CN107092859A - A kind of depth characteristic extracting method of threedimensional model - Google Patents

Abstract

The present invention provides a method for extracting depth features of a three-dimensional model. First, the polar view of the three-dimensional model is extracted as training input data for a deep convolutional neural network; secondly, a deep convolutional neural network is constructed and the polar view is trained; again , input the polar view into the deep convolutional neural network for training until the deep convolutional neural network converges, and realize the determination of the internal weights of the deep convolutional neural network after the training is completed; finally, in the trained deep convolutional neural network Input the polar view of the 3D model whose features need to be extracted, and calculate the feature vector of the fully connected layer in the deep convolutional neural network, which is used as the depth feature of the 3D model whose features need to be extracted. The present invention builds a deep convolutional neural network and reduces residual errors through iterative correction of weights, so that the network converges. After the training is completed, the fully connected layer of the convolutional neural network is extracted as the depth feature of the extreme view of the 3D model.

Description

A deep feature extraction method for 3D models

technical field

The invention relates to the technical field of three-dimensional model processing, and more specifically, relates to a deep feature extraction method of a three-dimensional model.

Background technique

With the rapid development of 3D model processing technology and computer hardware and software, as well as the promotion of multimedia technology and Internet technology, a large number of 3D models are used in various fields, and people's demand for 3D model applications is also increasing. 3D models play an important role in many fields such as e-commerce, architectural design, industrial design, advertising film and television, and 3D games. The 3D models of large-scale data sets need to reuse design and model retrieval in all aspects of social production and life. Therefore, how to quickly and accurately retrieve the target 3D model from the existing various types of 3D model data sets has become the key to be solved urgently. question.

In recent years, 3D model analysis based on deep learning has become a research hotspot. It combines computer vision, artificial intelligence, intelligent computing and other research content, and can solve the visual tasks of 3D models, including 3D model feature extraction, classification, recognition, detection and prediction, etc. problem. The use of deep learning technology can automatically learn the hidden features of the 3D model, and can be trained on a large-scale data set to enhance the generalization ability of the learning model.

The current feature extraction method based on deep learning has the following problems: such as the features extracted by the deep learning framework cannot fully express the 3D model information, the high computational complexity brought by the deep network layer and the over-fitting problem of the network, and the network learning time Long and large memory storage space, etc. With the mature application of deep learning technology and the strong demand for the ability to express 3D model features, the use of deep learning to extract features will bring new breakthroughs in the classification, retrieval, detection and identification of 3D models.

Contents of the invention

The purpose of the present invention is to overcome the shortcomings and deficiencies in the prior art, and provide a method for extracting deep features of a three-dimensional model. The method constructs a deep convolutional neural network to train the extreme view of the three-dimensional model, and reduces the residual value after iterative correction of weights. Poor, making the network converge. After the training is completed, the fully connected layer of the convolutional neural network is extracted as the depth feature of the extreme view of the 3D model, so that the depth feature can be used for visual tasks such as classification, retrieval and recognition of the 3D model. This method constructs a deep convolutional neural network with rich layers, which can speed up network training and improve the accuracy of deep network fitting.

In order to achieve the above object, the present invention is achieved through the following technical solutions: a method for extracting depth features of a three-dimensional model, characterized in that:

First, extract the polar view of the 3D model as the training input data for the deep convolutional neural network;

Secondly, construct a deep convolutional neural network and train the polar view; among them, the deep convolutional neural network includes an input layer with the polar view as the training input data, which is used to learn the features of the polar view and obtain a two-dimensional feature map The convolutional layer, the pooling layer used to aggregate the two-dimensional feature maps of different positions and reduce the feature dimension, the fully connected layer used to arrange and link the two-dimensional feature maps to form a one-dimensional vector, and the output The output layer of the category prediction result;

Again, input the polar view into the deep convolutional neural network for training until the deep convolutional neural network converges, and realize the determination of the internal weights of the deep convolutional neural network after the training is completed;

Finally, input the polar view of the 3D model that needs to be extracted into the trained deep convolutional neural network, and calculate the feature vector of the fully connected layer in the deep convolutional neural network, which is used as the depth feature of the 3D model that needs to be extracted.

In the above solution, the deep feature extraction method of the 3D model of the present invention is to train the polar view of the 3D model through a deep convolutional neural network, and iteratively corrects the weights to reduce the residual, so that the network converges. After the training is completed, the fully connected layer of the convolutional neural network is extracted as the depth feature of the extreme view of the 3D model. Among them, the polar view is a global expression of the spatial collection structure of the 3D model, and the use of the polar view of the 3D model can simplify and reduce the training calculation of the deep convolutional neural network. The deep convolutional neural network constructed by the invention has rich layers, which can speed up the network training speed and improve the fitting accuracy of the deep network. Among them, the polar view refers to a group of sampling rays emitted from the center of mass of the three-dimensional model, and a two-dimensional sampling map formed by the distance from the intersection of the rays and the model to the center of mass.

