A 3D point cloud image classification method and system based on feature distance representation

Abstract

This invention relates to the field of image processing technology, and more particularly to a 3D point cloud image classification method and system based on a feature distance representation. The method includes acquiring a 3D object image and converting it into a 3D point cloud image; inputting the point cloud image into a core feature module and a dimension calibration module, whereby the core feature module obtains the core features of the point cloud image; the dimension calibration module obtains the classifier features of the point cloud image; performing dimension calibration on the classifier features and calculating the distance between them and the core features to obtain distance features; and integrating the distance features into the classifier features for classification. This invention addresses the problem of existing methods involving overly complex local feature extraction operations that neglect the relationship between shallow and deep features, leading to reduced network robustness.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a 3D point cloud image classification method and system based on a feature distance representation. Background Technology

[0002] With the introduction of PointNet and its improved network PointNet++, the problem of the difficulty in applying deep learning to point clouds was solved, and on this basis, the research depth of deep learning on point clouds was promoted.

[0003] To date, most 3D point cloud applications still rely on PointNet or PointNet++ methods. In recent years, to further refine point cloud deep learning networks, convolutional, graph, or attention mechanisms have been used to explore local geometric information. Convolutional methods use shape relationship data from the point cloud and learn parameters carrying shape relationship information within the convolutional kernels through an MLP network. Graph-based methods establish connections between all points in the point cloud data, extract fully connected dynamic feature maps, and progressively increase the receptive field. Attention-based methods simulate the selective information reception and processing mechanism of human visual perception, allowing the network to autonomously prioritize key information and improve its sensitivity to critical information. These methods essentially rely on a complex local feature extractor for feature analysis of point cloud data; while overly complex local feature extraction designs have achieved good results, they neglect the relationships between shallow features and deep-shallow features, leading to reduced network robustness. Summary of the Invention

[0004] To address the shortcomings of existing methods, this invention constructs a feature distance representation for exploring the relationships between features at different levels of point clouds. Considering that 3D objects have an inherent hierarchical structure and are essentially combinations of small parts, it can be assumed that the intrinsic attributes of 3D objects at a certain structural level are similar or even identical. Deeper level features are actually detrimental to object recognition and network performance. Existing methods focus on extracting local features, i.e., deep features, and do not handle the intrinsic relationships between features at different levels very well. Therefore, this invention uses a feature distance representation to find the similarity between point cloud features and classifier features in different dimensions, and explores the relationships between features at different levels of 3D objects.

[0005] The technical solution adopted in this invention is: a 3D point cloud image classification method based on a feature distance representation includes the following steps:

[0006] Step 1: Acquire 3D object images and convert them into 3D point cloud images;

[0007] Step 2: Input the point cloud image into the core feature module and the dimension calibration module respectively. The core feature module is used to obtain the core features of the point cloud image; the dimension calibration module is used to obtain the classifier features of the point cloud image.

[0008] Furthermore, the core feature modules include the FPS module, the Group module, and the Tucker decomposition module.

[0009] Furthermore, the dimension calibration module includes: a baseline module, an MLP module, and a LINEAR module.

[0010] Furthermore, benchmark modules include, but are not limited to, PointNet++, PointNet, DGCNN, and PointMLP.

[0011] Step 3: Perform dimensionality calibration on the classifier features and calculate the distance between them and the core features to obtain the distance features;

[0012] Furthermore, the formula for calculating the distance is:

[0013]

[0014] Among them, S n Let T be an n-dimensional vector of vector S. n Let T be an n-dimensional vector.

[0015] Step 4: Incorporate distance features into the classifier features for classification.

[0016] Furthermore, the output of the classification is:

[0017]

[0018] Where N is the number of repetitions of the cosine distance feature, Tucker() is the tensor Tucker decomposition operation, coss() is the cosine distance calculation operation, and F is the number of repetitions of the cosine distance feature. BB For the classification features of the baseline module, P Ob This is the point cloud set of the 3D point cloud object Ob after sampling from the farthest point.

[0019] Furthermore, the 3D point cloud image classification system based on the feature distance representation includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement the 3D point cloud image classification method based on the feature distance representation.

[0020] The beneficial effects of this invention are:

[0021] 1. In order to effectively and accurately represent the inherent attributes of data, the relationship between features at different levels is represented by calculating the distance, which enhances the difference of features of different object classifiers, deepens the relationship between features and makes it more interpretable.

