A spatial target recognition method based on capsule networks

Abstract

This invention discloses a spatial target recognition algorithm based on capsule networks, relating to the field of spatial target recognition technology. It proposes a capsule network based on a Dirichlet process hybrid model. This invention improves the capsule network by adding a Dirichlet process hybrid model to the capsule network based on neural networks. This method can realize the recognition of spatial targets, improve the model's generalization ability to target observation perspectives and anti-attack characteristics, and effectively improve the accuracy of spatial target recognition.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a spatial target recognition method based on capsule networks. Background Technology

[0002] With the development of aerospace technology and the exploitation of space resources by various countries, the number of spacecraft in space has increased dramatically. Therefore, observing and identifying non-cooperative targets in space, and even providing on-orbit services such as recovery and debris removal, is of great significance. Currently, space target observation and identification systems mainly consist of ground-based and space-based components. Space-based observation systems are less susceptible to interference than ground-based observation platforms, and optical detection offers advantages such as low operating costs and small size. Therefore, using optoelectronic systems mounted on microsatellites to observe space targets is more advantageous. The visible light images obtained by optical detection contain features such as the color, texture, shape, and spatial relationships of space targets. Utilizing these features to classify and identify space targets can provide support for further space missions.

[0003] To reduce the computational load of feature extraction in spatial target recognition, Cao et al. attempted to integrate manifold correlation theory, compressed sensing theory, wavelet transform, and singular value decomposition into the recognition process; Zhu et al. used binarized images combined with the K-nearest neighbor method and support vector machines to reduce computational load. However, these methods cannot capture deep semantic features of the image, and the recognition effect needs improvement. With the popularization of deep learning, Yang et al. proposed a recognition network that combines spatial target component features. This model fuses the features extracted from the input image with the component features after Mask R-CNN segmentation and inputs them into a ResNet network to classify spatial targets. This method can extract the features of the desired target relatively accurately; however, since the effect of deep learning largely depends on the dataset, the accuracy and generalization of the model are uncertain. In addition, considering that traditional neural networks are vulnerable to attacks in practical applications, even a small modification to the test image according to a specific method can lead to a significant decrease in the model's prediction accuracy. Therefore, this method cannot be used in spatial and other applications with high security requirements. Thus, it is necessary to adopt certain adversarial measures to enhance the reliability of the method. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes a spatial target recognition method based on capsule networks.

[0005] To achieve the above-mentioned technical objectives, the technical solution of the present invention is as follows:

[0006] A spatial target recognition method based on capsule networks includes the following steps:

[0007] S1. Establish a satellite simulation model and acquire image data of space targets;

[0008] S2. Preprocess the image data of the spatial targets obtained in step S1 and create a dataset;

[0009] S3. Construct a capsule network and input the dataset constructed in step S2 into the capsule network to extract target features;

[0010] S4. Establish the initial weight matrix and activation matrix; and perform convolution operation with the features obtained in S3, with the outputs denoted as the pose matrix and activation values, respectively; combine the pose matrix and activation values ​​into a capsule.

[0011] S5. Initialize the new weight matrix by multiplying the attitude matrix obtained in step S4 with the weight matrix to construct the Dirichlet process hybrid model.

[0012] S6. Normalize the activation values ​​in the final capsule layer. The normalized activation values ​​are the probability values ​​corresponding to each category, thus obtaining the spatial target recognition results.

[0013] Preferably, the space target image in step S1 is modeled using Blender software, and similar simulated satellite images are obtained by imitating publicly available satellite models, and an image database is established.

[0014] Preferably, in step S2, the spatial targets in the image need to be selected and marked in advance by the marking software. (1) The image is marked by LabelME marking software, in which the satellite parts in the image need to be selected and classified by polygons. After the marking is completed, each image will generate a corresponding .json file with a label, which records the type and location information of the marking.

[0015] (2) Organize the obtained .json files into the VOC dataset format, which includes dividing these images into training and test sets according to requirements and proportions. Here, the training and test sets are divided in a 9:1 ratio and corresponding .json files are generated to record the labeling information of the training and test sets; thus obtaining the spatial target image set and the VOC format dataset required for training.