Specifically, the method includes the following steps:

Step s101: Extract the polar view of the three-dimensional model as the training input data of the deep convolutional neural network, wherein the training input data is set to x ⁽ⁱ⁾ ∈ χ, where χ is the polar view of N three-dimensional models in total; the i-th model corresponds to The category label of y ⁽ⁱ⁾ ∈ {1,2,...,K}, K is the number of categories of the three-dimensional model;

Step s102: build a deep convolutional neural network; in the deep convolutional neural network, the deep convolutional neural network includes: polar view x ⁽ⁱ⁾ as the input layer I of training input data, 4 convolutional layers C (t), t=1,2,3,4, 1 pooling layer P, two fully connected layers FC(1) and FC(2), and output layer O; among them, 4 convolutional layers C(t), t = Each convolutional layer in 1, 2, 3, 4 and each fully connected layer in the two fully connected layers FC(1) and FC(2) use a modified linear activation function f(a)=max(0, a) replace the sigmoid function f(a)=1/(1+e ^-a ) in the deep convolutional neural network;

Step s103: Set the parameters of the deep convolutional neural network, that is, initialize the weights of each layer of the deep convolutional neural network:

Input layer Ι: the input data is a polar view x ^{(i) whose size is (32×32)} ;

Convolutional layer C: 4 convolutional layers learn the features of the polar view in turn, and the number of two-dimensional feature maps obtained by each convolutional layer is F ^t = (6,8,10,12), and each is obtained according to the following formula The two-dimensional feature map of the convolutional layer:

in Represents the qth two-dimensional feature map of C(t), M represents the feature map set of the t-1 layer, when t-1 is 0, it means that the two-dimensional feature map is the polar view of the input, Represents the convolution kernel of the pth feature map of the t-1th layer to the qth two-dimensional feature map of the tth layer convolutional layer, a matrix of (5×5) generated by random numbers of [-1,1] as the initial convolution kernel. bias is the bias, the initial value is 0. (*) represents the operation of convolution, f(·) is the modified linear activation function; through the above formula, it can be obtained:

The convolutional layer C(1) calculates 6 two-dimensional feature maps with a size of (28×28);

The convolutional layer C(2) calculates 8 two-dimensional feature maps with a size of (24×24);

The convolutional layer C (3) calculates 10 two-dimensional feature maps with a size of (20×20);

The convolutional layer C (4) calculates 12 two-dimensional feature maps with a size of (16×16);

Pooling layer P: Perform maximum pooling processing on the two-dimensional feature map calculated by the convolutional layer C(4) by the following formula to obtain 12 feature matrices with a size of (8×8):

Among them, PM _p (u ₀ , v ₀ ) represents the coordinates corresponding to the feature matrix of the two-dimensional feature map maximum pooling process, is the two-dimensional feature map calculated by the convolutional layer C(4), and max() is the maximum value of the calculation matrix elements;

Fully-connected layer FC(1): Arrange PM _p into column vectors (64×1), and fully link the column vectors of each matrix to obtain one-dimensional vector L ⁰ (768×1), that is, one-dimensional vector L ⁰ (768× 1) as the input vector of the fully connected layer FC(1);

Fully connected layer FC(2): set the fully connected layer FC(1) to have 512 neurons, and the fully connected layer FC(2) to have 128 neurons; and calculate the fully connected layer FC by the following fully connected layer conduction formula The output vector L ¹ of (1) is used as the input vector of the fully connected layer FC(2), and the output vector L ² of the fully connected layer FC(2) is calculated as the input vector of the output layer O:

in for and Network weights, a matrix of (512×768) generated by a random number of [-1,1] is used as the initial weight W ¹ , and a matrix of (128×512) is generated as the initial weight W ² ; for The bias, the initial value is 0; f(·) is the modified linear activation function; Take 1 and 2;

Output layer O: Set the output layer to have K neurons, the calculation formula is as follows:

y'=f(W ³ L ² +b ² );

Among them, y′ ⁽ⁱ⁾ is the final output of the deep convolutional neural network, and i is the i-th model; W ³ is a matrix of (K×128) generated by random numbers [-1,1] as the initial weight; L ² is the output vector of the fully connected layer FC(2);

Step s104: train the polar view data set χ, set the learning rate η to 1, use the gradient random descent algorithm, and output the predicted result y′ ⁽ⁱ⁾ and the real category label y ⁽ⁱ⁾ according to the deep convolutional neural network The error of ∈{1,2,...,K} is backpropagated, and the algorithm can converge after 20 iterations, realizing the determination of the internal weights of the deep convolutional neural network after training;

Step s105: Input the polar view x ⁽ⁱ⁾ of the 3D model to be extracted into the trained deep convolutional neural network, and calculate the feature vector L ² output by the second fully connected layer FC(2), that is Depth features of the polar view of the 3D model for which features are to be extracted.