[0022] 2. It can be embedded into mainstream networks, significantly improving network performance and robustness. Furthermore, the proposed feature distance representation does not consume excessive resources and space, and is plug-and-play. Attached Figure Description

[0023] Figure 1 This is a block diagram of the 3D point cloud image classification method based on a feature distance representation according to the present invention;

[0024] Figure 2 This is a schematic diagram of the cosine distance of the point cloud image of the chair;

[0025] Figure 3 It is a point cloud feature extraction map of a 3D object;

[0026] Figure 4 This is a multi-level feature extraction structure diagram for PointMLP design;

[0027] Figure 5 This is a visualization of the core tensors of airplanes, palm trees, and tables in the ModelNet40 dataset. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0029] like Figure 1 As shown, the 3D point cloud image classification method based on feature distance representation includes the following steps:

[0030] Step 1: Acquire 3D object images and convert them into 3D point cloud images;

[0031] Step 2: Input the point cloud image into the core feature module and the dimension calibration module respectively. The core feature module is used to obtain the core features of the point cloud image; the dimension calibration module is used to obtain the classifier features of the point cloud image.

[0032] Step 3: Perform dimensionality calibration on the classifier features and calculate the distance between them and the core features to obtain the distance features;

[0033] Step 4: Incorporate distance features into the classifier features for classification.

[0034] The core feature modules include the FPS module, the Group module, and the Tucker decomposition module. The FPS and Group modules are based on the PointNet++ network. The FPS (Farthest Point Sampling) module selects n points from a set of N points, reducing the dataset size. The computation process of the FPS module is as follows: First, a point is randomly selected as the initial point and added to the initial point set. The Euclidean distances from the remaining points to the initial point are calculated, and the point with the furthest distance is added to the initial point set. Then, the distances from the remaining points to the initial point set are calculated. The smallest Euclidean distance from any remaining point to all points in the initial point set is taken as the distance from that point to the initial point set. The point with the largest distance from the remaining points to the initial point set is added to the initial point set, and so on until the length of the initial point set is n. The process of finding the initial point set and the FPS module is described below:

[0035]

[0036] Where P represents the initial point set, ||xP|| represents the Euclidean distance from the point to the initial point set, and x i The x represents points in the initial point set and those about to be added to the initial point set, ranging from 1 to n; j P represents the remaining points outside the initial point set, ranging from 1 to N-i+1; define P Ob ={x1,x2,...,x n} represents the point cloud set of a 3D point cloud object Ob after sampling from the farthest point.

[0037] The grouping operation forms a local neighborhood by connecting the centroid with local points within the same radius, facilitating the network's learning of spatial structural relationships between points. The sphere radius query method can find all points within the radius of the centroid. The input to the grouping operation is a set of points of size n×(d+c) and a set of centroid coordinates of size n′×d, where n′ is the number of centroids. Here, n×(d+c) represents n points with d-dimensional coordinates and c-dimensional point features. The output is a set of points of size n1×k×(d+c1), where each group corresponds to a local region, k is the number of points in the centroid's neighborhood, and n1 is the number of points in the grouped set.

[0038] The CP decomposition module and the Tucker decomposition module can be seen as higher-order extensions of tensor singular value decomposition: the CP decomposition module decomposes a tensor into the sum of rank tensors, and the Tucker decomposition is a higher-order form of principal component analysis (PCA); that is, the CP decomposition of a tensor decomposes it into the sum of many factor tensors, and the Tucker decomposition of a tensor essentially decomposes the original tensor into a core tensor and factor matrices corresponding to different dimensions. Each factor matrix can be viewed as a linear transformation operation corresponding to a different dimension. Taking a third-order tensor α as an example, if α∈R... I×J×O The formula for CP decomposition is as follows:

[0039]

[0040] in, Pointer outer product; R is the rank of CP and is a positive integer; a r ∈R I ,b r ∈R J c r ∈R O ;

[0041] The definition of Trucker decomposition is:

[0042]

[0043] Where A∈R I×X , B∈R J×Y , C∈R O×Z Core tensor The elements represent the level of interaction between different factor matrices; if X, Y, Z are less than I, J, O, the core tensor is equivalent to a compressed form of the original tensor, that is, the result of dimensionality reduction of the original tensor; therefore, Tucker decomposition can indeed be regarded as a higher-order form of PCA dimensionality reduction.