[0016] Preferably, in step S3, the processed VOC dataset is input into the neural network, and the target features are extracted in the convolutional layer of the network, which is a normal convolutional structure. The same convolutional kernel is used to filter the entire image, and the target part in the image is initially filtered out to obtain the shallow features of the target, that is, the image features without pose relationship.

[0017] Preferably, the capsule network in step S4 includes convolutional layers, convolutional capsules, digital capsules, a first fully connected layer, a second fully connected layer, and a third fully connected layer, with the capsule layers connected via a routing method; these are considered as two parts: an encoder and a decoder.

[0018] The encoder consists of three parts: convolutional layers, convolutional capsule layers, and digital capsule layers. The convolutional layers detect the basic features of the image. In the capsule network, 256 9x9x1 convolutional kernels are used with a stride of 1. The task of the convolutional capsules is to combine the basic features detected by the convolutional layers. The digital capsules transform the feature matrix obtained above into a multi-dimensional vector.

[0019] Decoder: The decoding part includes a first fully connected layer, a second fully connected layer, and a third fully connected layer; the multi-dimensional vectors obtained from the encoder are input into the decoder, and the fully connected layers decode these multi-dimensional vectors to extract the image.

[0020] The specific routing methods between each layer are as follows:

[0021] The routing method is modeled using a Dirichlet process hybrid model; the Dirichlet process G is truncated and an infinite sequence π is constructed. k The formula is expressed as:

[0022]

[0023] In the formula, v k It is a random variable, and v k ~Beta(1,α) c );v i For the i-th random variable

[0024] The formula for the Dirichlet process G is expressed as:

[0025]

[0026] In the formula, Represents an atomic measure;

[0027] Introduce a random variable Z that follows a multinomial distribution. n Z n From sequence π k The generated component represents the mixed component of the observed data, denoted as Z. n |π~Mult(π); The normal inverse Wissaud distribution in the exponential family of distributions is chosen as the base distribution G0;

[0028] Suppose that (m,S) follows a normal inverse Wieschat distribution, denoted as (m,S)~NIW(μ0,λ,Ψ,v), and its probability density function is expressed as follows:

[0029]

[0030] In the formula, μ0 is a d-dimensional vector over the real number field, θ is a positive real number, Ψ is a positive definite symmetric matrix, and v is a real number greater than d⁻¹. Represent the inverse Wieshard distribution; set an upper limit K for k, and when k exceeds the upper limit, let π k If the value is zero, the mean field assumption is introduced, and the mixture model is solved using an approximate inference method. Let W represent the latent variables in the model, so that the probability density function q(W) approximates the posterior distribution p(W|X) of the model.

[0031] The approximate probability density function q(W) of the latent variable W is decomposed and expressed by the following formula:

[0032]

[0033] In the formula, f n (z n ) represents the category distribution. The parameter is The Beta distribution, h i (m i ,S i |μ i ,λ i ,Ψ i ,v i Initialization parameters (μ0, λ0, Ψ0, v0, φ) n,i Solving the above equation yields the iterative equation system (5) and (6), and equation (5) is used to update the parameter (μ). i ,λ i ,Ψ i ,v i And calculate the parameters of the Beta distribution. Update parameter φ using equation (6) n,i Equations (5) and (6) are expressed as follows:

[0034]

[0035] In Formula 6, D represents the observed data x n The dimension of ψ(·) represents the double gamma function;

[0036] The activation function is defined as follows:

[0037]

[0038] In the formula, parameter β a and parameter β u For hyperparameters;

[0039] The steps of the routing method based on the Dirichlet process hybrid model are as follows:

[0040] (1) Initialize parameter φ n,i 1 / Ω j , where Ω j Let λ0, Ψ0, and v0 represent the number of capsules in layer L+1. Initialize the parameters λ0, Ψ0, and v0 to satisfy the conditions of a normal inverse Wieshard distribution, i.e., λ0 must be a real number greater than zero and Ψ0 must be a symmetric positive definite matrix. Assuming Ψ0 is a d×d matrix, then v0 is greater than d-1.