The polar view of the extracted three-dimensional model refers to:

First, preprocess the 3D point cloud model, calculate the centroid and scale of the 3D point cloud model, and translate the 3D point cloud model to the Cartesian coordinate system for scaling, so as to realize the normalization of the 3D point cloud model on the Cartesian coordinate system ;

Secondly, convert the scaled 3D point cloud model on the Cartesian coordinate system to spherical coordinates, and obtain the direction and distance attributes of each point of the 3D point cloud model;

Again, map the spherical coordinates of the point set to the pixel position of the polar view, calculate the maximum distance of each pixel sampling distance set, and use it as the ray sampling value of the direction interval;

Finally, the maximum distance of each pixel sampling distance set is arranged into a two-dimensional sampling map, which is the extracted polar view.

The extraction of the polar view of the three-dimensional model specifically includes the following steps:

Step S201: Input a 3D point cloud model, the scale of which is P={p _i ( _xi , y _i , z _i )|i=1,2,...,N};

Step S202: Calculate the centroid g(g _x , g _{y , g z ) of the 3D point cloud model according to the following formula, and translate the 3D point cloud model to a right angle through the obtained centroid g(g x , g y , g z )} coordinate system, the scale of the 3D point cloud model in the Cartesian coordinate system is p _i ′= _pi -g, i=1,2,...,N; the centroid of the 3D point cloud model p _i ′ after translation transformation at the origin of the Cartesian coordinate system;

Step s203: Calculate the scaling factor s of the 3D point cloud model, and scale the 3D point cloud model to a unit scale of , where the scaling factor s is

Step s204: Convert the scaled 3D point cloud model on the Cartesian coordinate system to the spherical coordinate Q. At this time, the scale of the 3D point cloud model on the spherical coordinate is The conversion formula is as follows:

Where θ∈[0,π], the elevation angle is 0 on the negative half axis of the Z axis;

Step s205: Mapping the spherical coordinate Q to the pixel position (u, v) of the polar view, the mapping relationship is Then the 3D point cloud model is calculated according to the following formula at the pixel position (u, v) of the polar view; where there is one spherical coordinate point, multiple spherical coordinate points or no spherical coordinate point at a pixel position (u, v) Coordinate points;

where n _u and n _v are the width and length of the polar view, respectively;

Step s206: The sampling distance set of each pixel (u, v) is Calculate the maximum value of each pixel sampling distance set according to the following formula as the maximum distance to obtain the pixel sampling value in the polar view, and arrange it into a two-dimensional acquisition image as the polar view I;

The above-mentioned depth feature extraction method directly performs feature extraction on the polar view of the 3D model, and the information of each position can be perceived by using the convolution operation for the two-dimensional polar view. The feature map obtained through multi-layer convolution processing can extract highly discriminative deep features in the fully connected layer.

When extracting the polar view, the preprocessing process needs to translate and scale the 3D model to ensure that the 3D model is normalized and standardized on the standard scale. Converting the 3D model point cloud into a spherical coordinate system is beneficial to map the spherical coordinate point to the corresponding pixel position of the 2D polar view. Through the mapping, the maximum value of the distance set of the point at the pixel position is counted, and the maximum sampling value is formed into a binary Dimensional sampling map, which is a new type of polar view for 3D models.

Compared with the prior art, the present invention has the following advantages and beneficial effects: the deep feature extraction method of the three-dimensional model of the present invention constructs a deep convolutional neural network to train the extreme view of the three-dimensional model, and the residual is reduced through iterative correction of the weight, so that Network convergence. After the training is completed, the fully connected layer of the convolutional neural network is extracted as the depth feature of the extreme view of the 3D model, so that the depth feature can be used for visual tasks such as classification, retrieval and recognition of the 3D model. This method constructs a deep convolutional neural network with rich layers, which can speed up network training and improve the accuracy of deep network fitting.

