Three-dimensional object recognition method based on multi-view object recognition network

Abstract

The present invention discloses a three-dimensional target recognition method based on a multi-view target recognition network, comprising: obtaining initial multi-views based on the diversified viewpoint positions of the three-dimensional target; encoding the initial multi-views using an image encoder of a pre-trained large model to obtain multi-view features; obtaining representative views and their semantic information loss based on the multi-view features; fusing the representative views based on a multi-view feature fusion network to obtain global features of the three-dimensional target; obtaining text features of the three-dimensional target using a text encoder of a pre-trained large model; obtaining prediction results and constructing a cross-entropy loss based on the text features and the global features, training and updating the parameters of the multi-view target recognition network based on the semantic information loss and the cross-entropy loss, and obtaining a trained multi-view target recognition network. The present invention optimizes the three-dimensional target representation from three levels: viewpoint position, number of views, and feature fusion, and realizes three-dimensional target recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of target visual recognition, and in particular relates to a three-dimensional target recognition method based on a multi-viewpoint target recognition network. Background Art

[0002] Due to the widespread application of 3D objects in industrial CAD, unmanned driving, VR / AR and other fields, as well as the rapid development of various model creation technologies such as 3D scanning and 3D reconstruction, massive 3D objects are created and used. Therefore, object recognition remains a research direction that has attracted much attention in the field of computer vision. However, due to the expensive collection and labeling costs, the diversity and coverage of 3D object databases are relatively small, resulting in a limited number of datasets available for object recognition tasks and difficulty in obtaining object semantic information. In comparison, multimodal pre-trained large models take advantage of large-scale image-text datasets extracted from the Internet, achieve outstanding performance in a variety of visual tasks, and have rich visual semantic information. Therefore, how to effectively manage and recognize 3D objects using the visual information in pre-trained large models with low training resources is a challenging task.

[0003] In addition, most traditional multi-viewpoint-based target recognition algorithms use a set of fixed viewpoint positions to obtain three-dimensional target view information, resulting in the loss of target detail information, and do not select and eliminate the initial multi-view information, resulting in a large number of redundant, semantically ambiguous, and low-quality views, which hinders the ability of the multi-view network to fully understand the target.

[0004] Based on the above analysis, some of the current challenges are as follows: how to effectively use the semantic information in the pre-trained large model to obtain a global feature of a three-dimensional target with discriminative information to achieve the target recognition task; secondly, how to effectively improve the viewpoint position and number of views in the multi-viewpoint algorithm to obtain a low number of high-quality view data to represent the three-dimensional target. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a three-dimensional target recognition method based on a multi-view target recognition network to solve the problems existing in the above prior art.

[0006] To achieve the above object, the present invention provides a three-dimensional target recognition method based on a multi-view target recognition network, comprising:

[0007] Acquire diverse viewpoint positions of the three-dimensional object, and obtain initial multiple views based on the diverse viewpoint positions;

[0008] Encode the initial multi-views using an image encoder of a pre-trained large model to obtain multi-view features; obtain representative views and their semantic information losses based on the multi-view features;

[0009] The representative views are fused based on a multi-view feature fusion network to obtain global features of the three-dimensional object;

[0010] A text encoder of a pre-trained large model is used to obtain text features of a three-dimensional target; based on the text features and the global features, a prediction result is obtained and a cross entropy loss is constructed; based on the semantic information loss and the cross entropy loss, parameters of a multi-view target recognition network are trained and updated to obtain a trained multi-view target recognition network;

[0011] A three-dimensional target to be identified is obtained, and a trained multi-view target recognition network is used to perform target recognition on the three-dimensional target to be identified, so as to obtain a recognition result of the three-dimensional target.

[0012] Preferably, the diversified viewpoint positions include: a basic viewpoint position and a supplementary viewpoint position; wherein the basic viewpoint position is a surrounding viewpoint of the three-dimensional target, and the supplementary viewpoint position is a viewpoint selected by adaptive regression of a viewpoint regression network.

[0013] Preferably, the process of obtaining the initial multi-view includes:

[0014] A rasterizer based on a differentiable renderer converts the diverse viewpoint positions into a two-dimensional view;

[0015] The two-dimensional view is shaded based on a shader of a differentiable renderer to obtain an initial multi-view.