[0041] Update φ using the activation value of capsule i. n,i As shown in formula (8):

[0042] φ n,i new =φ n,i old a i (8)

[0043] (2) Use equation (5) to update parameter μ i ,λ i ,Ψ i ,v i ,φ n,i ;

[0044] (3) Calculate using Formula 6 and Introducing φ n,i Same-dimensional matrix φ vote As a parameter, φ vote The value φ is generated during the training process. vote Multiplying by the exponent in formula (6), the normalized value of relative capsule j is the updated φ. n,i ;

[0045] (4) Repeat steps (2) and (3) to output the updated μ. i As the pose matrix of the capsule, the activation value a of capsule j is calculated using formula (7). j .

[0046] Preferably, step S5 specifically includes the following steps: after routing to the last layer, for the activation value a j Normalization is performed so that the value is within the interval [0,1], as shown in the following formula:

[0047]

[0048] In the formula, Vj represents the normalized value, and represents the probability of the corresponding class; the cross-entropy function is used as the loss function to train the model, and a j The activation value of capsule j.

[0049] The beneficial effects of adopting the above technical solution are as follows:

[0050] (1) The present invention uses a capsule network based on the Dirichlet process, which has the advantages of small parameters and strong generalization ability to observation perspective compared with the prior art.

[0051] (2) Compared with the prior art, the capsule network model in this invention has strong anti-attack characteristics and good robustness to data images of non-cooperative spatial targets. Attached Figure Description

[0052] Figure 1 It is a simulation space target dataset diagram;

[0053] Figure 2 This is a schematic diagram of a capsule network structure;

[0054] Figure 3 This is a graph showing the changes in the recognition accuracy and loss function of the capsule network under the simulation dataset. Figure 3 (a) in the figure is a graph showing the change in accuracy. Figure 3 (b) in the figure shows the change of the loss function. Detailed Implementation

[0055] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] This invention proposes a spatial target recognition method based on capsule networks, comprising the following steps:

[0057] Step S1: By establishing a satellite simulation model, an image database of space targets is obtained. Based on the different shapes and components of the satellite, 10 different satellite models are established. Each model generates 100 images from different angles, resulting in a total of 1000 sample images.

[0058] Step S2: Preprocess the spatial image data and create a suitable dataset, which includes three parts: a test set, a training set, and a validation set.

[0059] Step S3: Build a suitable capsule network and input the divided dataset into the convolutional layer of the network to extract target features;

[0060] Step S4: Set the initial weight matrix and activation matrix, and perform a convolution operation with the features obtained in S3. The outputs are denoted as the pose matrix and activation values, respectively. Combine the pose matrix and activation values ​​into a capsule.

[0061] Step S5: Initialize the new weight matrix by multiplying the pose matrix obtained in step S4 with the weight matrix to construct the Dirichlet process hybrid model.

[0062] Step S6: Normalize the activation values ​​in the final capsule layer to represent the probability values ​​corresponding to each category and obtain the spatial target recognition results;

[0063] The spatial target images required in step S1 are modeled using Blender software, simulated, and then used to obtain the corresponding images and establish a database.

[0064] Based on the 3D models of some satellites, materials and textures are added to the models, and rendering and imaging are performed to construct... Figure 1 The space target dataset shown has 10 different satellite models, each generating 100 images from different angles, for a total of 1000 sample images.

[0065] In step S2, the images need to be labeled with the target in advance using labeling software. Then, the dataset and training set are divided to create the required VOC dataset. For each type of satellite in the simulation dataset, 50 images are used as the training set, 20 images as the validation set, and 50 images as the test set.

[0066] In step S3, the processed training data needs to be input into the convolutional layer of the neural network to extract target features so that the target can be found through the features later.