Description of drawings

Fig. 1 is the flow chart of the depth feature extraction method of three-dimensional model of the present invention;

2 is a schematic diagram of a deep convolutional neural network in the depth feature extraction method of a three-dimensional model of the present invention;

Fig. 3 is the flowchart of extracting the extreme view of three-dimensional model in the method of the present invention;

Fig. 4 is the schematic diagram that extracts and obtains polar view by three-dimensional model in the method of the present invention;

detailed description

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Example

As shown in Figures 1 to 4, the depth feature extraction method of the three-dimensional model of the present invention is as follows:

First, extract the polar view of the 3D model as the training input data for the deep convolutional neural network;

Secondly, construct a deep convolutional neural network and train the polar view; among them, the deep convolutional neural network includes an input layer with the polar view as the training input data, which is used to learn the features of the polar view and obtain a two-dimensional feature map The convolutional layer, the pooling layer used to aggregate the two-dimensional feature maps of different positions and reduce the feature dimension, the fully connected layer used to arrange and link the two-dimensional feature maps to form a one-dimensional vector, and the output The output layer of the category prediction result;

Again, input the polar view into the deep convolutional neural network for training until the deep convolutional neural network converges, and realize the determination of the internal weights of the deep convolutional neural network after the training is completed;

Finally, input the polar view of the 3D model that needs to be extracted into the trained deep convolutional neural network, and calculate the feature vector of the fully connected layer in the deep convolutional neural network, which is used as the depth feature of the 3D model that needs to be extracted.

In the above solution, the deep feature extraction method of the 3D model of the present invention is to train the polar view of the 3D model through a deep convolutional neural network, and to reduce the residual after iterative correction of the weights, so that the network converges. After the training is completed, the fully connected layer of the convolutional neural network is extracted as the depth feature of the extreme view of the 3D model. Among them, the polar view is a global expression of the spatial collection structure of the 3D model, and the use of the polar view of the 3D model can simplify and reduce the training calculation of the deep convolutional neural network. The deep convolutional neural network constructed by the invention has rich layers, which can speed up the network training speed and improve the fitting accuracy of the deep network. Among them, the polar view refers to a group of sampling rays emitted from the center of mass of the three-dimensional model, and a two-dimensional sampling map formed by the distance from the intersection of the rays and the model to the center of mass.

Specifically, the method includes the following steps:

Step s101: Extract the polar view of the three-dimensional model as the training input data of the deep convolutional neural network, wherein the training input data is set to x ⁽ⁱ⁾ ∈ χ, where χ is the polar view of N three-dimensional models in total; the i-th model corresponds to The category label of y ⁽ⁱ⁾ ∈ {1,2,...,K}, K is the number of categories of the three-dimensional model;

Step s102: build a deep convolutional neural network; in the deep convolutional neural network, the deep convolutional neural network includes: polar view x ⁽ⁱ⁾ as the input layer I of training input data, 4 convolutional layers C (t), t=1,2,3,4, 1 pooling layer P, two fully connected layers FC(1) and FC(2), and output layer O; among them, 4 convolutional layers C(t), t = Each convolutional layer in 1, 2, 3, 4 and each fully connected layer in the two fully connected layers FC(1) and FC(2) use a modified linear activation function f(a)=max(0, a) replace the sigmoid function f(a)=1/(1+e ^-a ) in the deep convolutional neural network;

Step s103: Set the parameters of the deep convolutional neural network, that is, initialize the weights of each layer of the deep convolutional neural network:

Input layer Ι: the input data is a polar view x ^{(i) whose size is (32×32)} ;

Convolutional layer C: 4 convolutional layers learn the features of the polar view in turn, and the number of two-dimensional feature maps obtained by each convolutional layer is F ^t = (6,8,10,12), and each is obtained according to the following formula The two-dimensional feature map of the convolutional layer:

in Represents the qth two-dimensional feature map of C(t), M represents the feature map set of the t-1 layer, when t-1 is 0, it means that the two-dimensional feature map is the polar view of the input, Represents the convolution kernel of the pth feature map of the t-1th layer to the qth two-dimensional feature map of the tth layer convolutional layer, a matrix of (5×5) generated by random numbers of [-1,1] as the initial convolution kernel. bias is the bias, the initial value is 0. (*) represents the operation of convolution, f(·) is the modified linear activation function; through the above formula, it can be obtained:

The convolutional layer C(1) calculates 6 two-dimensional feature maps with a size of (28×28);

The convolutional layer C(2) calculates 8 two-dimensional feature maps with a size of (24×24);

The convolutional layer C (3) calculates 10 two-dimensional feature maps with a size of (20×20);

The convolutional layer C (4) calculates 12 two-dimensional feature maps with a size of (16×16);