[0016] Preferably, the process of obtaining a representative view and its semantic information loss based on the multi-view features includes:

[0017] Based on the multi-view features, obtaining cluster centers by selecting multi-view feature cluster centers;

[0018] Based on the coordinate values ​​of the cluster centers, a representative view is obtained by minimum distance calculation; wherein the representative view reflects the common features in each view feature cluster;

[0019] Based on the representative view, a semantic information loss of the representative view is obtained by a strategy of minimizing prediction entropy, wherein the semantic information loss constrains the representative view to have clear semantics.

[0020] Preferably, the semantic information loss of the representative view is:

[0021]

[0022] in, is the predicted entropy of the representative view.

[0023] Preferably, the process of obtaining the global features of the three-dimensional object includes:

[0024] Merging and compressing the representative views to obtain comprehensive view features;

[0025] Based on the attention mechanism, an attention score of the representative view relative to the comprehensive view feature is calculated;

[0026] Based on the attention score, a weighted comprehensive view feature is generated, and a residual connection is added to the multi-view feature fusion network. Based on the balance factor of the residual connection, the comprehensive view feature, and the weighted comprehensive view feature, a global feature of the three-dimensional target is fused.

[0027] Preferably, the cross entropy loss is:

[0028] Where CE(·) represents the cross entropy loss, y is the true label of the three-dimensional target, is the predicted probability of the three-dimensional target for all categories;

[0029] The overall loss function of the multi-view object recognition network is:

[0030] L=β1L info +β2L cls

[0031] Among them, β1 and β2 are the hyperparameter weights of representative view semantic information loss and cross entropy loss, respectively.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] The present invention provides a three-dimensional target recognition method based on a multi-view target recognition network, comprising: obtaining diversified viewpoint positions of the three-dimensional target, and obtaining initial multi-views based on the diversified viewpoint positions; encoding the initial multi-views using an image encoder of a pre-trained large model to obtain multi-view features; obtaining representative views and their semantic information losses based on the multi-view features; fusing the representative views based on a multi-view feature fusion network to obtain global features of the three-dimensional target; obtaining text features of the three-dimensional target using a text encoder of a pre-trained large model; obtaining prediction results and constructing cross-entropy losses based on the text features and the global features, training and updating parameters of the multi-view target recognition network based on the semantic information loss and the cross-entropy loss to obtain a trained multi-view target recognition network; finally, performing target recognition on the three-dimensional target to be recognized through the trained multi-view target recognition network to obtain a recognition result of the three-dimensional target.

[0034] The present invention makes up for the defect that the traditional fixed viewpoint may not cover the blind spots of the viewpoint through the technical solution of diversified viewpoint positions, so that the subsequent initial rendering obtains multiple views that retain the complete features and details of the three-dimensional target as much as possible. The present invention selects a group of representative multi-views to improve the computational efficiency of the subsequent network while reducing memory consumption. In addition, with the help of text features, the representative views are made semantically clear and information interference from redundant and semantically ambiguous views is avoided. Compared with the traditional multi-view pooling operation, the present invention optimizes the three-dimensional target representation from three levels: viewpoint position, number of views, and feature fusion, based on the powerful visual generalization ability in the pre-trained large model, to achieve three-dimensional target recognition. The fused global features are of higher quality, which greatly improves the discriminability of the global features of the three-dimensional target. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0036] Figure 1 A schematic diagram of a flow chart of a three-dimensional target recognition method based on a multi-view target recognition network according to an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of a framework of a three-dimensional target recognition method based on a multi-view target recognition network according to an embodiment of the present invention;

[0038] Figure 3 is a schematic diagram of viewpoint position scene parameters according to an embodiment of the present invention;

[0039] Figure 4 A schematic diagram of a framework of a representative view selection module according to an embodiment of the present invention;

[0040] Figure 5 Schematic diagram of the framework of a multi-view feature fusion network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0043] Embodiment 1

[0044] like Figure 1-2As shown, in this embodiment, a three-dimensional target recognition method based on a multi-view target recognition network is provided, which uses views rendered at multiple viewpoint positions to represent the three-dimensional target, adopts a multimodal pre-trained large model as a visual feature extractor to obtain view features with rich visual semantics, and supplements the view regression network, the representative view selection module, and the multi-view feature fusion network with text feature constraints, respectively, optimizes the three-dimensional target representation from three levels: viewpoint position, number of views, and feature fusion, and further mines the structural information of the three-dimensional target on the basis of low-quantity and high-quality view features, thereby improving the target recognition accuracy, including the following steps:

[0045] Step 1: Determine a set of basic viewpoint positions and supplementary viewpoint positions obtained by the viewpoint regression network for the 3D target data, use a variety of viewpoint positions to complement each other to fully capture the overall structure and detail information of the target, and use a differentiable renderer to render based on multiple viewpoint positions to obtain a set of multi-view representations of the 3D target;

[0046] Specifically, given any input 3D target S, N multi-viewpoint positions are used to render N views to represent the 3D target, where the viewpoint position data includes three scene parameters: azimuth, elevation, and distance, such as Figure 3 As shown, they are defined as This embodiment divides the multi-viewpoint positions into basic viewpoint positions and supplementary viewpoint positions. The basic viewpoint positions adopt conventional surround viewing angles to mine the basic structural features of the three-dimensional target. A viewpoint regression network is designed for the supplementary viewpoint positions, and the viewpoint positions are adaptively constrained according to subsequent losses to provide additional viewpoint position rendering views, so as to understand and observe the three-dimensional target from more angles, thereby enriching the model's initial multi-view information and contributing to a more comprehensive characterization of the three-dimensional target.

[0047] The basic viewpoint position is a set of predefined fixed viewing angles. This embodiment follows the similar surround viewpoint setting in the previous multi-viewpoint algorithm. It is assumed that each input model is along a fixed axis (z axis) as the vertical direction, and rises 30° in the horizontal position. M viewpoint positions are evenly spaced in the vertical direction of the surround model, which are defined as

[0048] Only using the traditional predefined fixed viewpoint position may not be able to fully capture all the features and details of the three-dimensional target. Some perspectives may lack key information, resulting in the inability of the subsequently extracted multi-view features to accurately characterize the model, making the algorithm less robust to three-dimensional target posture changes and occlusions. Therefore, this embodiment adds a set of supplementary viewpoint positions on the basis of the basic viewpoint position, and adaptively regresses and selects a set of supplementary viewpoint positions through a viewpoint regression network to enrich the rendered multi-views, mine the structure and feature information that is not easily observed from conventional viewpoint positions, and improve the multi-view feature representation. Specifically, this embodiment introduces a learnable viewpoint regression network D, which inputs randomly initialized view scene parameters Through the subsequent hierarchical text constraint loss training network parameters, the output M' supplementary viewpoint positions are defined as

[0049]

[0050]

[0051]

[0052]

[0053] in, is the normalized viewpoint scene parameter of the output of the viewpoint regression network D, tanh(·) is the activation function, They are the constraints of azimuth, elevation and distance, and take values ​​of 180, 90 and 2 respectively, so that the scene parameters of the supplementary viewpoint position meet the parameter range.

[0054] Therefore, the final multi-viewpoint position parameters of the three-dimensional target are composed of the basic viewpoint position and the supplementary viewpoint position:

[0055]

[0056]

[0057]

[0058] In this embodiment, we use the above multi-viewpoint positions and use a differentiable renderer R to render a set of N initial multi-views X = {x1, x2, ..., x N} represents the three-dimensional target S. Since the differentiable renderer R is parameter differentiable compared to the traditional renderer, the gradient An end-to-end deep learning framework can be established to train the viewpoint regression network by propagating back from each rendered image to the scene parameters of the viewpoint position. Specifically, the differentiable renderer consists of a rasterizer and a shader. The rasterizer maps the surface of the 3D target table data to the pixels of the 2D view according to the multi-viewpoint position parameters to form a view. The shader colors the view according to the rendering parameters based on the correspondence between the table surface of the 3D target and the 2D pixels to output the final multi-view:

[0059] X=R(u a ,u e ,u d ,S)

[0060] Step 2: For a set of multi-view data of three-dimensional targets, use the image encoder in the pre-trained large model to encode the multi-view data to obtain multi-view features. Then, at the feature level and decision level, high-quality representative views are obtained through multi-view feature cluster center selection and semantic information loss constraints based on predicted entropy, which improves the feature correlation of views, highlights the clear semantic information of views, and reduces the interference of redundant views on recognition results.

[0061] This embodiment uses the image encoder E in the pre-trained large model v It replaces the two-dimensional convolutional network in the traditional multi-view algorithm to extract features from multiple views, obtains rich visual information in the pre-trained large model, and obtains the initial multi-view features with strong discriminative power.

