Space target ISAR image classification method under small sample conditions based on XGBoost with multi-learner optimization

Abstract

The invention relates to a method for classifying ISAR images of space targets under small sample conditions based on XGBoost optimized by multiple learners, and belongs to the technical field of ISAR image processing. The invention is to solve the problem of low recognition rate of ISAR images of space targets under small sample conditions. The invention targets ISAR spin space target images to be classified, and classifies ISAR images of space targets using an XGBoost model optimized by multiple learners; the XGBoost model is an XGBoost network model formed by integrating a stacked network with an XGBoost algorithm, and the parallel basic learners in the stacked network include a graph convolutional network, a capsule network, and an Alexnet network with an attention mechanism module based on rotation invariance; wherein the capsule network includes a basic feature extraction module, a vector feature extraction module, and a dynamic routing layer, and the capsule network does not contain a pooling layer. It is mainly used for the classification of ISAR images.

Description

Technical Field

[0001] The invention relates to a space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners, and belongs to the technical field of ISAR image processing. Background Art

[0002] Recently, deep learning models have achieved great success in Inverse Synthetic Aperture Radar (ISAR) image recognition. However, due to the contradiction between the deep learning model parameter space and the insufficient labeled samples of ISAR space target images, most deep learning models cannot obtain satisfactory results under small sample conditions.

[0003] For ISAR images of space targets, due to problems such as target inversion and occlusion, the images may be missing or have different forms of expression. In addition, due to the influence of angular glint and coherent speckle noise, the image quality is reduced, which affects ISAR image recognition. For space targets, there are also differences in target structure, which leads to completely different representations of the same type of targets in the image. In addition, the rotation of the target with respect to the radar causes the imaging plane to change, which causes the radar to obtain different sides of the target. These reasons lead to the emergence of small sample problems.

[0004] Existing deep learning networks generally require a large amount of labeled sample data. Even networks that learn from small samples have the problem of low recognition rate, and there is a lack of network models specifically for ISAR image recognition. Summary of the invention

[0005] In view of the problem of low recognition rate of ISAR images of space targets under small sample conditions, the present invention provides a classification method of ISAR images of space targets under small sample conditions based on XGBoost optimized by multiple learners.

[0006] The space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners, for the ISAR spinning space target images to be classified, the space target ISAR image classification is performed using the XGBoost model based on multi-learner optimization;

[0007] The XGBoost model based on multi-learner optimization is an XGBoost network model formed by integrating a stacked network using the XGBoost algorithm;

[0008] The stacked network is a stacked network based on meta-learner integration, and the parallel basic learners in the stacked network include a graph convolutional network, a capsule network and an Alexnet network with an attention mechanism module based on rotation invariance added; the Alexnet network with an attention mechanism module based on rotation invariance added is a network model formed by adding the attention mechanism module based on rotation invariance to the back of the second convolutional layer of the Alexnet network; the capsule network includes a basic feature extraction module, a vector feature extraction module and a dynamic routing layer, and the capsule network does not contain a pooling layer.

[0009] Furthermore, it is characterized in that before using the XGBoost model based on multi-learner optimization to classify the space target ISAR image, the ISAR spin space target image to be classified needs to be preprocessed, and the preprocessing process includes the following steps:

[0010] Firstly, enhance the Lee filter on the ISAR image of the space target;

[0011] Then contrast enhancement and energy normalization are performed.

[0012] Furthermore, the specific structure of the capsule network is as follows:

[0013] (A) Basic feature extraction module:

[0014] The basic feature extraction module consists of three convolution units, each of which consists of a convolution layer, a batch normalization layer, and a ReLU activation layer;

[0015] (B) Vector feature extraction module:

[0016] The vector feature extraction module contains 8 convolutional layers, which are in parallel relationship and do not use activation functions. Each convolutional layer has the same size and performs convolution on the input feature map respectively.

[0017] Then the obtained results are organized into tensor form;

[0018] (C) Dynamic Routing Layer

[0019] The dynamic routing layer is implemented using a dynamic routing method.