Pooling layer P: Perform maximum pooling processing on the two-dimensional feature map calculated by the convolutional layer C(4) by the following formula to obtain 12 feature matrices with a size of (8×8):

Among them, PM _p (u ₀ , v ₀ ) represents the coordinates corresponding to the feature matrix of the two-dimensional feature map maximum pooling process, is the two-dimensional feature map calculated by the convolutional layer C(4), and max() is the maximum value of the calculation matrix elements;

Fully-connected layer FC(1): Arrange PM _p into column vectors (64×1), and fully link the column vectors of each matrix to obtain one-dimensional vector L ⁰ (768×1), that is, one-dimensional vector L ⁰ (768× 1) as the input vector of the fully connected layer FC(1);

Fully connected layer FC(2): set the fully connected layer FC(1) to have 512 neurons, and the fully connected layer FC(2) to have 128 neurons; and calculate the fully connected layer FC by the following fully connected layer conduction formula The output vector L ¹ of (1) is used as the input vector of the fully connected layer FC(2), and the output vector L ² of the fully connected layer FC(2) is calculated as the input vector of the output layer O:

in for and Network weights, a matrix of (512×768) generated by a random number of [-1,1] is used as the initial weight W ¹ , and a matrix of (128×512) is generated as the initial weight W ² ; for The bias, the initial value is 0; f(·) is the modified linear activation function; Take 1 and 2;

Output layer O: Set the output layer to have K neurons, the calculation formula is as follows:

y'=f(W ³ L ² +b ² );

Among them, y′ ⁽ⁱ⁾ is the final output of the deep convolutional neural network, and i is the i-th model; W ³ is a matrix of (K×128) generated by random numbers [-1,1] as the initial weight; L ² is the output vector of the fully connected layer FC(2);

Step s104: train the polar view data set χ, set the learning rate η to 1, use the gradient random descent algorithm, and output the predicted result y′ ⁽ⁱ⁾ and the real category label y ⁽ⁱ⁾ according to the deep convolutional neural network The error of ∈{1,2,...,K} is backpropagated, and the algorithm can converge after 20 iterations, realizing the determination of the internal weights of the deep convolutional neural network after training;

Step s105: Input the polar view x ⁽ⁱ⁾ of the 3D model to be extracted into the trained deep convolutional neural network, and calculate the feature vector L ² output by the second fully connected layer FC(2), that is Depth features of the polar view of the 3D model for which features are to be extracted.

The polar view of the above-mentioned extracted 3D model refers to:

First, preprocess the 3D point cloud model, calculate the centroid and scale of the 3D point cloud model, and translate the 3D point cloud model to the Cartesian coordinate system for scaling, so as to realize the normalization of the 3D point cloud model on the Cartesian coordinate system ;

Secondly, convert the scaled 3D point cloud model on the Cartesian coordinate system to spherical coordinates, and obtain the direction and distance attributes of each point of the 3D point cloud model;

Again, map the spherical coordinates of the point set to the pixel position of the polar view, calculate the maximum distance of each pixel sampling distance set, and use it as the ray sampling value of the direction interval;

Finally, the maximum distance of each pixel sampling distance set is arranged into a two-dimensional sampling map, which is the extracted polar view.

Specifically, extracting the polar view of the 3D model specifically includes the following steps:

Step S201: Input a 3D point cloud model, the scale of which is P={p _i ( _xi , y _i , z _i )|i=1,2,...,N};

Step S202: Calculate the centroid g(g _x , g _{y , g z ) of the 3D point cloud model according to the following formula, and translate the 3D point cloud model to a right angle through the obtained centroid g(g x , g y , g z )} coordinate system, the scale of the 3D point cloud model in the Cartesian coordinate system is p _i ′= _pi -g, i=1,2,...,N; the centroid of the 3D point cloud model p _i ′ after translation transformation at the origin of the Cartesian coordinate system;

Step s203: Calculate the scaling factor s of the 3D point cloud model, and scale the 3D point cloud model to a unit scale of , where the scaling factor s is

Step s204: Convert the scaled 3D point cloud model on the Cartesian coordinate system to the spherical coordinate Q. At this time, the scale of the 3D point cloud model on the spherical coordinate is The conversion formula is as follows:

Where θ∈[0,π], the elevation angle is 0 on the negative half axis of the Z axis;

Step s205: Mapping the spherical coordinate Q to the pixel position (u, v) of the polar view, the mapping relationship is Then the 3D point cloud model is calculated according to the following formula at the pixel position (u, v) of the polar view; where there is one spherical coordinate point, multiple spherical coordinate points or no spherical coordinate point at a pixel position (u, v) Coordinate points;

where n _u and n _v are the width and length of the polar view, respectively;

Step s206: The sampling distance set of each pixel (u, v) is Calculate the maximum value of each pixel sampling distance set according to the following formula as the maximum distance to obtain the pixel sampling value in the polar view, and arrange it into a two-dimensional acquisition image as the polar view I;

The above-mentioned depth feature extraction method directly performs feature extraction on the polar view of the 3D model, and the information of each position can be perceived by using the convolution operation for the two-dimensional polar view. The feature map obtained through multi-layer convolution processing can extract highly discriminative deep features in the fully connected layer.