[0044] It is easy to see from formulas (2) and (3) that CP decomposition is a special case of Tucker decomposition, namely the case of core tensor superdiagonal tensor; CP decomposition is limited to decomposing tensors into multiple rank-1 matrices, while Tucker decomposition is more general.

[0045] Cosine distance, also known as cosine similarity, can be used to calculate distance. It's a method that assesses the similarity between vectors by calculating the cosine value between them. Let s:(s1,s2) and t:(t1,t2) be any two non-zero vectors in vector space. The cosine distance between them, which is the cosine of the angle between vectors s and t, is:

[0046]

[0047] It has been proven that this method of calculating cosine also holds true for n-dimensional vectors; assume that S and T are two n-dimensional vectors, and S is [S1, S2, ..., ST]. n ], T is [T1,T2,...,T n If the angle θ between S and T is θ, then the cosine of the angle θ is equal to:

[0048]

[0049] Cosine similarity algorithm is commonly used for data comparison. The closer the cosine value is to 1, the closer the angle between the two vectors is to 0 degrees, which means that the two vectors are more similar and contain more similar information. Similarly, tensors are also applicable to cosine similarity algorithm.

[0050] Cosine distance uses the cosine of the angle between two vectors to measure the difference between the two individuals. Compared to other distance methods, cosine distance focuses more on the difference in direction between the two vectors and is not sensitive to the numerical value. Figure 2 Point cloud data is mostly the contour data of objects. To some extent, point cloud data is sparse on the same cross plane. Therefore, the cosine angle between points on opposite planes or not close to each other contains distance information. For close points, since the benchmark model uses sampling and grouping operations of the farthest point to establish the adjacent region features, the points in the same adjacent region already have a connection. It can be seen that using cosine distance to apply to point cloud information has good advantages and interpretability.

[0051] The dimension calibration module includes a baseline module, an MLP module, and a LINEAR module. The baseline model is based on the PointNet++ network and includes an FPS module and a Group module.

[0052] Because different benchmark modules have different performance and functional focuses, the same object can extract different features using different network structures. Therefore, different benchmark modules can be used for testing. 3D point cloud data is fed into the benchmark module and aligned in terms of feature dimensions through a set of MLP modules and Linear modules. The output data size is N×d. out1 N is the number of groups, d out1 The size is after linear channel alignment. The FPS module uses the farthest distance point sampling to reduce the size of the point cloud data and obtains the point cloud set that best represents the outer contour of the point cloud object. Then, the Group module is used to divide the adjacent regions of the point cloud. At this point, the point cloud set is the closest to the original structural features after preprocessing. Therefore, the point cloud set is decomposed into tensor Tucker to obtain a core tensor and factor matrices corresponding to different dimensions. The point cloud set can be regarded as the core tensor obtained by linear transformation of different dimensions.

[0053] like Figure 3 The structure of a 3D object is fixed; it is essentially a combination of unchanging small parts. A single local feature extraction operation can be likened to a process of finding the next stage of features. The final stage features of each object are not unique. Figure 3Taking the desk, chair, table, and bed as examples, the features of the desk and chair at a certain stage are similar to those of the table. The features of these three 3D objects at a certain stage may be similar or even the same. Furthermore, the stage features of another branch of the chair are similar to those of the bed at a certain stage. This indicates that focusing solely on the ability to extract local features can lead to difficulties in distinguishing similar objects. The original 3D object does not significantly improve the overall recognition ability by extracting only coarse features, and the coarse and insufficiently refined features may even reduce the network's performance. After each local feature extraction, a 3D object enters a deeper feature level, but it also moves further away from the original structural features. Therefore, it is necessary to establish the relationship between hierarchical features, and distance is the best representation method. Since the ability of the local feature extractor is fixed, the level of feature extraction is also fixed. At the same or different levels, the distance between similar local features of different objects and their respective original feature levels is different. By incorporating the distance features between the two into the classifier features, the performance and robustness of the existing network can be further improved.

[0054] Existing local feature extractors cannot clearly identify the level at which the extracted local features are located. They mostly rely on existing experience or experimental results to verify the quality of classifier features. However, the feature distance representation of this invention, regardless of the capabilities of the local feature extractor, integrates the distance between the extracted levels as a distance feature into the classifier features. The distance feature has excellent identification and representation.

