Three-dimensional object identification method and system based on recursive gated convolution

Through the three-dimensional object recognition method based on recursive gating convolution, the convolution neural network and feature interaction module are used to enhance view feature interaction, which solves the problem of insufficient view information exchange and redundancy, and improves the accuracy of three-dimensional object recognition.

CN119068246APending Publication Date: 2024-12-03GUANGDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411141273.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing three-dimensional object recognition methods have limitations in insufficient information exchange between views and redundancy of views, resulting in low recognition accuracy, and the existing methods ignore the differentiation and diversity information of views when reducing redundancy.

Method used

Using a method based on recursive gating convolution, multi-scale view features are extracted through convolution neural networks, and the local feature interaction module and global feature optimization module are used to enhance view feature interaction, including low-order recursive gating convolution and high-order recursive gate The control convolution module preferably has a view feature and is fused to generate a three-dimensional shape descriptor for identification.

Benefits of technology

It improves the accuracy of three-dimensional object recognition, enhances the interactive relationship of view features, reduces the interference of duplication of similar features and single information in the view, and improves recognition performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068246A_ABST
    Figure CN119068246A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of three-dimensional object recognition, and discloses a three-dimensional object recognition method and system based on recursive gating convolution, and the method comprises the steps: obtaining a plurality of two-dimensional views of a to-be-recognized three-dimensional object; inputting the plurality of two-dimensional views into a convolutional neural network to obtain multi-scale view features of each view; inputting the multi-scale local view features of each two-dimensional view into a local feature interaction module to obtain local interaction features; inputting the global view feature of each two-dimensional view into a global feature optimization module to obtain a global optimization feature; adding the local interaction feature and the global view feature corresponding to each view, and fusing the added result with the global optimization feature to obtain a three-dimensional shape descriptor; and performing identification based on the three-dimensional shape descriptor to obtain an identification result. According to the method, the feature interaction relationship between the views is enhanced, meanwhile, the interference of similar feature repetition and single information amount in the views is reduced, and the accuracy of three-dimensional object recognition is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional object recognition, and in particular to a three-dimensional object recognition method and system based on recursive gated convolution. Background Art

[0002] In recent years, three-dimensional data has appeared more and more in people's lives and work, showing broad prospects in the fields of autonomous driving and robot navigation, intelligent transportation, analysis and diagnosis of medical images, etc. Three-dimensional object recognition has therefore become a new research hotspot, and a large number of related research methods have emerged. The view-based method is currently the most mainstream method. It can obtain more comprehensive and accurate object information by observing objects from different perspectives. At the same time, since the view is two-dimensional data, it can avoid the errors and complexity brought by three-dimensional reconstruction and greatly reduce the amount of calculation. Since multiple views change through angles, object features from different perspectives usually provide complementary information. Therefore, there is a correlation between views. If the view features are simply fused directly, the features of different views will lack sufficient information interaction. Therefore, how to strengthen the information exchange between views is one of the main research issues at present. On the other hand, the feature information provided by views under similar perspectives may overlap to a certain extent. If all views are combined, the feature information provided by views under similar perspectives may overlap to a certain extent. Figure 1 Processing them together will bring about the problem of view redundancy, so how to reduce view redundancy is also one of the main issues that need to be studied at present.

[0003] In solving the problem of insufficient correlation between views, most existing studies are based on self-attention mechanisms and long short-term memory networks, which have the advantage of long-term dependencies and are conducive to mining the correlation between views. However, their feature interaction capabilities in spatial positions are poor and limited, resulting in low recognition accuracy. In solving the problem of view redundancy, most methods only focus on the richness of view information. For example, OVPT calculates the information entropy of the view, selects a part of the views with the largest information entropy value, and simplifies multiple views. However, this may result in multiple views with similar features and similar perspectives. There is still a risk of view redundancy, and the distinctiveness and diversity information conveyed by multiple views are ignored. These shortcomings will also have a negative impact on the performance of 3D object recognition. Summary of the invention

[0004] The primary purpose of the present invention is to overcome the problems existing in the prior art and provide a three-dimensional object recognition method based on recursive gated convolution. The present invention can improve the accuracy of three-dimensional object recognition.

[0005] As another object of the present invention, a system adapted to the method based on the aforementioned object is provided.