[0020] Furthermore, the Alexnet network with the attention mechanism module based on rotation invariance is as follows:

[0021] The attention mechanism module based on rotation invariance includes a rotation processing module and a spatial attention mechanism module;

[0022] Rotation processing module: Based on the Alexnet network, the output feature map of the ReLU activation function of the second convolutional layer is rotated to obtain the rotated feature map;

[0023] The spatial attention mechanism module consists of three convolutional units, namely Conv1, Conv2 and Conv3;

[0024] Conv1: The rotated feature map first passes through the convolution layer and then through the ReLU activation;

[0025] Conv2 includes Conv2-1 and Conv2-2 subunits. The feature maps processed by Conv1 are sent to Conv2-1 and Conv2-2 subunits respectively. In Conv2-1 and Conv2-2 subunits, they are first passed through the convolution layer and then deactivated using Softmax to obtain the output feature map F 2-1 and F 2-2 ;

[0026] The Conv3 unit includes Conv3-1 and Conv3-2 subunits. The rotation feature map obtained by the rotation processing module is directly sent to the Conv3-1 and Conv3-2 subunits respectively. It first passes through the convolutional layers in the Conv3-1 and Conv3-2 subunits, and then undergoes batch normalization to obtain G 3-j ,j=1,2;

[0027] For the feature map F 2-j Transpose to get F′ 2-j , put F′ 2-j And feature map G 3-j With randomly initialized coefficients α j Multiply and add to get the output feature map

[0028] Finally, the output feature maps obtained by rotating at each angle are averaged to obtain the output feature map of the RIAM module. K represents the number of rotation operations performed by the rotation processing module.

[0029] Furthermore, the graph convolutional network is as follows:

[0030] The graph convolutional network includes a feature extraction module and a graph neural network processing module;

[0031] (a) Feature extraction module:

[0032] The feature extraction module contains three convolution units, each of which contains: convolution layer, batch normalization layer, maximum pooling layer and ReLU activation function layer;

[0033] The last convolution unit outputs the feature expression of the image in vector form and combines it with the one-hot encoding formed by the image category as the input of the graph neural network module;

[0034] (b) Graph neural network module:

[0035] The graph neural network module includes two neighbor matrix calculation units and two graph convolution units; the first neighbor matrix calculation unit is connected to the first graph convolution unit, and then connected to the second neighbor matrix calculation unit and the second graph convolution unit;

[0036] The input of the first calculation neighbor matrix unit is the two-dimensional feature map obtained by subtracting the node representation vector of the feature extraction module; the input of the second calculation neighbor matrix unit is the feature map obtained by subtracting the node representation vector output by the first graph convolution unit; both calculation neighbor matrix units contain three convolution modules, the first two convolution modules contain convolution layers, batch normalization layers, and ReLU activation layers, and the last convolution module has only one convolution layer;

[0037] The first graph convolution unit contains a fully connected layer and a batch normalization layer; the second graph convolution unit contains two fully connected layers and a batch normalization layer, and the output of the second graph convolution unit is equal to the classification probability of the five categories of targets.

[0038] Furthermore, the process of integrating the stacked networks using the XGBoost algorithm to form an XGBoost network model includes the following steps:

[0039] 2.4.1 Calculating the weights of the base learner

[0040] The input of the XGBoost network is D{x i,α ,l i ,e i,α}, where i = 1, 2, ..., n represents the number of meta-features, x i,α is a meta-feature, l i It is a label, i,α is the weight of the meta-feature, α=1,2,3 is the serial number of the basic learner, and the weight q corresponding to the learner is introduced α , the weight expression of the meta-feature is as follows:

[0041]

[0042] Among them, x i,α,1 ,……,x i,α,N They are meta-features x i,α Elements in is the mean of the meta-feature; n is the total number of meta-features, and N is the number of elements in the meta-feature;

[0043] 2.4.2 Calculation of the partial derivative of the weight to the loss function

[0044] First define the loss function L, including the predicted value and label l i The cross entropy loss l, and the tree structure f k The corresponding regularization loss Ω();

[0045] The loss function L is defined as:

[0046]

[0047] l is defined as follows:

[0048]

