical manufacturing enterprises. If the above models already contain CAD model.sldprt files, they can be used as a preparation; screen the models to be used. Design scripts using the SOLIDWORKS API to screen out models without a reasonable feature tree, and then further manually screen the remaining models. At the same time, use the Toolbox tool of SOLIDWORKS to randomly generate models within the category range to supplement the existing model categories. Finally, obtain a dataset of.sldprt model files with a feature structure.

[0024] Preferably, a hierarchical pooling module based on edge contraction is added to the classification model. Edge contraction pooling realizes pooling by following the edge relationship between nodes, specifically: calculate scores for edges according to the nodes at both ends of the edges in the graph, contract the edges in the graph in descending order of scores, aggregate the nodes at both ends of the edges in the graph, and at the same time retain the original edge relationship between the two end nodes. Control the pooling node ratio at 0.5 during the contraction process, and finally output the pooled graph.

[0025] The beneficial results of the present invention:

[0026] 1. This technology proposes a method for converting a CAD model based on constructive geometry into a feature-level graph descriptor. This method utilizes the feature structure of the CAD model to construct a structured three-dimensional model representation, accurately representing the geometric attributes of different structures of the CAD model and the dependency relationships between structures, making up for the deficiencies of other three-dimensional representations in terms of structural accuracy and engineering semantic expression ability.

[0027] 2. This technology proposes a classification model based on feature-level graph descriptors and graph convolutional networks to achieve efficient and accurate classification of CAD models. The graph convolutional network uses neighbor sampling for node message passing, improving the classification accuracy while enabling the network to perform graph embedding tasks on graphs of different scales. The edge contraction pooling mechanism can effectively extract the features of key structures in the graph, and the skip knowledge network can enrich the graph representation information. Combining the two further improves the classification accuracy and efficiency. Brief Description of the Drawings

[0028] Figure 1 is the overall process of the CAD model classification method based on feature-level graph descriptors and graph convolutional networks;

[0029] Figure 2 is the category concept map of the self-constructed 18-class CAD model dataset;

[0030] Figure 3 is the relevant information on the sketch part attributes design of the feature-level graph descriptor node attribute vector;

[0031] Figure 4Information related to the attribute design of the feature interface part of the feature-level graph descriptor node attribute vector;

[0032] Figure 5 For the array relationalization algorithm process.

[0033] Figure 6 For the overall architecture diagram of the classification model based on the feature-level graph descriptor and the graph convolutional network Specific implementation manners

[0034] The present invention proposes a method for generating a feature-level graph descriptor based on the feature structure of a CAD model, and at the same time proposes a model classification method based on the feature-level graph descriptor and the graph convolutional network. Among them, the feature-level graph descriptor generation part uses the.sldprt model file of the CAD model to convert the CAD model into a feature-level graph descriptor with features as nodes, which accurately represents the attributes and associations of the structure, making up for the defects of the conventional three-dimensional representation with single characterization information and low description accuracy; at the same time, a message passing mechanism for neighbor sampling, a hierarchical pooling mechanism for qualified edge contraction, and a skip connection network are introduced into the graph convolutional network to construct a CAD model classification network. Neighbor sampling controls the neighborhood range and the number of nodes of the convolution, adapting to graph descriptors of different scales. The pooling mechanism of edge contraction realizes the gradual aggregation of graph information without loss of structural information to achieve graph embedding. The skip connection network enriches the graph representations at each level, further improving the classification accuracy of the CAD model.

[0035] The following combines the accompanying drawings and a self-constructed 18-class CAD model data set to describe the present invention in detail. The overall process of the present invention is as shown in the accompanying Figure 1 drawing, and the specific steps are as follows:

[0036] Step 1, Obtaining a CAD model file.

[0037] Step 2, Performing generation processing on the CAD model to obtain a feature-level graph descriptor.