[0006] As another object of the present invention, a non-volatile storage medium suitable for storing a computer program implemented according to the method described is provided.

[0007] In order to achieve the above object, the present invention provides a three-dimensional object recognition method based on recursive gated convolution, comprising: Acquire multiple two-dimensional views of the three-dimensional object to be identified; Inputting multiple two-dimensional views into a convolutional neural network to obtain multi-scale view features of each view, wherein the multi-scale view features include multi-scale local view features and global view features; Input the multi-scale local view features of each two-dimensional view into the local feature interaction module to obtain local interaction features; Inputting the global view features of each two-dimensional view into a global feature optimization module to obtain a global optimization feature, wherein the global feature optimization module includes a preferred view module and a high-order recursive gated convolution module, wherein the preferred view module is used to obtain a preferred global view feature according to the global view features, and the high-order recursive gated convolution module is used to obtain a global optimization feature according to the preferred global view features; Adding the local interaction features and the global view features corresponding to each view, and fusing the added result with the global optimization feature to obtain a three-dimensional shape descriptor; Recognition is performed based on the three-dimensional shape descriptor to obtain a recognition result.

[0008] Further, the obtaining of multiple two-dimensional views of the three-dimensional object to be identified includes: projecting the three-dimensional object to be identified from multiple equidistant viewpoints to obtain multiple two-dimensional views, wherein the height of the viewpoint is 30 degrees relative to the horizontal plane.

[0009] Furthermore, the convolutional neural network is a ResNet network, and multiple two-dimensional views are input into the ResNet network to obtain multi-scale view features of each view.

[0010] Furthermore, the local feature interaction module is a low-order recursive gated convolution, which inputs the multi-scale local view features of each two-dimensional view into the low-order recursive gated convolution to obtain local interaction features, including: Project local view features of different scales to the same dimension; The local view features after projection are multiplied element by element to obtain the local features, as follows:

[0011] in, and These are all linear projection operations. It is a local feature; The local features are input into a low-order recursive gated convolution to obtain local interaction features.

[0012] Furthermore, the step of inputting the global view features of each two-dimensional view into a global feature optimization module to obtain the global optimization features includes: The global view features of each 2D view are divided into Group, where , is the number of viewpoints; Calculate the average gray value of each global view feature after classification , calculated as follows:

[0013] in, For the View eigenvalues, is the width of the view feature, is the height of the view feature; The variance of each global view feature is calculated based on the average grayscale value as follows:

[0014] According to the variance from Select the first several global view features in the group as the preferred global view features; The optimized global view features are input into the high-order recursive gated convolution module to obtain the globally optimized features.

[0015] Furthermore, the high-order recursive gated convolution module includes a first normalization layer, a recursive gated convolution, a second normalization layer and a multilayer perceptron connected in sequence, wherein the first normalization layer is used to normalize the optimized global view features to obtain the normalized global view features; the recursive gated convolution is used to obtain the interacted global view features based on the normalized global view features; the second normalization layer is used to obtain a normalized result based on the sum of the optimized global view features and the interacted global view features, and the multilayer perceptron is used to process the normalized result, and the obtained result is added to the interacted global view features to obtain the global optimization feature.

[0016] Furthermore, the local interaction features and the global view features corresponding to each view are added, and the result of the addition is fused with the global optimization feature to obtain a three-dimensional shape descriptor, as follows:

[0017] in, is the adaptive average pooling operation, To average the view features, For fusion splicing operations, is the global view feature, is the local interaction feature, It is a global optimization feature.

[0018] Furthermore, the performing recognition based on the three-dimensional shape descriptor to obtain a recognition result includes: inputting the three-dimensional shape descriptor into a fully connected layer to recognize the three-dimensional object.