[0049] Among them, γ is the regularization parameter of the number of leaves, T is the number of leaf nodes, λ is the regularization parameter of the leaf weight, and w represents the weight of the leaf;

[0050] About q α Partial derivative of:

[0051]

[0052]

[0053] 2.4.3 Adaptive weight update

[0054] Use the partial derivative and update rate η calculated above to update the weights of the learner:

[0055]

[0056] The weights corresponding to the meta-features are updated by the changes in the learner weights to achieve adaptive weight updates;

[0057] 2.4.4 Calculation of network gain

[0058] The gain of the network is calculated by the first and second derivatives of the predicted value of l, g i,α and h i,α is the first-order derivative and second-order derivative of the meta-feature, L and R are the left and right nodes of the decision tree. The gain is defined as:

[0059]

[0060]

[0061]

[0062] The tree is split according to the maximum gain to obtain the final classification result.

[0063] Furthermore, the training process of the XGBoost network model includes the following steps:

[0064] Meta-feature output of graph convolutional network: the classification probability of the data in the validation set is used as the training meta-feature;

[0065] The meta-feature output of the Alexnet network with the rotation-invariant attention mechanism module: The classification confidence of the validation sample set is combined to become the training meta-feature;

[0066] Meta-feature output of capsule network: Take the L2 norm of the output vector to get the classification probability, and use the classification probability of the data in the validation sample set as the training meta-feature;

[0067] The multi-learner optimized XGBoost network is trained using the training meta-features.

[0068] Furthermore, the training process of graph convolutional network, capsule network and Alexnet network is as follows:

[0069] The Alexnet network model training process includes the following steps:

[0070] During the training process, each category of the training sample library is divided into K equal parts, of which K-1 parts are training sample sets and the remaining part is the validation sample set. K-fold cross-validation is performed to ensure that all K parts of data have been used as validation samples.

[0071] The Alexnet network model is trained using the training sample set and the validation sample set, and the classification confidence of the validation sample set is combined to become a training meta-feature;

[0072] The graph convolutional network model training process includes the following steps:

[0073] The training sample library is divided into a training set and a validation set. The training set has qshot images for each category, and the validation set has K-qshot images for each category. During the training process, 5×qshot+1 images are input into the network for each iteration, where 1 means taking one image from the validation set as validation in turn, and its label is known, which is used for supervised learning.

[0074] The graph convolutional network model is trained using the training set and validation set, and the classification probability of the data in the validation set is used as the training meta-feature;

[0075] The capsule network model training process includes the following steps:

[0076] The training sample library is divided into a training set and a validation set. The capsule network model is trained using the training set and the validation set. The L2 norm of the output vector is taken to obtain the classification probability. The classification probability of the data in the validation sample set is used as the training meta-feature.

[0077] The training sample library stores ISAR spin space target samples.

[0078] Furthermore, the ISAR spinning space target samples are obtained based on the three-dimensional space target model and the range Doppler algorithm simulation.

[0079] Furthermore, the ISAR spin space target samples stored in the training sample library are preprocessed samples, and the preprocessing process includes the following steps:

[0080] Firstly, enhance the Lee filter on the ISAR image of the space target;

[0081] Then contrast enhancement, energy normalization and data expansion are performed, where data expansion includes left-right flipping and up-down flipping.

[0082] Beneficial effects of the present invention:

[0083] The method of the present invention draws on the idea of ​​integrated networks and uses three complementary small sample learning networks, namely, a network of the Alexnet+RIAM module that focuses on the target position and semantic information of the ISAR image, a graph convolutional network that focuses on the deep feature relationship of the image, and a capsule network that focuses on the relative position relationship of key components in the space target. Finally, the network is integrated using an innovative multi-learner optimized XGBoost decision tree. Since the method adds an attention mechanism module based on rotation invariance, the influence of the rotation of the target on the ISAR image on the recognition is reduced, and the network focuses on different information on the ISAR image, such as semantic information, deep features, etc., so that the final recognition rate is improved.