[0038] Step 3, Constructing a classification model based on the feature-level graph descriptor and the graph convolutional network

[0039] Step 4, Classifying the CAD model by using the classification model based on the feature-level graph descriptor and the graph convolutional network trained in Step 3.

[0040] Further, in step 1, the present invention obtains the.sldprt model files from the public model website and part of the model libraries of mechanical manufacturing enterprises, and at the same time generates models of specified categories by using the Toolbox toolkit of SOLIDWORKS. Among the.sldprt files obtained from the public model website and the model libraries of mechanical manufacturing enterprises, models without a reasonable feature tree need to be screened out, and files with only a single Import feature in the CAD model feature tree need to be screened out. The CAD models generated by the Toolbox toolkit are batch-generated according to the templates of specified categories by randomly configuring parameters. There are a total of 6,922 models, belonging to 18 conventional categories: spur gear, flat gasket, steel, hex nut, screw, bearing, flange, bolt, grooved pin, stud, key, etc. The external shapes of partial models of all categories are as shown in Figure 2 as follows.

[0041] Further, in step 2, a feature-level graph descriptor of the converted CAD model is generated, and the generation process is as shown in Figure 3 as follows. The.sldprt file obtained in step 1 is converted into a feature-level graph descriptor in TUDataset format. The feature tree of the CAD model file describes the compositional dependencies of the various feature structures of the model based on the modeling logic of constructive geometry. Therefore, the present invention selects to construct a feature dependency graph to represent the CAD model structure. Considering that features have a large number of complex parameters and it is difficult to effectively express the structural geometric properties, the parameters are dimensionally compressed, and an attribute vector of the feature is designed. There is a type of array feature in the features, which is manifested as an array copy of multiple source features. Considering the need to accurately represent the structures of various parts of the CAD model by the feature-level graph descriptor, the present invention designs a conversion method of "array relationalization" to convert the array features. Based on the above analysis, the specific generation steps of the feature-level graph descriptor are as follows:

[0042] Step 2-1: Read the feature tree in the.sldprt model file, extract the features of the specified feature interface, and then extract the parent-child dependency relationships of the features to construct a feature dependency graph with features as nodes and feature dependency relationships as edges.

[0043] Step 2-2: Design the attribute vector of the feature-level graph descriptor, extract the parameters of the corresponding sketch and feature interface of the feature, and after dimensional compression, generate an attribute vector with a dimension of 16 to describe the geometric attributes of the descriptor node. Partial attributes of the sketch are as shown in Figure 3 as follows, and partial attributes of the feature are as shown in Figure 4 as follows.

[0044] Step 2-3: Perform a conversion of "array relationalization" on the feature dependency graph, expand the array features into multiple source features of their array copies, and reasonably connect them to the feature dependency graph according to the dependency relationships of the source features, and finally generate the feature-level graph descriptor. The "array relationalization" algorithm is as shown inFigure 5 as shown

[0045] Furthermore, in step 3, the present invention constructs a classification model based on feature-level graph descriptors and graph convolutional networks, which integrates edge contraction pooling and skip connection networks, denoted as CSG-GraphNet. The overall network architecture of CSG-GraphNet is as shown in the appendix Figure 6 as shown. GSG-GraphNet consists of a total of 5 convolutional layers (conv1 to conv5). At the same time, edge contraction pooling layers pool1 and pool2 are connected behind conv2 and conv4 respectively. Pool1 and pool2 are connected in front of conv3 and conv5 respectively to ensure that the edge contraction pooling mechanism gradually aggregates and fully learns the graph representation. A skip connection network JK-Net is set to connect the outputs of all network layers to enrich multi-level graph characterization information.

[0046] Among them, the convolutional part (conv1 to conv5): is used to perform convolutional operations on the input or the graph representation of the previous layer. Among these, conv1 receives the feature-level graph descriptor, and conv2 to conv5 receive the graph representation of the previous layer.

[0047] The aggregation formula for the convolutional operation is as follows:

[0048]

[0049] where σ represents the ReLU function, and N(v i ) represents the neighbor nodes sampled when node v i is the central node.