When extracting the polar view, the preprocessing process needs to translate and scale the 3D model to ensure that the 3D model is normalized and standardized on the standard scale. Converting the 3D model point cloud into a spherical coordinate system is beneficial to map the spherical coordinate point to the corresponding pixel position of the 2D polar view. Through the mapping, the maximum value of the distance set of the point at the pixel position is counted, and the maximum sampling value is formed into a binary Dimensional sampling map, which is a new type of polar view for 3D models.

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims (4)

1. A depth feature extraction method of a three-dimensional model, characterized in that: First, extract the polar view of the 3D model as the training input data for the deep convolutional neural network; Secondly, construct a deep convolutional neural network and train the polar view; among them, the deep convolutional neural network includes an input layer with the polar view as the training input data, which is used to learn the features of the polar view and obtain a two-dimensional feature map The convolutional layer, the pooling layer used to aggregate the two-dimensional feature maps of different positions and reduce the feature dimension, the fully connected layer used to arrange and link the two-dimensional feature maps to form a one-dimensional vector, and the output The output layer of the category prediction result; Again, input the polar view into the deep convolutional neural network for training until the deep convolutional neural network converges, and realize the determination of the internal weights of the deep convolutional neural network after the training is completed; Finally, input the polar view of the 3D model that needs to be extracted into the trained deep convolutional neural network, and calculate the feature vector of the fully connected layer in the deep convolutional neural network, which is used as the depth feature of the 3D model that needs to be extracted. 2. the depth feature extraction method of three-dimensional model according to claim 1, is characterized in that: comprise the following steps: Step s101: Extract the polar view of the three-dimensional model as the training input data of the deep convolutional neural network, wherein the training input data is set to x ⁽ⁱ⁾ ∈ χ, where χ is the polar view of N three-dimensional models in total; the i-th model corresponds to The category label of y ⁽ⁱ⁾ ∈ {1,2,...,K}, K is the number of categories of the three-dimensional model; Step s102: build a deep convolutional neural network; in the deep convolutional neural network, the deep convolutional neural network includes: polar view x ⁽ⁱ⁾ as the input layer I of training input data, 4 convolutional layers C (t), t=1,2,3,4, 1 pooling layer P, two fully connected layers FC(1) and FC(2), and output layer O; among them, 4 convolutional layers C(t), t = Each convolutional layer in 1, 2, 3, 4 and each fully connected layer in the two fully connected layers FC(1) and FC(2) use a modified linear activation function f(a)=max(0, a) replace the sigmoid function f(a)=1/(1+e ^-a ) in the deep convolutional neural network; Step s103: Set the parameters of the deep convolutional neural network, that is, initialize the weights of each layer of the deep convolutional neural network: Input layer Ι: the input data is a polar view x ^{(i) whose size is (32×32)} ; Convolutional layer C: 4 convolutional layers learn the features of the polar view in turn, and the number of two-dimensional feature maps obtained by each convolutional layer is F ^t = (6,8,10,12), and each is obtained according to the following formula The two-dimensional feature map of the convolutional layer: <mrow> <msubsup> <mi>x</mi> <mi>q</mi> <mi>t</mi> </msubsup> <mo>=</mo> <mi>f</mi> <mrow> <mo>(</mo> <munder> <mi>&amp;Sigma;</mi> <mrow> <mi>x</mi> <mo>&amp;Element;</mo> <mi>M</mi> </mrow> </munder> <msubsup> <mi>x</mi> <mi>p</mi> <mrow> <mi>t</mi> <mo>-</mo> <mn>1</mn> </mrow> </msubsup> <mo>*</mo> <msubsup> <mi>k</mi> <mrow> <mi>p</mi> <mi>q</mi> </mrow> <mi>t</mi> </msubsup> <mo>+</mo> <msubsup> <mi>bias</mi> <mi>q</mi> <mi>t</mi> </msubsup> <mo>)</mo> </mrow> </mrow> in Represents the qth two-dimensional feature map of C(t), M represents the feature map set of the t-1 layer, when t-1 is 0, it means that the two-dimensional feature map is the polar view of the input, Represents the convolution kernel of the pth feature map of the t-1th layer to