[0055] for Figure 3 Different 3D objects can be viewed as mappings of different core features in different spaces; the core tensor of the point cloud is considered to be the feature closest to the original structure space of the object, with a data size of N×(k×d). out2 ), d out2 Let d be the size of one dimension after truker decomposition. Since different objects may have the same core features, or the same core features can be reconstructed into different objects through different mappings, the cosine distance between the features of the baseline module after feature dimension alignment and the core tensor features is calculated to represent the relationship between the features. The size of the cosine distance is 1×d. out d out =d out1 =k×d out2 Each number in the cosine distance feature represents the weight of a certain adjacent region in the core tensor and all local features extracted by the baseline network. The cosine distance feature is repeated N times to obtain a result of size N×d. out The characteristic relationships between features are represented and incorporated into the classification features F of the baseline module. BB In the process, the input is fed into a fully connected layer for behavior recognition, and the overall expression is as follows:

[0056]

[0057] Where F represents the final classifier features, Tucker() is the tensor Tucker decomposition operation, coss() is the cosine distance calculation operation, and MLP and Linear layers are used for channel alignment.

[0058] Experimental comparison

[0059] Table 1 compares the performance of the benchmark module with the embedded feature distance representation with that of other networks.

[0060]

[0061]

[0062] This study investigates the performance of the feature distance representation on the synthetic dataset ModelNet40 and the real dataset ScanObjectNN. The feature distance representation of this invention is applied to various classification architectures, including DGCNN, PointNet++ models, and PointMLP models. During experiments, [the following is a list of parameters to be inserted]. Figure 1 The BACKBONE in the model is replaced with the baseline model. Then, the features of the baseline model are aligned and fused with the cosine distance features of this invention through a set of MLP modules and Linear modules. The parameters of the baseline model are set by default to the official settings.

[0063] During tensor decomposition, point cloud data is not compressed, and the core tensor size is the same as the original size, which makes it easier to retain more original information; based on the PyTorch architecture, the model is trained on a Tesla V100-PCIE 32G graphics card in a Linux environment.

[0064] Table 1 shows the results of this invention in ModelNet40 classification and a comparison with other state-of-the-art methods, using class average accuracy (mAcc) and overall accuracy (OA) as evaluation metrics. It can be seen that the baseline network with embedded feature distance representations shows varying degrees of improvement in both mAcc and OA recognition rates compared to the original network. When applied to PointMLP, the network's OA recognition rate surpasses the state-of-the-art performance on the ModelNet40 dataset, and its mAcc recognition rate is second only to the PointNet++ network with embedded feature distance representations.

[0065] The parameters of the baseline network model are the same as those in the official PointNet++ documentation, and each experiment is trained for 200 epochs.

[0066] Table 2: ModelNet40-PointNet++

[0067]

[0068]

[0069] As shown in Table 2, the performance of PointNet++ with embedded feature distance representations surpasses that of the original network under various settings. Compared to the approximately 0.2% improvement in OA recognition rate, the improvement in mAcc is more significant. PointNet++ with embedded feature distance representations achieves the highest improvement under the settings of SSG feature extraction and feature normalization, improving the mAcc recognition rate by 1.66% compared to the original network under the same settings and by 2.15% compared to the original network under SSG settings. The feature distance representation integrates the distance between features at different levels into the classifier features, enhancing the differences between different categories of objects and significantly improving the average accuracy and classification accuracy in recognition and classification.

[0070] Secondly, the performance of DGCNN with embedded feature distance representations is evaluated on the ModelNet40 dataset. The baseline network model parameters are the same as the official parameters of DGCNN, and each experiment is trained for 300 epochs.

[0071] Table 3: ModelNet40-DGCNN

[0072] method Points mAcc (%) OA (%) *DCGNN 1024 89.41 92.58 *DGCNN+ Feature Distance Representation 1024 90.02 92.63 *DCGNN 2048 90.16 93.07 *DGCNN+ Feature Distance Representation 2048 90.79 93.15

[0073] As shown in Table 3, under the same point count setting, the DGCNN baseline model with embedded feature distance representation has higher mAcc and OA recognition rates than the original DGCNN model. Specifically, the DGCNN baseline model with embedded feature distance representation improves mAcc recognition accuracy by 0.61% and 0.63% with 1024 and 2048 points, respectively. Similar to PointNet++, embedding cosine feature representation significantly improves average accuracy and classification accuracy, with an overall accuracy improvement that is not significant. This further demonstrates the correctness of the kernel tensor cosine feature representation.