[0084] The present invention generally includes three parts, ISAR spin space target samples, stacked network architecture and target category estimation; the ISAR spin space target sample part is responsible for simulating and generating a large number of ISAR images of space targets in motion state, forming a training sample library and a test sample library, the stacked network architecture gives the design structure of the network, and the target category estimation part is based on the stacked network, and trains and tests the ISAR spin space target sample library to realize the target category estimation function.

[0085] The stacked network model trained with the training sample library of the present invention can make target category estimation for ISAR space target images under the conditions of a small number of labeled samples in ISAR observation scenarios, thereby improving the correct classification probability of targets by basic learners such as graph convolutional networks, meeting actual needs and being easy to implement.

[0086] The present invention obtains the scattering point information required for ISAR imaging through the three-dimensional model of the space target and electromagnetic parameter simulation. The ISAR imaging results are almost consistent with the measured ISAR images. A small number of ISAR simulation images of five types of targets are obtained, which ensures the sample requirements of the stacked network method in this paper and the smooth application of the algorithm.

[0087] The present invention innovatively proposes a stacked network to estimate the ISAR spin space target type. The input of the stacked network is an ISAR space target image with a small number of labeled samples, and the output is the category to which the target belongs. The introduction of the attention mechanism module based on rotation invariance and the proposed XGBoost method for multi-learner optimization provide guarantees for improving the recognition rate, and the method has certain superiority.

[0088] The present invention utilizes ISAR imaging technology and deep learning technology to better meet the requirements of complex ISAR target category estimation in different imaging scenarios. The stacked network maintains a certain degree of stability to the perspective change, affine transformation and noise of the input image. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is an overall flow chart of the space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners according to the present invention;

[0090] Figure 2 It is a schematic diagram of the RIAM module design;

[0091] Figure 3 This is the architecture diagram of the graph convolutional network in the base learner;

[0092] Figure 4 This is the architecture diagram of the capsule network in the basic learner;

[0093] Figure 5 It is a flowchart for obtaining the training sample library and the test sample library. DETAILED DESCRIPTION

[0094] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0095] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0096] Specific implementation method 1. Combination Figures 1 to 5 To explain this embodiment,

[0097] The method for classifying space target ISAR images under small sample conditions based on XGBoost optimized by multiple learners described in this embodiment includes the following steps:

[0098] S1. Based on the three-dimensional space target model and the range Doppler algorithm simulation, an ISAR spin space target sample sequence is obtained, and the ISAR spin space target samples are grouped into a training sample library and a test sample library;

[0099] After obtaining the ISAR spinning space target sample sequence, the image data needs to be preprocessed: first, the space target ISAR image is enhanced by Lee filtering to remove the coherent speckle noise; then contrast enhancement, energy normalization and data expansion are performed, where data expansion includes left-right flipping and up-down flipping;

[0100] In the process of grouping ISAR spin space target samples into training sample library and test sample library, a K / N value needs to be set. This value represents the ratio of the number of images in the training sample library to the number of images in the test sample library. When K / N=1, the number of images in the training sample library is equal to the number of images in the test sample library.

[0101] S2. Construct a stacked network based on meta-learner integration. The basic learners of the stacked network include the Alexnet network with an attention mechanism module based on rotation invariance, a graph convolutional network, and a capsule network. The basic learners of the stacked network are learners that work in parallel.

[0102] The final meta-learner uses the multi-learner optimized XGBoost algorithm to integrate the above three basic learners;

[0103] The Alexnet network with the attention mechanism module based on rotation invariance is added: the attention mechanism module based on rotation invariance is added after the second convolution layer of the Alexnet network. Due to the imaging mechanism of ISAR space target images, the same target is often displayed at different angles in the image, so the rotation invariance of feature extraction is enhanced to the network to enhance the recognition effect. The construction process of the Alexnet network with the attention mechanism module based on rotation invariance includes the following steps:

[0104] 2.1 Construct an Alexnet network model with an attention mechanism module based on rotation invariance:

[0105] 2.1.1 During the training process, each category of the training sample library is divided into K equal parts, of which K-1 parts are training sample sets and the remaining part is the validation sample set. K-fold cross-validation is performed to ensure that all K parts of data have been used as validation samples;