the qth two-dimensional feature map of the tth layer convolutional layer, a matrix of (5×5) generated by random numbers of [-1,1] as the initial convolution kernel. bias is the bias, the initial value is 0. (*) represents the operation of convolution, f(·) is the modified linear activation function; through the above formula, it can be obtained: The convolutional layer C(1) calculates 6 two-dimensional feature maps with a size of (28×28); The convolutional layer C(2) calculates 8 two-dimensional feature maps with a size of (24×24); The convolutional layer C (3) calculates 10 two-dimensional feature maps with a size of (20×20); The convolutional layer C (4) calculates 12 two-dimensional feature maps with a size of (16×16); Pooling layer P: Perform maximum pooling processing on the two-dimensional feature map calculated by the convolutional layer C(4) by the following formula to obtain 12 feature matrices with a size of (8×8): <mrow> <msub> <mi>PM</mi> <mi>p</mi> </msub> <mrow> <mo>(</mo> <mrow> <msub> <mi>u</mi> <mn>0</mn> </msub> <mo>,</mo> <msub> <mi>v</mi> <mn>0</mn> </msub> </mrow> <mo>)</mo> </mrow> <mo>=</mo> <mi>max</mi> <mfenced open = "(" close = ")"> <mtable> <mtr> <mtd> <mrow> <msubsup> <mi>x</mi> <mi>p</mi> <mn>4</mn> </msubsup> <mrow> <mo>(</mo> <mrow> <mn>2</mn> <msub> <mi>u</mi> <mn>0</mn> </msub> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mn>2</mn> <msub> <mi>v</mi> <mn>0</mn> </msub> <mo>-</mo> <mn>1</mn> </mrow> <mo>)</mo> </mrow> </mrow> </mtd> <mtd> <mrow> <msubsup> <mi>x</mi> <mi>p</mi> <mn>4</mn> </msubsup> <mrow> <mo>(</mo> <mrow> <mn>2</mn> <msub> <mi>u</mi> <mn>0</mn> </msub> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mn>2</mn> <msub> <mi>v</mi> <mn>0</mn> </msub> </mrow> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msubsup> <mi>x</mi> <mi>p</mi> <mn>4</mn> </msubsup> <mrow> <mo>(</mo> <mrow> <mn>2</mn> <msub> <mi>u</mi> <mn>0</mn> </msub> <mo>,</mo> <mn>2</mn> <msub> <mi>v</mi> <mn>0</mn> </msub> <mo>-</mo> <mn>1</mn> </mrow> <mo>)</mo> </mrow> </mrow> </mtd> <mtd> <mrow> <msubsup> <mi>x</mi> <mi>p</mi> <mn>4</mn> </msubsup> <mrow> <mo>(</mo> <mrow> <mn>2</mn> <msub> <mi>u</mi> <mn>0</mn> </msub> <mo>,</mo> <mn>2</mn> <msub> <mi>v</mi> <mn>0</mn> </msub> </mrow> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> <mi>p</mi> <mo>=</mo> <mn>1</mn> <mo>,</mo> <mn>2</mn> <mo>,</mo> <mn>...</mn> <mo>,</mo> <msup> <mi>F</mi> <mn>4</mn> </msup> <mo>;</mo> </mrow> Among them, PM _p (u ₀ , v ₀ ) represents the coordinates corresponding to the feature matrix of the two-dimensional feature map maximum pooling process, is the two-dimensional feature map calculated by the convolutional layer C(4), and max() is the maximum value of the calculation matrix elements; Fully-connected layer FC(1): Arrange PM _p into column vectors (64×1), and fully link the column vectors of each matrix to obtain one-dimensional vector L ⁰ (768×1), that is, one-dimensional vector L ⁰ (768× 1) as the input vector of the fully connected layer FC(1); Fully connected layer FC(2): set the fully connected layer FC(1) to have 512 neurons, and the fully connected layer FC(2) to have 128 neurons; and calculate the fully connected layer FC by the following fully connected layer conduction formula The output vector L ¹ of (1) is used as the input vector of the fully connected layer FC(2), and the output vector L ² of the fully connected layer FC(2) is calculated as the input vector of the output layer O: L ^l ＝f(W ^l L ^l-1 +b ^l-1 ) Among them, W ^l is the network weight of FC(l-1) and FC(l), and the matrix of (512×768) generated by the random number of [-1,1] is used as the initial weight W ¹ to generate (128×512) The matrix of is used as the initial weight W ² ; b ^l-1 is the bias of FC(l-1), the initial value is 0; f(·) is the modified linear activation function; l takes 1 and 2; Output layer O: Set the output layer to have K neurons, the calculation formula is as follows: y'=f(W ³ L ² +b ² ); Among them, y′ ⁽ⁱ⁾ is the final output of the deep convolutional neural network, and i is the i-th model; W ³ is a matrix of (K×128) generated by random numbers [-1,1] as the initial weight; L ² is the output vector of the fully connected layer FC(2); Step s104: train the polar view data set χ, set the learning rate η to 1, use the gradient random descent algorithm, and output the predicted result y′ ⁽ⁱ⁾ and the real category label y ⁽ⁱ⁾ according to the deep convolutional neural network The error of ∈{1,2,...,K} is backpropagated, and the algorithm can converge after 20 iterations, realizing the determination of the internal weights of the deep convolutional neural network after training; Step s105: Input the polar view x ⁽ⁱ⁾ of the 3D model to be extracted into the trained deep convolutional neural network, and calculate the feature vector L ² output by the second fully connected layer FC(2), that is Depth features of the polar view of the 3D model for which features are to be extracted. 