[0074] Finally, the performance of PointMLP with embedded feature distance representations was evaluated on the ModelNet40 and ScanObjectNN datasets; the baseline network model parameters were the same as the official parameters of PointMLP, and each experiment was trained for 300 epochs.

[0075] Table 4: ModelNet40-PointMLP

[0076] method mAcc OA *PointMLP 91.0 93.4 *PointMLP (voting) 91.0 93.7 *PointMLP+ Feature Distance Representer 91.1(+0.1) 93.5(+0.1) *PointMLP+ Feature Distance Representation (voting) 91.2(+0.2) 94.0(+0.3)

[0077] Table 4 shows that compared to the PointMLP baseline model, the PointMLP model with embedded feature distance representations improved both mAcc and OA accuracy before and after using the voting mechanism. Specifically, the PointMLP model with embedded feature distance representations showed a 0.3% improvement in OA accuracy compared to the original model. Overall, while PointMLP's accuracy improved after embedding feature distance representations, the improvement was not significant. PointMLP consists of multiple layers of pure residual MLP structures. MLP is a simplified spatiotemporal local feature extractor, where structural information contains distance information. While extracting features step by step, PointMLP continuously inserts spatiotemporal structural information from the previous layer through the residual structure, thus continuously integrating and fusing distance information from each layer. Therefore, embedding feature distance representations in PointMLP does not yield significant results.

[0078] Table 5: ScanObjectNN-PointMLP

[0079] method mAcc OA *PointMLP 84.2 85.6 *PointMLP+MLP Feature Representation 84.2(+0.0) 85.7(+0.1) *PointMLP+ Feature Distance Representer 84.2(+0.0) 86.0(+0.4)

[0080] The results shown in Table 5 are similar to those in Table 4. The PointMLP model with embedded feature distance representation does not perform well on the ScanObjectNN dataset, which also excludes the influence of different datasets on our conjecture.

[0081] Fusion Experiment

[0082] To verify the issue of the PointMLP model's ineffective performance in feature distance representation, PointMLP was designed as a multi-level feature extraction structure, such as... Figure 4 The residual structure remains unchanged. Taking the official 4-stage structure of PointMLP as an example, in the original structure, spatiotemporal structural information (distance information) can only be transmitted between blocks in a stage. However, the residual structure is not used between stages, which leads to information gaps and affects the transmission of distance information. Experiments are conducted after filling the gaps.

[0083] Table 6:

[0084] method mAcc OA PointMLP+ Multi-level Feature Extraction 91.1(+0.1) 93.4(+0.0) *PointMLP+ Multi-level Feature Extraction (voting) 91.1(+0.1) 93.8(+0.1) *PointMLP+ Feature Distance Representer 91.1(+0.1) 93.5(+0.1) *PointMLP+ Feature Distance Representation (voting) 91.2(+0.2) 94.0(+0.3)

[0085] The purpose of improving the fault is not to improve the PointMLP structure, but to fill in the gaps in the transmission of distance information in the PointMLP, which is in line with the idea of ​​the connection between features at different levels.

[0086] To investigate whether data channels should be compressed during tensor decomposition, the following experiments were conducted. Table 7 shows the compression of the 3D coordinate channel of the point cloud data, and Table 8 shows the compression of the point number channel of the point cloud data. Both experiments used pointMLP as the baseline network, and Table 7 adopted the vote mechanism.

[0087] Table 7: Tucker decomposition channel compression (pointMLP uses backbone, xyz3 channel compression)

[0088] method mAcc (%) OA (%) *PointMLP+ Feature Distance Representer / Compressed to 1 Channel 91.1 93.4 *PointMLP+ Feature Distance Representer / Compressed to 2 Channels 89.8 93.0 *PointMLP+ Feature Distance Representation / Invariable (3 Channels) 91.1 93.5 *PointMLP+ Feature Distance Representation / Compressed to 1 Channel (voting) 91.5 93.7 *PointMLP+ Feature Distance Representation / Compressed to 2 Channels (voting) 90.2 93.3 *PointMLP+ Feature Distance Representation / Invariant (3 Channels) (voting) 91.2 94.0