[0106] 2.1.2 Attention mechanism module based on rotation invariance RIAM includes a rotation processing module and a spatial attention mechanism module;

[0107] Rotation processing module: Based on the Alexnet network, a series of rotation operations are performed on the output feature map of the ReLU activation function of the second convolutional layer to convert it into a rotated feature map; is a series of rotation operations, using The function transforms the input feature map at an angle of λ i ∈λ,i=1,....,K rotation transformation, get the rotated feature map x, x∈R {C×W×H} , where R represents a real number set, W×H is the size of the feature map, and C is the number of feature channels, which is consistent with the number of input feature map channels;

[0108] The spatial attention mechanism module consists of three convolutional units, namely Conv1, Conv2 and Conv3;

[0109] First, set the convolution kernel size and number of channels of Conv1. The rotation feature map first passes through the convolution layer and then passes through the ReLU deactivation, as shown in the following formula:

[0110] F(x) = ReLU(Conv1(R λ (x))) (1)

[0111] Conv2 includes Conv2-1 and Conv2-2 subunits. The feature maps processed by Conv1 are sent to Conv2-1 and Conv2-2 subunits respectively. For Conv2-1 and Conv2-2 subunits, the input feature map is F(x). The convolution kernel size and the number of channels are set to be consistent with the number of input feature channels. In Conv2-1 and Conv2-2 subunits, the convolution layer is first used and then deactivated using Softmax to obtain the output feature map F 2-1 and F 2-2 , which means the following:

[0112] F 2-j =Softmax(Conv 2-j (F(x))),j=1,2 (2)

[0113] The Conv3 unit includes Conv3-1 and Conv3-2 subunits. The rotated feature map obtained by the rotation processing module is directly sent to the Conv3-1 and Conv3-2 subunits respectively. It first passes through the convolution layer in the Conv3-1 and Conv3-2 subunits, uses hollow convolution, the convolution kernel size is 3×3, the expansion step is set to 2, and the padding is set to keep the output feature map size W×H (so that the RIAM module can be added without changing the Alexnet structure); then batch normalization is performed, as shown below:

[0114] G 3-j =BN(Conv 3-j (R λ (x))),j=1,2 (3)

[0115] For the feature map F 2-j Transpose to get F′ 2-j , put F′ 2-j And feature map G 3-j With randomly initialized coefficients α j After multiplication and addition, the output feature map is obtained, which is expressed as:

[0116]

[0117] Finally, the output feature maps obtained by rotating at each angle are averaged to obtain the output feature map of the RIAM module, which is expressed as:

[0118]

[0119] After putting the built attention mechanism module based on rotation invariance into the second convolutional layer of Alexnet, because the features in the deep layer do not require high rotation invariance, and the features extracted by the first convolutional layer are too sensitive to rotation, putting it after the first layer leads to poor final classification results;

[0120] 2.1.3 Combine the classification confidence of the validation sample set to form training meta-features; input the test sample library as a test set into the constructed Alexnet network for training to obtain test meta-features.

[0121] 2.2 Constructing graph convolutional network:

[0122] 2.2.1 Data processing:

[0123] The training sample library is further divided into a training set and a validation set. The training set has qshot images for each category, and the validation set has K-qshot images for each category. During the training process, 5×qshot+1 images are input into the network for each iteration, where 1 means taking an image from the validation set as validation in turn, and its label is known, which is used for supervised learning.

[0124] 2.2.2 The graph convolutional network includes a feature extraction module and a graph neural network processing module;