[0062] F=E v (X)

[0063] According to step 1, the multi-viewpoint positions and the rendering of the differentiable renderer are determined, and a set of multi-view data with comprehensive information is obtained to ensure that the initial multi-view data can fully represent the overall and detailed structure of the three-dimensional object. However, in the initial multi-view set of the three-dimensional object, it is inevitable that some views lack clear semantic content, and too much redundant data will also hinder the accurate recognition of the object. This embodiment proposes a representative view selection module, such as Figure 4 As shown, for the initial multiple views, feature-level selection and decision-level constraints are used to obtain a set of representative views, so as to improve the processing efficiency of the subsequent feature fusion network for multiple views with a small amount of high-quality view data, capture the main features in multiple views, and avoid information interference from redundant and semantically ambiguous views, thereby optimizing the overall target recognition performance.

[0064] The representative view selection module uses a spectral clustering algorithm based on graph theory at the feature level for the initial multi-views, clusters them into multiple feature clusters according to the similarity between the multiple views, and determines the multiple feature cluster centers of the initial multi-views. Based on the cluster centers, the view representing the common features of each cluster at the feature level is selected. Specifically, we construct N nodes based on N multi-view features, use the similarity between nodes as the weight of the edge, and calculate the similarity of each pair of views:

[0065]

[0066] Among them, S(i,j) represents the similarity between view i and view j, is the similarity matrix composed of the similarities between all views, which is represented as a graph, f i and f j are the feature vectors of view i and view j respectively.

[0067] Next, we use the spectral clustering algorithm to construct the Laplacian matrix L of the graph to capture the geometric structure of the graph:

[0068] L=DS

[0069] in, is the degree matrix, which is a diagonal matrix with diagonal elements D(i,i)=∑ j S(i,j) represents the sum of the weights of all connections of the i-th node.

[0070] Compute the eigenvalues ​​and corresponding eigenvectors of the Laplacian matrix L. Assume {λ1,λ2,…,λ N} is the eigenvalue and satisfies λ1≤λ2≤…≤λ N , and their corresponding eigenvectors are {z1,z2,…,z N}. If we assume that the number of selected views is N seclec , then select the first N seclec The eigenvector corresponding to the smallest non-zero eigenvalue To reflect the more prominent cluster structure of multi-view features, and construct the matrix as a column vector Mapping to another low-dimensional space is equivalent to transforming the original view feature points, so that views of the same cluster are more closely clustered together and views of different clusters are farther apart, making clustering more effective.

[0071] Normalize the matrix Z to get Each row vector in Z' is considered as a point, and K-means clustering is applied to select the number of views required and divide them into N seclec Feature Clusters Then, the mean of all feature points in each feature cluster is calculated as the cluster center.seclec Feature Clusters For example, the cluster center coordinates The calculation is as follows:

[0072]

[0073] in, represents the coordinate vector of the i-th point in the feature cluster, is the corresponding cluster center coordinate vector.

[0074] This embodiment uses each point (the transformation of the original view feature) to measure the distance from the cluster center based on the cluster center coordinates, and then determines that the view corresponding to the point with the smallest distance is the representative view of the cluster:

[0075]

[0076] Among them, Z'(i) is the transformed and standardized view feature, and its corresponding original view feature f i For nth seclec Feature Clusters Representative view of .

[0077] After the above operations, we can get the selected representative views which can be expressed as Reflect the common features in each view feature cluster and remove views with duplicate information.

[0078] This embodiment also uses a strategy based on minimizing the prediction entropy to semantically constrain the selected views, so that the selected view features are more consistent with the visual space of the pre-trained large model, making the semantics clearer at the decision level, and avoiding the view's category semantics being too vague and affecting the subsequent feature fusion network's recognition results of the target. The prediction entropy reflects the prediction uncertainty and can represent the prediction confidence of the pre-trained large model under all category text descriptions of the view feature. Therefore, at the decision level, the representative view is made to have a higher prediction confidence as much as possible. Specifically, a fixed text template is first used in combination with K category names to form a text description containing clear category hints, such as "A rendered view of [CLASS]", and then the text encoder E in the pre-trained large model is used. t Encode it to get its corresponding text features, defined as Using this text feature to evaluate the semantic representation of the view, the predicted entropy of a single view is calculated as:

[0079]

[0080]