[0089] As shown in Table 7, when using Point MLP as the baseline model, the performance of Point MLP with embedded feature distance representation is better than that with the coordinate channels compressed to 2 channels when the 3 coordinate channels of the point cloud data remain unchanged or are compressed to 1 channel. Specifically, Point MLP with embedded feature distance representation, after compressing the 3 coordinate channels to 1 channel and employing the vote mechanism, achieves the highest mAcc (91.5%), which is 0.3% higher than the mAcc when the number of channels remains unchanged under the same conditions. Point MLP with embedded feature distance representation and unchanged channel number, employing the vote mechanism, achieves the highest OA (94.0%), which is also 0.3% higher than the OA when the number of channels is compressed to 1 channel under the same conditions.

[0090] Table 8: Tucker decomposition point compression (pointMLP is backbone, all coordinate channels are 3)

[0091] method mAcc OA *PointMLP / 1024 points 91.1 93.5 *PointMLP / 512 points 90.6 93.5 *PointMLP / 256 points 89.9 93.5

[0092] * indicates the result of running in the experimental environment.

[0093] As shown in Table 8, when pointMLP is used as the baseline model and the number of coordinate channels is 3, the number of points is compressed during Tucker decomposition to 1024, 512, and 256 points respectively. In all three cases, the OA accuracy of pointMLP with embedded feature distance representation is the same, which is 93.5%. However, as the number of points decreases, the mAcc accuracy of pointMLP with embedded feature distance representation also decreases. When the number of points decreases from 1024 to 256, the mAcc accuracy decreases by 1.2%. This is because the purpose of using tensor decomposition is to obtain more information that is close to the original data. As the number of points decreases, the overall information contained also decreases, and a small number of points cannot well represent the core information.

[0094] Furthermore, to further demonstrate that distance information between features is useful for building better networks, in Figure 5The text demonstrates the visualization of the core tensor of a portion of the ModelNet40 data. To more intuitively compare the differences between the core tensor and point cloud data, some processing was performed on the core tensor, using different shapes to divide different block or point regions and connecting them with straight lines. It can be seen that the core tensors of the airplane and table both consist of a central region and several adjacent point regions, representing the main body and torso respectively. The number of torso regions corresponds, with the main body region having a large number of clustered points, while the torso region has only a single point. The points in the main body of the airplane are relatively concentrated, while the points in the main body of the table, although concentrated, are relatively dispersed, corresponding to the characteristics of a tabletop. The four torso regions also correspond to the four table legs. The core tensor of the coconut tree in the Plants consists entirely of several point regions, without any concentrated parts. Connecting these regions reveals the original outline of the coconut tree, proving that the most core features of the 3D object were extracted from the core tensor space.

[0095] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A 3D point cloud image classification method based on a feature distance representation, characterized in that, Includes the following steps: Step 1: Acquire 3D object images and convert them into 3D point cloud images; Step 2: Input the point cloud image into the core feature module and the dimension calibration module respectively. The core feature module is used to obtain the core features of the point cloud image; the dimension calibration module is used to obtain the classifier features of the point cloud image. Step 3: Perform dimensionality calibration on the classifier features and calculate the distance between them and the core features to obtain the distance features; The formula for calculating distance is: in, S n For vectors S of n dimensional vector, T n For vectors T of n dimensional vector; Step 4: Incorporate distance features into the classifier features for classification; The output of the classification is: (6) in, N Cosine distance feature repetition count Tucker () represents the tensor Tucker decomposition operation. coss () is the operation for calculating cosine distance. The classification features of the baseline module, This is the point cloud set of the 3D point cloud object Ob after sampling from the farthest point.

2. The 3D point cloud image classification method based on a feature distance representation according to claim 1, characterized in that, The core feature modules include the FPS module, the Group module, and the Tucker decomposition module.

3. The 3D point cloud image classification method based on a feature distance representation according to claim 1, characterized in that, The dimension calibration module includes: a baseline module, an MLP module, and a LINEAR module.

4. The 3D point cloud image classification method based on feature distance representation according to claim 1, characterized in that, The benchmark modules include PointNet++, PointNet, DGCNN, and PointMLP.

5. A 3D point cloud image classification system based on a feature distance representation, characterized in that, include: Memory is used to store instructions that can be executed by the processor; A processor for executing instructions to implement the 3D point cloud image classification method based on a feature distance representation as described in any one of claims 1-4.

Images