[0019] In order to achieve another object of the present invention, a three-dimensional object recognition system based on recursive gated convolution is provided, the system comprising: The first acquisition module is used to acquire multiple two-dimensional views of the three-dimensional object to be identified; A second acquisition module is used to input multiple two-dimensional views into a convolutional neural network to obtain multi-scale view features of each view, wherein the multi-scale view features include multi-scale local view features and global view features; The third acquisition module is used to input the multi-scale local view features of each two-dimensional view into the local feature interaction module to obtain the local interaction features; A fourth acquisition module: used for inputting the global view features of each two-dimensional view into the global feature optimization module to obtain the global optimization features, wherein the global feature optimization module includes a preferred view module and a high-order recursive gated convolution module, wherein the preferred view module is used for obtaining the preferred global view features according to the global view features, and the high-order recursive gated convolution module is used for obtaining the global optimization features according to the preferred global view features; Fusion module: used to add the local interaction features and global view features corresponding to each view, and fuse the result of the addition with the global optimization feature to obtain a three-dimensional shape descriptor; Recognition module: used for performing recognition based on the three-dimensional shape descriptor to obtain a recognition result.

[0020] In order to achieve another purpose of the present invention, the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer-stored program is executed by a processor, the three-dimensional object recognition method based on recursive gated convolution is implemented.

[0021] Compared with the prior art, the present invention has the following beneficial effects: The present invention extracts view features of a two-dimensional view at multiple scales through a convolutional neural network, takes the extracted view features as a basis, and uses a local feature interaction module and a global feature optimization module to strengthen the view features from two aspects, locally and globally, respectively. For the local feature interaction module, the local features of multiple scales are input to realize the interaction of features in spatial positions by using low-order recursive gated convolution, thereby enhancing the information exchange between local view features; the global feature optimization module includes a preferred view module and a high-order recursive gated convolution module, the preferred view module is used to obtain the preferred global view features according to the global view features, and the high-order recursive gated convolution module is used to obtain the global optimization features according to the preferred global view features, thereby enhancing the feature interaction relationship between views, and at the same time reducing the interference of similar feature duplication and single information in the view, thereby improving the accuracy of three-dimensional object recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of a three-dimensional object recognition method based on recursive gated convolution according to Embodiment 1 of the present invention; Figure 2 is a block diagram of a three-dimensional object recognition system based on recursive gated convolution according to Embodiment 2 of the present invention; Figure 3 is a flow chart of a three-dimensional object recognition method based on recursive gated convolution according to Embodiment 1 of the present invention; Figure 4 This is a structural diagram of obtaining a two-dimensional view in Embodiment 1 of the present invention; Figure 5 It is a structural diagram of a high-order recursive gated convolution module according to Embodiment 1 of the present invention. DETAILED DESCRIPTION

[0023] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0024] In the description of the present invention, it should be noted that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0025] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0026] Furthermore, in the description of the present invention, unless otherwise specified, “plurality” means two or more.

[0027] Example 1 like Figure 1 and 3 As shown, a three-dimensional object recognition method based on recursive gated convolution in a preferred embodiment of the present invention includes: S1: Acquire multiple two-dimensional views of the three-dimensional object to be identified; like Figure 4 As shown, from The 3D object to be identified is projected by equidistant viewpoints. The height of the viewpoints is 30 degrees relative to the horizontal plane. Finally, the 3D object to be identified can obtain a set of A view, represented by .

[0028] S2: Input multiple 2D views into the convolutional neural network to obtain multi-scale view features of each view, where the multi-scale view features include multi-scale local view features and global view features. In a feasible embodiment, the convolutional neural network selects a ResNet network. Input into the ResNet network, extract the feature maps of the output of the first 4 layers, the first 6 layers, and the first 9 layers in the network respectively; each view obtains three multi-view features of different scales , where and As a local view feature, As a global view feature.

[0029] S3: Input the multi-scale local view features of each 2D view into the local feature interaction module to obtain local interaction features; In a feasible embodiment, the local feature interaction module is a low-order recursive gated convolution, which inputs the multi-scale local view features of each two-dimensional view into the low-order recursive gated convolution to obtain local interaction features, including: Project local view features of different scales to the same dimension; The local view features after projection are multiplied element by element to obtain the local features, as follows:

[0030] in, and These are all linear projection operations. It is a local feature; The local features are input into a low-order recursive gated convolution to obtain local interaction features, including: ;

[0031]

[0032] in, and These are all linear projection operations. After projection, the channel becomes twice as large as before, and then it is evenly divided into and , It is a deep convolution operation. After convolution calculation, Again with Perform a dot product, is the result of the first-order feature interaction. It is a local interactive feature that enhances view communication.