3. the depth feature extraction method of three-dimensional model according to claim 1, is characterized in that: the polar view of described extraction three-dimensional model refers to: First, preprocess the 3D point cloud model, calculate the centroid and scale of the 3D point cloud model, and translate the 3D point cloud model to the Cartesian coordinate system for scaling, so as to realize the normalization of the 3D point cloud model on the Cartesian coordinate system ; Secondly, convert the scaled 3D point cloud model on the Cartesian coordinate system to spherical coordinates, and obtain the direction and distance attributes of each point of the 3D point cloud model; Again, map the spherical coordinates of the point set to the pixel position of the polar view, calculate the maximum distance of each pixel sampling distance set, and use it as the ray sampling value of the direction interval; Finally, the maximum distance of each pixel sampling distance set is arranged into a two-dimensional sampling map, which is the extracted polar view. 4. the depth feature extraction method of three-dimensional model according to claim 3, is characterized in that: the polar view of described extraction three-dimensional model specifically comprises the following steps: Step S201: Input a 3D point cloud model, the scale of which is P={p _i ( _xi , y _i , z _i )|i=1,2,...,N}; Step S202: Calculate the centroid g(g _x , g _{y , g z ) of the 3D point cloud model according to the following formula, and translate the 3D point cloud model to a right angle through the obtained centroid g(g x , g y , g z )} coordinate system, the scale of the 3D point cloud model in the Cartesian coordinate system is p′ _i =p _i -g,i=1,2,...,N; the centroid of the 3D point cloud model p _i ′ after translation transformation at the origin of the Cartesian coordinate system; <mrow> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <msub> <mi>g</mi> <mi>x</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mi>N</mi> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <msub> <mi>x</mi> <mi>i</mi> </msub> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>g</mi> <mi>y</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mi>N</mi> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <msub> <mi>y</mi> <mi>i</mi> </msub> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>g</mi> <mi>z</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mi>N</mi> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <msub> <mi>x</mi> <mi>z</mi> </msub> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>;</mo> </mrow> Step s203: Calculate the scaling factor s of the 3D point cloud model, and scale the 3D point cloud model to a unit scale of , where the scaling factor s is <mrow> <mi>s</mi> <mo>=</mo> <mi>m</mi> <mi>a</mi> <mi>x</mi> <mrow> <mo>(</mo> <mo>|</mo> <mo>|</mo> <msubsup> <mi>p</mi> <mi>i</mi> <mo>&amp;prime;</mo> </msubsup> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> <mo>)</mo> </mrow> <mo>;</mo> </mrow> Step s204: Convert the scaled 3D point cloud model on the Cartesian coordinate system to the spherical coordinate Q. At this time, the scale of the 3D point cloud model on the spherical coordinate is The conversion formula is as follows: Where θ∈[0,π], the elevation angle is 0 on the negative half axis of the Z axis; Step s205: Mapping the spherical coordinate Q to the pixel position (u, v) of the polar view, the mapping relationship is Then the 3D point cloud model is calculated according to the following formula at the pixel position (u, v) of the polar view; where there is one spherical coordinate point, multiple spherical coordinate points or no spherical coordinate point at a pixel position (u, v) Coordinate points; where n _u and n _v are the width and length of the polar view, respectively; Step s206: The sampling distance set of each pixel (u, v) is Calculate the maximum value of each pixel sampling distance set according to the following formula as the maximum distance to obtain the pixel sampling value in the polar view, and arrange it into a two-dimensional acquisition image as the polar view I;