[0050] Edge contraction pooling mechanism: The present invention adds hierarchical pooling modules pool1 and pool2 based on edge contraction between conv2 and conv3, and between conv4 and conv5 respectively. They receive the graph representation and perform hierarchical pooling to output the aggregated graph representation. The pooling module based on edge contraction focuses on the features of the key structure, calculates the score s ij = softmax j (w T [h i ||h i + b), and selects the two nodes with the highest scores that have not been selected for contraction in descending order, merges the nodes without repetition, and controls the merging ratio at 0.5;

[0051] The formula for the feature vector of the new node after merging:

[0052] h ij = s(h i + h j ), s = max(s ij + sji ) (2)

[0053] Among them, s ij is the edge score of node v i pointing to v j and s ji is the edge score of node v j pointing to v i .

[0054] Skip connection network: In the present invention, the outputs of all network layers from conv1 to conv5 and from pool1 to pool2 are aggregated using the skip connection network to enrich multi-level graph feature information, strengthen the learning of graph embedding, connect a global average readout layer after each network layer, output the overall feature representation of the intermediate graph, and obtain the multi-level graph feature representation through concatenation aggregation.

[0055] Among them, the formula of the readout layer is as follows:

[0056]

[0057] Among them, σ is the Sigmoid function, and h i is the feature vector of a single node.

[0058] The JK-Net aggregation formula is as follows:

[0059]

[0060] Among them, || is the concatenation aggregation operation.

[0061] Finally, a fully connected layer is used to fit the features output after the aggregation of JK-Net; the softmax function layer completes the classification of the CAD model.

[0062] Furthermore, in step 4, the CAD model is classified using the feature-level graph descriptor and the graph convolutional network classification model trained in step 3 to complete the correct classification of the input undetermined CAD model.

[0063] As shown in Table 1, based on the method of the present invention, the classification accuracy of the CAD model on the self-constructed 18-class dataset is 0.99206, which is greater than the average accuracy of methods based on multi-views or point clouds such as MVCNN, DGCNN, and CurveNet, indicating that the method of the present invention has a better classification effect compared with other classical 3D model classification methods.

[0064] Table 1 Classification effect of 3D CAD models

[0065]