[0033] S4: inputting the global view features of each two-dimensional view into a global feature optimization module to obtain a global optimization feature, wherein the global feature optimization module includes a preferred view module and a high-order recursive gated convolution module, wherein the preferred view module is used to obtain a preferred global view feature according to the global view features, and the high-order recursive gated convolution module is used to obtain a global optimization feature according to the preferred global view features; In a feasible embodiment, S4 includes: The global view features of each 2D view are divided into Group, where , is the number of viewpoints; Calculate the average gray value of each global view feature after classification , calculated as follows:

[0034] in, For the View eigenvalues, is the width of the view feature, is the height of the view feature; The variance of each global view feature is calculated based on the average grayscale value as follows:

[0035] According to the variance from The first several global view features in the group are selected as the preferred global view features. Specifically, the first several global view features are selected from each group according to the size of the variance value. View, after optimization The group view is represented as , ; The optimized global view features are input into the high-order recursive gated convolution module to obtain the globally optimized features.

[0036] It should be noted that the high-order recursive gated convolution module includes a first normalization layer, a recursive gated convolution, a second normalization layer and a multi-layer perceptron connected in sequence, and the global view features after optimization are Expressed as , The first normalization layer is used to normalize the optimized global view features to obtain the normalized global view features, which is specifically as follows:

[0037] in, is the normalization layer; Recursive gated convolution is used to obtain the interactive global view features based on the normalized global view features, as shown in the following formula:

[0038]

[0039]

[0040]

[0041]

[0042] in, and is a linear projection operation, is the dimension mapping function in the operation process to ensure the consistency of the dimension in the spatial interaction process. The number of channels is doubled and divided into four parts, namely and ,right Perform deep convolution processing to obtain , , Then they are input in sequence to interact spatially with adjacent features to achieve recursive operations. , and They are the recursive results obtained by the first to third order interactions of features, Through the projection layer Get the output of the recursive gated convolution .

[0043] The second normalization layer is used to obtain the normalized result according to the sum of the optimized global view features and the interactive global view features, specifically: ; The multilayer perceptron is used to add the normalized result to the global view feature after interaction to obtain the global optimization feature, as follows:

[0044] in, To optimize the features globally, the high-order recursive gated convolution module structure is shown in Figure 5 .

[0045] S5: Add the local interaction features and global view features corresponding to each view, and fuse the result of the addition with the global optimization features to obtain a 3D shape descriptor. In a feasible embodiment, S5 is specifically as follows:

[0046] in, is the adaptive average pooling operation, To average the view features, For fusion splicing operations, is the global view feature, is the local interaction feature, It is a global optimization feature.

[0047] S6: Perform recognition based on the three-dimensional shape descriptor to obtain a recognition result.

[0048] In a feasible embodiment, the three-dimensional shape descriptor is fed to a fully connected layer to identify and classify the three-dimensional object to obtain a recognition result.

[0049] Example 2 This embodiment applies a three-dimensional object recognition method based on recursive gated convolution proposed in Example 1, and conducts experiments on modelnet40, modelnet10 and shapenet55, three datasets widely used in three-dimensional object classification and retrieval tasks, to evaluate the performance of the method proposed in the present invention.

[0050] The modelnet40 dataset contains 40 categories such as airplanes, cars, and computers, with a total of 12,311 3D shapes, including 9,843 training samples and 2,468 test samples. The modelnet10 dataset, as a subset of modelnet40, has 10 categories and a total of 4,899 3D shapes. The shapenet55 dataset is divided into 55 categories, with a total of 51,162 3D models, with training sets, validation sets, and test sets accounting for 70%, 10%, and 20%, respectively.

[0051] The method proposed in this embodiment (Ours) is compared with other current view-based methods on modelnet40 and modelnet10. The method proposed in this embodiment inputs 12 views for processing, so it is compared with other methods that also input 12 views. The compared methods include: MVCNN, GVCNN, Ma et al., RotationNet and other methods. Instance-level classification accuracy and retrieval mAP are indicators for evaluating the performance of these methods. The specific results are shown in the following table:

[0052] It can be seen from the results in the above table that, whether on the modelnet40 or modelnet10 dataset, the method proposed in this embodiment has the highest instance-level classification accuracy, and the retrieval mAP also achieves excellent results, which verifies the effectiveness and superiority of the proposed method.