[0067] In step S4, the network initialization weight matrix and activation matrix are set. A zero matrix is ​​set as the initialization. The features from step S3 are convolved with the weight matrix and activation matrix respectively, and the outputs are denoted as the pose matrix and activation values ​​respectively. The pose matrix and activation values ​​are combined into a capsule.

[0068] Capsule network structure such as Figure 2 As shown, it generally consists of convolutional layers, basic capsules, convolutional capsules, and digital capsules. The image is input into the convolutional layer to extract features. The basic capsule layer transforms the features into capsules. The weight matrix and activation matrix in the basic capsule layer are initialized. The features are then convolved with the weight matrix and activation matrix respectively, outputting the pose matrix M and activation values ​​a.

[0069] Capsule layers are connected via a routing method. Taking the routing method between the base capsule layer and the convolutional capsule layer as an example, a new weight matrix is ​​initialized. Taking the routing process from base capsule i to the next capsule j as an example, first, the pose matrix and weight matrix W of capsule i are... ij Multiplying them together yields the voting matrix V. ij By V ij The h-th component V ij h The process of getting the h-th component of capsule j is similar to that of V ij h The process of estimating model parameters from observation data.

[0070] In step S5, a new weight matrix is ​​initialized. The pose matrix of the capsule from step S4 is multiplied by the weight matrix to obtain the voting matrix. Prior probabilities are added to the voting matrix, and a Dirichlet process hybrid model is constructed. The iterative process is solved by approximate inference, and the features are routed to subsequent capsule layers.

[0071] The routing method is modeled using a Dirichlet process hybrid model. A truncated stick construction is performed on the Dirichlet process G, transforming the infinite sequence π... k Represented as:

[0072]

[0073] Where v k It is a random variable, and v k ~Beta(1,α) c If ), then the Dirichlet process G can be expressed as:

[0074]

[0075] in Let Z represent an atomic measure. Introduce a random variable Z that follows a multinomial distribution. n Z n From sequence π k The generated component represents the mixed component of the observed data, denoted as Z. n |π~Mult(π). The Normal-Inverse-Wishart Distribution from the exponential family of distributions is chosen as the base distribution G0. Assume (m,S) follows a Normal-Inverse-Wishart Distribution, denoted as (m,S)~NIW(μ0,λ,Ψ,ν), whose probability density function satisfies the equation:

[0076]

[0077] Where μ0 is a d-dimensional vector over the real number field, λ is a positive real number, Ψ is a positive definite symmetric matrix, and v is a real number greater than d⁻¹. This represents the inverse Wieshard distribution. An upper limit K is set for k. When k exceeds the upper limit, let π... k The value is zero. Introducing the mean-field assumption, we solve the mixture model using an approximate inference method. Let W represent the latent variables in the model, and thus the probability density function q(W) approximates the posterior distribution p(W|X) of the model.

[0078] The approximate probability density function q(W) of the latent variable W is decomposed as shown in Equation 4:

[0079]

[0080] Where f n(z n ) represents the category distribution. The parameter is The Beta distribution, h i (m i ,S i |μ i ,λ i ,Ψ i ,v i Initialization parameters (μ0, λ0, Ψ0, v0, φ) n,i Solving these equations yields iterative equation sets 5 and 6:

[0081]

[0082] Update the parameters (μ) using equation system 5. i ,λ i ,Ψ i ,v i And calculate the parameters of the Beta distribution. Update parameter φ using equation system 6 n,i .

[0083] In Formula 6, D represents the observed data x n The dimension of ψ(·) represents the double gamma function.

[0084] For capsule j, the activation function is defined as shown in Equation 7:

[0085]

[0086] Where parameter β a and parameter β u For hyperparameters. The complete steps of the routing method based on the Dirichlet process hybrid model are as follows:

[0087] Initialize parameter φ n,i 1 / Ω j , where Ω j This represents the number of capsules in layer L+1. Initialize parameters λ0, Ψ0, and v0 to satisfy the conditions of a normal inverse Wieshard distribution, i.e., λ0 must be a real number greater than zero, and Ψ0 must be a symmetric positive definite matrix. Assuming Ψ0 is a d×d matrix, then v0 should be greater than d⁻¹.