Claims

1. A CAD model classification method based on feature-level graph descriptors and graph convolutional networks, characterized in that The method is as follows: Step 1: Obtain a feature-oriented model file as the basic data set; Step 2: Convert the.sldprt file of the CAD model in the data set into a feature-level graph descriptor in TU format. The specific operations are as follows: Step 2-1: Read the feature tree of the CAD model in the.sldprt model file, extract the features of the specified feature interfaces, where the feature interfaces include ExtrudeFeatureData, LoftFeatureData, SweepFeatureData, RevolveFeatureData, LinearPatternFeatureData, CircularPatternFeatureData, IMirrorPatternFeatureData, ISimpleFilletFeatureData, IVariableFilletFeatureData, IChamferFeatureData, WizardHoleFeatureData, SimpleHoleFeatureData, ShellFeatureData, ThreadFeatureData, and then extract the parent-child dependency relationships in the feature attributes. For the extracted features, construct a feature dependency graph G with features as nodes and the parent-child dependency relationships of features as edges fea , as the basis for the feature-level graph descriptor; Step 2-2: Design the attribute vector of the feature-level graph descriptor node, and read the dependency relationship and interface parameters of the feature, so as to define what attributes the descriptor node has; Read the feature dependency graph G fea Read the sub-dependency relationships of the features in it, and sequentially read the first sketch as the sketch for generating the feature. If there is no sketch in the sub-dependency relationship, set the corresponding feature sketch to none. Extract the parameters of the sketch line segments according to the predefined attributes, and analyze the features of the line segments to obtain the attribute vectors of the sketch line segments. Combine the attribute vectors of the sketch in this way. If there is no sketch, set the corresponding dimension to E; Preprocess the model dataset, obtain all features according to the method of capturing features in step 2-1, read all parameters of the feature corresponding to the feature interface, and convert them into quantifiable numerical values. Extract the numerical values of all parameters for the dataset, construct a numerical distribution for the same parameter. If the distribution of the same numerical value of the same parameter is higher than 90% of the whole, it is determined that the parameter does not have geometric expressiveness and is screened out. At the same time, perform weighted combination on the parameters of the same class, x d = w1x1 + w2x2 + … + w n x n , where x n represents the nth parameter, and w n represents the weighting coefficient of the nth parameter; to compress the dimension of the parameter, finally, according to the feature type divided by the feature interface, design an attribute vector for each feature, and for the attribute vector with inconsistent dimensions, assign the value E to the undefined dimension; Step 2-3: Perform "array relationship conversion" on the array features in the feature dependency graph G fea to expand the array features into multiple source features of their array copies, and connect them to G fea according to the dependency relationships of the source features, forming a feature-level graph descriptor G fea-level ; Step 3: Combine the skip connection network and the edge contraction-based hierarchical pooling mechanism to construct a graph convolutional classification network for the CAD model to realize the classification of the CAD model. The specific operations are as follows: The CAD model classification network uses the GraphSAGE graph convolutional network as the backbone network, and integrates the skip connection network and the edge contraction pooling mechanism. For the CAD model oriented to constructive geometry, this network is denoted as CSG-GraphNet; CSG-GraphNet contains a total of 5 convolutional parts conv1 to conv5, and hierarchical pooling parts pool1 and pool2 are added after its conv2 and conv4 convolutional parts; Add the residual unit JK-Net of the skip connection network to aggregate the network layer outputs of conv1 to conv5 and pool to pool2; The GraphSAGE graph convolutional network realizes node message passing based on the spatial domain, and proposes a neighbor sampling mechanism to enable the graph convolutional network to adapt to graphs of different sizes. When performing convolution, sample neighbor nodes, and control the number of sampled nodes by setting the sampling magnification Si; Step 3-1: Add an edge contraction-based hierarchical pooling module to the classification model, and the edge contraction pooling realizes pooling following the edge relationship between nodes; Step 3-2: Add a skip connection network module to the classification model to aggregate intermediate graph representations at different levels; The operation of the skip connection network module is as follows: First, take the intermediate graph representation output by each network layer as the input, and respectively pass through a global average readout layer to read the graph structure as a canonical feature matrix; Then, splice and merge the feature matrices output by all readout layers into a multi-level feature matrix to better represent the hierarchical changes of the graph.

2. The CAD model classification method based on feature-level graph descriptors and graph convolutional networks according to claim 1, characterized in that: The specific operation of obtaining a feature-oriented model file as the basic data set is as follows: Obtain model files from public model websites and partial model libraries of machinery manufacturing enterprises. If the model file contains a CAD model.sldprt file, use it as the model to be prepared for use; Screen the models to be prepared for use, design a script using the SOLIDWORKS API, and screen out models without a reasonable feature tree, and then further manually screen the remaining models; At the same time, use the Toolbox tool of SOLIDWORKS to randomly generate models within the category range to supplement the existing model categories; Finally, obtain a data set of.sldprt model files with a feature structure.

3. The CAD model classification method based on the feature-level graph descriptor and the graph convolutional network according to claim 1, characterized in that: The hierarchical pooling module based on edge contraction is added to the classification model. The hierarchical pooling based on edge contraction realizes pooling by following the edge relationships between nodes. Specifically, according to the nodes at both ends of the edges in the graph, scores are calculated for the edges, the edges in the graph are contracted in descending order of the scores, the nodes at both ends of the edges in the graph are aggregated, and the original edge relationships between the two end nodes are retained. During the contraction process, the pooling node ratio is controlled at 0.5, and finally the pooled graph is output.

Images