[0125] Overall, the graph convolution network has a total of 9 convolutional layers, including 3 feature extraction modules and 3 computational neighbor matrix units in the graph neural network module; the convolution kernel size in the first convolution layer is 9×9 pixels, the step size is 2 pixels, and there are 32 convolution kernels in total; the convolution kernel size in the second convolution layer is 5×5 pixels, the step size is 1 pixel, and there are 64 convolution kernels in total; the convolution kernel size in the third convolution layer is 3×3 pixels, the step size is 1 pixel, and there are 128 convolution kernels in total; the convolution kernel size in the fourth convolution layer is 3×3 pixels, the step size is 1 pixel, and there are 64 convolution kernels in total; the convolution kernel size in the fifth convolution layer is 3 ×3 pixels, the stride is 1 pixel, and there are 32 convolution kernels in total; the convolution kernel size in the sixth convolution layer is 3×3 pixels, the stride is 1 pixel, and there are 1 convolution kernel in total; the convolution kernel size in the seventh convolution layer is 3×3 pixels, the stride is 1 pixel, and there are 64 convolution kernels in total; the convolution kernel size in the eighth convolution layer is 3×3 pixels, the stride is 1 pixel, and there are 32 convolution kernels in total; the convolution kernel size in the ninth convolution layer is 3×3 pixels, the stride is 1 pixel, and there is 1 convolution kernel in total; there is no activation function only after the last convolution layer for calculating the neighboring matrix units, and the activation function of each of the remaining convolution layers is the ReLU activation function.

[0126] There are three pooling layers in the graph convolutional network, using maximum pooling with a size of 2×2 pixels and a step size of 2 pixels.

[0127] (a) Feature extraction module:

[0128] The feature extraction module contains three convolution units, each of which contains: convolution layer, batch normalization layer, maximum pooling layer, and ReLU activation function layer;

[0129] The input of the convolution layer in the first convolution unit is the training set image and the validation set image, and the input of the next two convolution units is the feature map output by the previous convolution unit;

[0130] Use batch normalization to process the weight coefficients;

[0131] Use the maximum pooling layer to perform pooling operations;

[0132] Use the ReLU function to activate the weight coefficient;

[0133] The last convolution unit outputs the feature expression of the image in the form of a vector, and combines it with the one-hot encoding formed by the image category and records it as the node representation vector of the feature extraction module, which serves as the input of the graph neural network module.

[0134] (b) Graph neural network module:

[0135] The graph neural network module includes two neighbor matrix calculation units and two graph convolution units; the first neighbor matrix calculation unit is connected to the first graph convolution unit, and then connected to the second neighbor matrix calculation unit and the second graph convolution unit;

[0136] The input of the first calculation neighbor matrix unit is the two-dimensional feature map obtained by subtracting the node representation vector of the feature extraction module (the initial input is a series of images, and the node representation vectors are obtained after the previous processing of each image, and the two-dimensional feature map is obtained by subtracting the node representation vectors); the input of the second calculation neighbor matrix unit is the feature map obtained by subtracting the node representation vector output by the first graph convolution unit (similar to the processing of subtracting the node representation vector of the first calculation neighbor matrix unit); both calculation neighbor matrix units contain three convolution modules, the first two convolution modules contain convolution layers, batch normalization layers and ReLU activation layers, and the last convolution module has only one convolution layer;

[0137] The first graph convolution unit contains a fully connected layer and a batch normalization layer; the second graph convolution unit contains two fully connected layers and a batch normalization layer, and the output of the second graph convolution unit is equal to the classification probability of the five categories of targets.

[0138] 2.2.4 Output of meta-features:

[0139] The classification probability of the data in the validation set is used as the training meta-feature, and the classification probability of the test sample library as the test set data is used as the test meta-feature.

[0140] 2.3 Constructing capsule network:

[0141] 2.3.1 Data processing:

[0142] The training sample library is further divided into a training set and a validation set.

[0143] 2.3.2 The capsule network consists of a basic feature extraction module, a vector feature extraction module and a dynamic routing layer;

[0144] Overall, the capsule network has 11 convolutional layers, including 3 basic feature extraction modules and 8 vector feature extraction modules; the convolution kernel size in the first convolutional layer is 9×9 pixels, the step size is 2 pixels, and there are 64 convolution kernels in total; the convolution kernel size in the second convolutional layer is 9×9 pixels, the step size is 2 pixels, and there are 128 convolution kernels in total; the convolution kernel size in the third convolutional layer is 7×7 pixels, the step size is 1 pixel, and there are 128 convolution kernels in total; the convolution kernel size from the fourth to the eleventh convolutional layer is 9×9 pixels, the step size is 2 pixels, and there are 32 layers. These 8 convolutional layers are in parallel and form vector features in the new dimension; the activation functions of the first three convolutional layers are all ReLU activation functions, and the remaining convolutional layers do not use activation functions. The capsule network does not contain a pooling layer.