[0081] Among them, s n Each bit in represents the semantic similarity score between the n-th view and the hierarchical text of K categories. In addition, represents the predicted probability of the kth class, Used to quantify information. Compared with other views of the same 3D object, a lower prediction entropy means that the view is more semantically clear. We measure the semantic clarity of the view based on the prediction entropy, and define the semantic information loss of the representative view based on the prediction entropy of the representative view:

[0082]

[0083] Step 3: For a set of representative views, this embodiment designs a multi-view feature fusion network based on the attention mechanism to learn the global structure of the 3D target and fuse multiple view features to generate the final global feature representation of the 3D target. The global feature aggregates the association and supplementary information between representative views of the 3D target from different viewpoints and is used to construct the cross entropy loss of downstream tasks.

[0084] After the representative view selection process, each representative view contains a large amount of model private information, which may be visually different. Simply using average view pooling and maximum view pooling may be meaningless and the fusion quality is poor.

[0085] This embodiment uses a cross-attention mechanism to construct a global structural relationship between multi-view feature fusion network learning samples to obtain a global feature representation of a three-dimensional object. Figure 5 Specifically, firstly, all representative view features are transformed into Combine and compress to get comprehensive view features f repre :

[0086]

[0087] in, represents the compressed comprehensive view features, concat(·) is the vertical feature concatenation operation, and MLP(·) is a multi-layer perceptron.

[0088] Then, the attention mechanism is used to build attention associations for multiple representative views. For each representative view feature and comprehensive view feature, the attention score of the representative view relative to the comprehensive view feature is calculated. The representative view feature and comprehensive view feature query Q, key K and value V are generated through different linear transformations:

[0089] Q repre =W Q f repre

[0090]

[0091]

[0092] Among them, Q repre A query representing the characteristics of a comprehensive view, and The nth seclec The keys and values ​​representing the view, W Q , W K and W V is the learnable parameter matrix.

[0093] No.n seclec Attention scores of representative views relative to comprehensive view features:

[0094]

[0095] Among them, Softmax represents the normalization function, is the transpose of the key, Scales the attention score as a scaling factor.

[0096] The attention weight is used to weight the corresponding value vector Generate a weighted comprehensive view feature representation f repre' :

[0097]

[0098] In order to retain the identification information extracted from the representative views, this embodiment adds a residual connection to the multi-view feature fusion network, so that the network can learn the residual mapping between input and output and enhance the intrinsic connection between multiple views:

[0099] f Global =f repre +αf repre'

[0100] Among them, f Global It represents the global feature representation of the final three-dimensional target, and the non-learnable hyperparameter α serves as the balancing factor of the residual connection.

[0101] Step 4: Use the text features encoded by the text encoder of the pre-trained large model to match the global features representing the three-dimensional target to obtain the prediction result of the three-dimensional target. Based on the prediction result and the true label, a cross entropy loss is constructed, which is combined with the semantic information loss in step 2 to jointly constrain and realize the training of the multi-viewpoint target recognition network.

[0102] In this embodiment, the multi-view object recognition network includes: a viewpoint regression network, a representative view selection module and a multi-view feature fusion network.

[0103] Based on the above steps, we obtain the global feature representation of the three-dimensional target. and text information containing category hints The predicted probability of each three-dimensional target sample for all categories can be calculated:

[0104]

[0105] The cross entropy loss function is used to constrain the model training:

[0106]

[0107] Among them, CE(·) represents the cross entropy loss, y is the true label of the three-dimensional target, is the predicted probability of the three-dimensional target.

[0108] The overall loss function of the multi-view object recognition network is:

[0109] L=β1L info +β2L cls

[0110] Among them, β1 and β2 are the hyperparameter weights of representative view semantic information loss and 3D object classification cross entropy loss, respectively.

[0111] Finally, a three-dimensional target to be identified is obtained; and the three-dimensional target to be identified is identified using the trained multi-view target recognition network to obtain a recognition result of the three-dimensional target.

[0112] To summarize, this embodiment uses the target overall loss function to train the multi-view target recognition network and update its parameters so that the network can complete the target recognition of the three-dimensional target to be identified, implement the multi-view target recognition method, improve the accuracy of three-dimensional target recognition, and meet various needs in practical applications.