[0088] Update φ using the activation value of capsule i. n,i As shown in Formula 8:

[0089] φ n,i new =φ n,i old a i (8)

[0090] Use Equation 5 to update the parameter μ. i ,λ i ,Ψ i ,v i ,φ n,i .

[0091] Calculate using formula 6 and Introducing φ n,i Same-dimensional matrix φ vote As a parameter, φ vote The value of φ is generated during the training process. vote Multiplying by the exponent in Equation 4.27, the normalized value of relative capsule j is the updated φ. n,i .

[0092] Repeat steps (2) and (3) to output the updated μ. i The pose matrix of capsule j is used as the basis for calculation. The activation value a of capsule j is then calculated using Equation 7. j .

[0093] After routing to the last layer, the activation values ​​are normalized to represent the corresponding class probabilities. The cross-entropy function is used as the loss function to train the model.

[0094] In step S6, the activation values ​​in the final capsule layer obtained in step S5 are normalized to represent the probability values ​​corresponding to each category.

[0095] Using 50 images of each satellite class in the simulation dataset as the training set, 20 as the validation set, and 50 as the test set, without using convolutional capsule layers, the test accuracy reached 91.41%, the F1 score reached 0.9652, and the model parameters were only 0.5M. During training, the changes in recognition accuracy and loss function are as follows... Figure 3 As shown.

[0096] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0097] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0101] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A spatial target recognition method based on capsule networks, characterized in that, Includes the following steps: S1. Establish a satellite simulation model and acquire image data of space targets; S2. Preprocess the image data of the spatial targets obtained in step S1 and create a dataset; S3. Construct a capsule network and input the dataset constructed in step S2 into the capsule network to extract target features; S4. Establish the initial weight matrix and activation matrix; and perform convolution operation with the features obtained in S3, with the outputs denoted as the pose matrix and activation values, respectively; combine the pose matrix and activation values ​​into a capsule. S5. Initialize the new weight matrix by multiplying the attitude matrix obtained in step S4 with the weight matrix to construct the Dirichlet process hybrid model. S6. Normalize the activation values ​​in the final capsule layer. The normalized activation values ​​are the probability values ​​corresponding to each category, and the spatial target recognition results are obtained. In step S4, the capsule network includes convolutional layers, convolutional capsules, digital capsules, a first fully connected layer, a second fully connected layer, and a third fully connected layer. The capsule layers are connected through a routing method; these can be considered as two parts: an encoder and a decoder. (1) Encoder: The encoder consists of three parts: convolutional layer, convolutional capsule layer and digital capsule layer. The convolutional layer detects the basic features of the image. In the capsule network, 256 9x9x1 convolutional kernels are used and the stride is set to 1. The task of the convolutional capsule is to combine the basic features detected by the convolutional layer. The digital capsule transforms the previously obtained feature matrix into a multidimensional vector; (2) Decoder: The decoding part includes a first fully connected layer, a second fully connected layer and a third fully connected layer; the multi-dimensional vectors obtained from the encoder part are input into the decoder, and the images are deduced from these multi-dimensional vectors through the fully connected layer.

2. The spatial target recognition method based on capsule networks according to claim 1, characterized in that, In step S1, the space target image is modeled using Blender software, and similar simulated satellite images are obtained by imitating publicly available satellite models, and an image database is established.

3. The spatial target recognition method based on capsule networks according to claim 1, characterized in that, In step S2, the spatial targets in the image need to be selected and marked in advance using marking software. (1) Use LabelME tagging software to tag the images, which requires selecting and classifying satellite components in the images using polygons; after tagging, each image will generate a corresponding .json file with tags, which records the types and locations of the tags; (2) Organize the obtained .json files into the VOC dataset format, which includes dividing these images into training and test sets according to requirements and proportions. Here, the training and test sets are divided in a 9:1 ratio and corresponding .json files are generated to record the labeling information of the training and test sets. This allows us to obtain a set of spatial target images, which in turn provides the VOC format dataset required for training.