[0145] (A) Basic feature extraction module:

[0146] The purpose of the basic feature extraction module is to transform the image into a feature map through a convolution layer. The basic feature extraction module includes three convolution units, each of which includes a convolution layer, a batch normalization layer, and a ReLU activation layer.

[0147] The input of the convolutional layer is the training set images and the test set images;

[0148] Use batch normalization to process the weight coefficients;

[0149] Use the ReLU function to activate the weight coefficient;

[0150] Does not contain pooling layers;

[0151] (B) Vector feature extraction module:

[0152] The vector feature extraction module contains 8 convolutional layers, each of which has the same size. The input feature maps are convolved separately, and the results are in the form of tensors.

[0153] The input of the convolutional layer is the training set images and the test set images;

[0154] Set the convolution kernel size, number of channels and stride of the convolution layer respectively;

[0155] The output results are combined in the new dimension and the feature vector is output;

[0156] The length of the feature vector changes after it passes through the posture transformation matrix;

[0157] (C) Dynamic Routing Layer

[0158] Similar to the fully connected layer in the convolutional neural network (dynamic routing and fully connected layer work differently but have the same function), this part is implemented using dynamic routing in the capsule network.

[0159] The number of output vectors is consistent with the number of types of spatial targets;

[0160] 2.3.3 Output of Meta-Features

[0161] The L2 norm of the output vector is taken to obtain the classification probability. The classification probability of the data in the validation sample set is used as the training meta-feature, and the classification probability of the data in the test sample set is used as the test meta-feature.

[0162] 2.4 Establishing an XGBoost network optimized by multiple learners:

[0163] The XGBoost network optimized by multiple learners includes calculating the weights of the base learners, calculating the partial derivative of the weights with respect to the loss function, adaptively updating the weights, and calculating the network gain.

[0164] 2.4.1 Calculating the weights of the base learner

[0165] The input of the network is D{x i,α ,l i ,e i,α}, where i = 1, 2, ..., n represents the number of training meta-features, x i,α is the training meta-feature, l i It is a label, i,α is the weight of the meta-feature, α=1,2,3 is the serial number of the basic learner, and the weight q corresponding to the learner is introduced α , the weight expression of the meta-feature is as follows:

[0166]

[0167] Among them, x i,α,1 ,……,x i,α,N They are meta-features x i,α Elements in is the mean of the meta-feature; n is the total number of meta-features, and N is the number of elements in the meta-feature;

[0168] 2.4.2 Calculation of the partial derivative of the weight to the loss function

[0169] First define the loss function L, including the predicted value and label l i The cross entropy loss l, and the tree structure f k The corresponding regularization loss Ω();

[0170] The loss function L is defined as:

[0171]

[0172] l is defined as follows:

[0173]

[0174] Among them, γ is the regularization parameter of the number of leaves, T is the number of leaf nodes, λ is the regularization parameter of the leaf weight, and w represents the weight of the leaf;

[0175] About q α Partial derivative of:

[0176]

[0177]

[0178] 2.4.3 Adaptive weight update

[0179] Use the partial derivative and update rate η calculated above to update the weights of the learner:

[0180]

[0181] The weights corresponding to the meta-features are updated by the changes in the learner weights to achieve adaptive weight updates;

[0182] 2.4.4 Calculation of network gain

[0183] The gain of the network is calculated by the first and second derivatives of the predicted value of l, g i,α and h i,α is the first-order derivative and second-order derivative of the meta-feature, L and R are the left and right nodes of the decision tree. The gain is defined as:

[0184]

[0185]

[0186]

[0187] The tree is split according to the maximum gain to obtain the final classification result.

[0188] S3, using the training sample library and the test sample library to train the basic learner respectively, to obtain the training meta-feature and the test meta-feature;

[0189] The multi-learner optimized XGBoost network is trained using the training meta-features to obtain the XGBoost network model, and the test meta-features are tested to obtain a result with a higher recognition rate.