[0113] Beneficial effects of this embodiment:

[0114] Based on the traditional viewpoint, this embodiment uses the viewpoint regression network to add a set of additional viewpoint positions for adaptive regression, which makes up for the defect that the traditional fixed viewpoint may not cover the viewpoint blind spot. Therefore, the viewpoint regression network is used to ensure the comprehensiveness of the viewpoint position, so that the subsequent initial rendering obtains multiple views that retain the complete features and details of the three-dimensional target as much as possible.

[0115] This embodiment uses a pre-trained large model to extract features from multiple views. By using the rich semantic information already available in the pre-trained large model, multiple feature cluster centers are obtained based on spectral clustering at the feature level, and then a group of representative multi-views are selected to improve the computational efficiency of the subsequent network while reducing memory consumption. In addition, with the help of text features, representative views are made semantically clear and information interference from redundant and semantically ambiguous views is avoided.

[0116] This embodiment proposes a multi-view feature fusion network based on the attention mechanism, which captures the association and complementary information between views by calculating the attention weights of each representative view relative to the comprehensive view features, and introduces the design of residual connection to enhance the intrinsic connection between multiple views, ensuring that the fused global features learn the global structural information of the three-dimensional target. Compared with the traditional multi-view pooling operation, the fused global features are of higher quality and the discriminability of the global features of the three-dimensional target is greatly improved.

[0117] Therefore, this embodiment can optimize the three-dimensional target representation from three levels: viewpoint position, number of views, and feature fusion, fully explore the structural information of the three-dimensional target, and improve the accuracy of target recognition.

[0118] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A three-dimensional target recognition method based on a multi-view target recognition network, characterized in that: The following steps are involved: Obtaining diversified viewpoint positions of a three-dimensional object, and obtaining initial multi-views based on the diversified viewpoint positions; the diversified viewpoint positions include: a basic viewpoint position and a supplementary viewpoint position; wherein the basic viewpoint position is a surrounding viewpoint of the three-dimensional object, and the supplementary viewpoint position is a viewpoint selected by adaptive regression of a viewpoint regression network; The process of obtaining the initial multi-view includes: A rasterizer based on a differentiable renderer converts the diverse viewpoint positions into a two-dimensional view; Coloring the two-dimensional view based on a shader of a differentiable renderer to obtain an initial multi-view; Encoding the initial multi-views using an image encoder of a pre-trained large model to obtain multi-view features with rich visual information; obtaining representative views and their semantic information losses based on the multi-view features; Based on the multi-view features, the process of obtaining a representative view and its semantic information loss includes: Based on the multi-view features, obtaining cluster centers by selecting multi-view feature cluster centers; Based on the coordinate values ​​of the cluster centers, a representative view is obtained by minimum distance calculation; wherein the representative view reflects the common features in each view feature cluster; Based on the representative view, a semantic information loss of the representative view is obtained by a strategy of minimizing prediction entropy, wherein the semantic information loss constrains the representative view to have clear semantics; The semantic information loss of the representative view is: in, is the predicted entropy of the representative view, N seclec is the number of representative views selected, n seclec is the index of a representative view, indicating the position of a view in a set of representative views; The representative views are fused based on a multi-view feature fusion network to obtain global features of the three-dimensional object; The process of obtaining the global features of a three-dimensional target includes: Merging and compressing the representative view features to obtain comprehensive view features; Based on the attention mechanism, an attention score of the representative view relative to the comprehensive view feature is calculated; Based on the attention score, a weighted comprehensive view feature is generated, and a residual connection is added to the multi-view feature fusion network, and a global feature of the three-dimensional object is obtained based on a balance factor of the residual connection, the comprehensive view feature, and the weighted comprehensive view feature; A text encoder of a pre-trained large model is used to obtain text features of a three-dimensional target; based on the text features and the global features, a prediction result is obtained and a cross entropy loss is constructed; based on the semantic information loss and the cross entropy loss, parameters of a multi-view target recognition network are trained and updated to obtain a trained multi-view target recognition network; A three-dimensional target to be identified is obtained, and a trained multi-view target recognition network is used to perform target recognition on the three-dimensional target to be identified, so as to obtain a recognition result of the three-dimensional target.

2. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 1, characterized in that: The cross entropy loss is: Where CE(·) represents the cross entropy loss, y is the true label of the three-dimensional target, is the predicted probability of the three-dimensional target for all categories; The overall loss function of the multi-view object recognition network is: L=β1L info +β2L cls Among them, β1 and β2 are the hyperparameter weights of representative view semantic information loss and cross entropy loss, respectively.

Images