4. The spatial target recognition method based on capsule networks according to claim 1, characterized in that, In step S3, the processed VOC dataset is input into the neural network. The target features are extracted in the convolutional layer of the network, which is a regular convolutional structure. The same convolutional kernel is used to filter the entire image, and the target part in the image is initially filtered out to obtain the shallow features of the target, which are image features without pose relationships.

5. The spatial target recognition method based on capsule networks according to claim 1, characterized in that, The specific routing methods between each layer are as follows: The routing method is modeled using a Dirichlet process hybrid model; the Dirichlet process G is truncated and an infinite sequence π is constructed. k The formula is expressed as: In the formula, v k It is a random variable, and v k ~Beta(1,α) c );v i For the i-th random variable The formula for the Dirichlet process G is expressed as: In the formula, Represents an atomic measure; Introduce a random variable Z that follows a multinomial distribution. n Z n From sequence π k The generated component represents the mixed component of the observed data, denoted as Z. n |π~Mult(π); The normal inverse Wissaud distribution in the exponential family of distributions is chosen as the base distribution G0; Suppose that (m,S) follows a normal inverse Wieschat distribution, denoted as (m,S)~NIW(μ0,λ,Ψ,v), and its probability density function is expressed as follows: In the formula, μ0 is a d-dimensional vector over the real number field, λ is a positive real number, Ψ is a positive definite symmetric matrix, and v is a real number greater than d⁻¹. Represent the inverse Wieshard distribution; set an upper limit K for k, and when k exceeds the upper limit, let π k If the value is zero, the mean field assumption is introduced, and the mixture model is solved using an approximate inference method. Let W represent the latent variables in the model, so that the probability density function q(W) approximates the posterior distribution p(W|X) of the model. The approximate probability density function q(W) of the latent variable W is decomposed and expressed by the following formula: In the formula, f n (z n ) represents the category distribution. The parameter is The Beta distribution, h i (m i ,S i |μ i ,λ i ,Ψ i ,v i Initialization parameters (μ0, λ0, Ψ0, v0, φ) n,i Solving the above equation yields the iterative equation system (5) and (6), and equation (5) is used to update the parameter (μ). i ,λ i ,Ψ i ,v i And calculate the parameters of the Beta distribution. Update parameter φ using equation (6) n,i Equations (5) and (6) are expressed as follows: In Formula 6, D represents the observed data x n The dimension of ψ(·) represents the double gamma function; The activation function is defined as follows: In the formula, parameter β a and parameter β u For hyperparameters; The steps of the routing method based on the Dirichlet process hybrid model are as follows: (1) Initialize parameter φ n,i 1 / Ω j , where Ω j Let λ0, Ψ0, and ν0 represent the number of capsules in layer L+1. Initialize the parameters λ0, Ψ0, and ν0 to satisfy the conditions of a normal inverse Wieshard distribution, i.e., λ0 must be a real number greater than zero and Ψ0 must be a symmetric positive definite matrix. Assuming Ψ0 is a d×d matrix, then ν0 is greater than d-1. Update φ using the activation value of capsule i. n,i As shown in formula (8): f n,i new =φ n,i old a i (8) (2) Use equation (5) to update parameter μ i ,λ i ,Ψ i ,v i ,φ n,i ; (3) Calculate using Formula 6 and Introducing φ n,i Same-dimensional matrix φ vote As a parameter, φ vote The value φ is generated during the training process. vote Multiplying by the exponent in formula (6), the normalized value of relative capsule j is the updated φ. n,i ; (4) Repeat steps (2) and (3) to output the updated μ. i As the pose matrix of the capsule, the activation value a of capsule j is calculated using formula (7). j .

6. The spatial target recognition method based on capsule networks according to claim 1, characterized in that, Step S5 specifically includes the following steps: After routing to the last layer, for the activation value a j Normalization is performed so that the value is within the interval [0,1], as shown in the following formula: In the formula, V j The normalized value represents the probability of the corresponding class; the cross-entropy function is used as the loss function to train the model, a. j The activation value of capsule j.

Images