[0053] The method (Ours) proposed in this embodiment is compared with other view-based methods on shapenet55. The compared methods include View-GCN, CAR-Net and HMVCM. The evaluation criteria are divided into two versions: micro average (Micro) and macro average (Macro). The evaluation indicators include precision (P@N), recall rate (R@N), F-score (F1@N), average precision (mAP) and normalized discounted cumulative gain (NDCG). Micro average is to directly count the results of each sample, and then directly calculate the overall indicators of the entire data set. Macro average is to evaluate the performance of the entire data set by the average value of the indicators of each category. The specific results are shown in the following table:

[0054] As can be seen from the above table, the method proposed in Example 1 achieves the best results in three of the indicators. Except for the NDCG of Micro, the other indicators are also second only to the best performance, which verifies the strong competitiveness of the proposed method.

[0055] Example 3 like Figure 2 As shown, a three-dimensional object recognition system based on recursive gated convolution according to an embodiment of the present invention includes: The first acquisition module is used to acquire multiple two-dimensional views of the three-dimensional object to be identified; The second acquisition module is used to input multiple two-dimensional views into the convolutional neural network to obtain multi-scale view features of each view, where the multi-scale view features include multi-scale local view features and global view features; The third acquisition module is used to input the multi-scale local view features of each two-dimensional view into the local feature interaction module to obtain the local interaction features; A fourth acquisition module: used for inputting the global view features of each two-dimensional view into the global feature optimization module to obtain the global optimization features, the global feature optimization module includes a preferred view module and a high-order recursive gated convolution module, the preferred view module is used for obtaining the preferred global view features according to the global view features, and the high-order recursive gated convolution module is used for obtaining the global optimization features according to the preferred global view features; Fusion module: used to add the local interaction features and global view features corresponding to each view, and fuse the added results with the global optimization features to obtain a 3D shape descriptor; Recognition module: used to perform recognition based on three-dimensional shape descriptors and obtain recognition results.

[0056] It can be understood that the system of this embodiment corresponds to the method of the above-mentioned embodiment 1, and the options in the above-mentioned embodiment 1 are also applicable to this embodiment, so they will not be described repeatedly here.

[0057] Example 4 An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer-stored program is executed by a processor, the three-dimensional object recognition method based on recursive gated convolution is implemented.

[0058] In summary, the embodiment of the present invention provides a three-dimensional object recognition method and system based on recursive gated convolution, which extracts view features of two-dimensional views at multiple scales through a convolutional neural network, takes the extracted view features as a basis, and uses a local feature interaction module and a global feature optimization module to strengthen the view features from two aspects, respectively, locally and globally. For the local feature interaction module, the local features of multiple scales are input to realize the interaction of features in spatial positions using low-order recursive gated convolution, thereby enhancing the information exchange between local view features; the global feature optimization module includes a preferred view module and a high-order recursive gated convolution module, the preferred view module is used to obtain the preferred global view features according to the global view features, and the high-order recursive gated convolution module is used to obtain the global optimization features according to the preferred global view features, thereby enhancing the feature interaction relationship between views, and at the same time reducing the interference of similar feature duplication and single information in the view, thereby improving the accuracy of three-dimensional object recognition.

[0059] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.

Claims

1. A three-dimensional object recognition method based on recursive gated convolution, characterized in that: The method comprises the following steps: Acquire multiple two-dimensional views of the three-dimensional object to be identified; Inputting multiple two-dimensional views into a convolutional neural network to obtain multi-scale view features of each view, wherein the multi-scale view features include multi-scale local view features and global view features; Input the multi-scale local view features of each two-dimensional view into the local feature interaction module to obtain local interaction features; Inputting the global view features of each two-dimensional view into a global feature optimization module to obtain a global optimization feature, wherein the global feature optimization module includes a preferred view module and a high-order recursive gated convolution module, wherein the preferred view module is used to obtain a preferred global view feature according to the global view features, and the high-order recursive gated convolution module is used to obtain a global optimization feature according to the preferred global view features; Adding the local interaction features and the global view features corresponding to each view, and fusing the added result with the global optimization feature to obtain a three-dimensional shape descriptor; Recognition is performed based on the three-dimensional shape descriptor to obtain a recognition result.

2. The three-dimensional object recognition method based on recursive gated convolution according to claim 1, characterized in that: The obtaining of multiple two-dimensional views of the three-dimensional object to be identified includes: projecting the three-dimensional object to be identified from multiple equidistant viewpoints to obtain multiple two-dimensional views, wherein the height of the viewpoint is 30 degrees relative to the horizontal plane.

3. The three-dimensional object recognition method based on recursive gated convolution according to claim 1, characterized in that: The convolutional neural network is a ResNet network, and multiple two-dimensional views are input into the ResNet network to obtain multi-scale view features of each view.

4. The three-dimensional object recognition method based on recursive gated convolution according to claim 3, characterized in that: The local feature interaction module is a low-order recursive gated convolution, which inputs the multi-scale local view features of each two-dimensional view into the low-order recursive gated convolution to obtain local interaction features, including: Project local view features of different scales to the same dimension; The local view features after projection are multiplied element by element to obtain the local features, as follows: in, and These are all linear projection operations. It is a local feature; The local features are input into a low-order recursive gated convolution to obtain local interaction features.

5. The three-dimensional object recognition method based on recursive gated convolution according to claim 1, characterized in that: The step of inputting the global view features of each two-dimensional view into a global feature optimization module to obtain the global optimization features includes: The global view features of each 2D view are divided into Group, where , is the number of viewpoints; Calculate the average gray value of each global view feature after classification , calculated as follows: in, For the View eigenvalues, is the width of the view feature, is the height of the view feature; The variance of each global view feature is calculated based on the average grayscale value as follows: According to the variance from Select the first several global view features in the group as the preferred global view features; The optimized global view features are input into the high-order recursive gated convolution module to obtain the globally optimized features.

6. The three-dimensional object recognition method based on recursive gated convolution according to claim 5, characterized in that: The high-order recursive gated convolution module includes a first normalization layer, a recursive gated convolution, a second normalization layer and a multi-layer perceptron connected in sequence, wherein the first normalization layer is used to normalize the optimized global view features to obtain normalized global view features; The recursive gated convolution is used to obtain the global view features after interaction according to the normalized global view features; the second normalization layer is used to obtain the normalized result according to the sum of the optimized global view features and the global view features after interaction; the multilayer perceptron is used to process the normalized result, and the obtained result is added to the global view features after interaction to obtain the global optimization feature.

7. The three-dimensional object recognition method based on recursive gated convolution according to claim 1, characterized in that: The local interaction features and global view features corresponding to each view are added, and the result of the addition is fused with the global optimization feature to obtain a three-dimensional shape descriptor, as follows: in, is the adaptive average pooling operation, To average the view features, For fusion splicing operations, is the global view feature, is the local interaction feature, It is a global optimization feature.

8. The three-dimensional object recognition method based on recursive gated convolution according to claim 5, characterized in that: The performing recognition based on the three-dimensional shape descriptor to obtain a recognition result includes: inputting the three-dimensional shape descriptor into a fully connected layer to recognize the three-dimensional object.

9. A three-dimensional object recognition system based on recursive gated convolution, application claims Any one of the three-dimensional object recognition methods based on recursive gated convolution, characterized in that: The system comprises: The first acquisition module is used to acquire multiple two-dimensional views of the three-dimensional object to be identified; A second acquisition module is used to input multiple two-dimensional views into a convolutional neural network to obtain multi-scale view features of each view, wherein the multi-scale view features include multi-scale local view features and global view features; The third acquisition module is used to input the multi-scale local view features of each two-dimensional view into the local feature interaction module to obtain the local interaction features; A fourth acquisition module: used for inputting the global view features of each two-dimensional view into the global feature optimization module to obtain the global optimization features, wherein the global feature optimization module includes a preferred view module and a high-order recursive gated convolution module, wherein the preferred view module is used for obtaining the preferred global view features according to the global view features, and the high-order recursive gated convolution module is used for obtaining the global optimization features according to the preferred global view features; Fusion module: used to add the local interaction features and global view features corresponding to each view, and fuse the result of the addition with the global optimization feature to obtain a three-dimensional shape descriptor; Recognition module: used for performing recognition based on the three-dimensional shape descriptor to obtain a recognition result.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer stored program is executed by a processor, the three-dimensional object recognition method based on recursive gated convolution described in any one of claims 1 to 8 is implemented.