[0190] The training samples and the test samples are both pre-processed ISAR images;

[0191] During the training process, the pixel size of the ISAR sample input to the Alexne network of the stacked network is 224×224, the pixel size of the ISAR sample input to the graph convolution network is 32×32, and the pixel size of the ISAR sample input to the capsule network is 128×128, including a total of 5 types of space targets.

[0192] For graph convolutional networks, the input image size is 32×32 pixels;

[0193] After the feature extraction layer processes the ISAR image, it is passed to the first convolutional layer;

[0194] The first convolutional layer processes the input image and outputs 32 feature maps of 32×32 pixels;

[0195] After batch normalization and ReLU processing, the first pooling layer processes the feature map output by the first convolutional layer and outputs a three-dimensional feature matrix of 16×16×32 pixels;

[0196] The second convolutional layer processes the three-dimensional feature matrix output by the first pooling layer and outputs 64 feature maps of 16×16 pixels.

[0197] After batch normalization and ReLU processing, the second pooling layer processes the feature map output by the second convolutional layer and outputs a three-dimensional feature matrix of 8×8×64 pixels;

[0198] The third convolutional layer processes the three-dimensional feature matrix output by the second pooling layer and outputs 128 feature maps of 8×8 pixels;

[0199] After batch normalization and ReLU processing, the third pooling layer processes the feature map output by the third convolutional layer and outputs a three-dimensional feature matrix of 4×4×128 pixels;

[0200] The feature map is converted into a vector and added with one-hot encoding as the vertex feature representation, with a size of 21×1×128. The vertex representation is subtracted to obtain the neighboring feature matrix, with a size of 21×21×128 pixels;

[0201] The fourth convolutional layer processes the neighboring feature matrix and outputs 64 feature maps of 21×21 pixels;

[0202] After batch normalization and ReLU processing, the fifth convolutional layer processes the feature map output by the fourth convolutional layer and outputs 32 feature maps of 21×21 pixels;

[0203] After batch normalization and ReLU processing, the sixth convolutional layer processes the feature map output by the fifth convolutional layer and outputs a feature map of 21×21 pixels;

[0204] The first fully connected layer contains 16 neurons, which are fully connected to the one-dimensional vector mapped by the three-dimensional feature matrix output by the sixth convolutional layer;

[0205] After being processed by the fully connected layer and the batch normalization layer, a vertex representation vector of size 16×1 is output; the vertex representations are subtracted to obtain a neighboring feature matrix of size 16×16×1 pixels;

[0206] The seventh convolutional layer processes the neighboring feature matrix and outputs 64 feature maps of 16×16 pixels;

[0207] After batch normalization and ReLU processing, the eighth convolutional layer processes the feature map output by the seventh convolutional layer and outputs 32 feature maps of 16×16 pixels;

[0208] After batch normalization and ReLU processing, the ninth convolutional layer processes the feature map output by the eighth convolutional layer and outputs a feature map of 16×16 pixels;

[0209] The second fully connected layer contains 16 neurons, which are fully connected to the one-dimensional vector mapped by the three-dimensional feature matrix output by the fourth pooling layer;

[0210] The second fully connected layer contains 5 neurons, which are fully connected to the 16 neurons in the first fully connected layer;

[0211] For capsule networks, the input image size is 128×128 pixels;

[0212] The first convolutional layer processes the input image and outputs 64 feature maps of 60×60 pixels;

[0213] After batch normalization and ReLU processing;

[0214] The second convolutional layer processes the three-dimensional feature matrix and outputs 128 feature maps of 26×26 pixels;

[0215] After batch normalization and ReLU processing;

[0216] The third convolutional layer processes the three-dimensional feature matrix and outputs 128 feature maps of 20×20 pixels;

[0217] After batch normalization and ReLU processing;

[0218] The fourth to eleventh convolutional layers process the three-dimensional feature matrix, and the obtained feature maps form feature vectors in the new dimension with a size of 6×6×8×32;

[0219] After being processed by 6×6×32 transformation matrices of size 8×16, 1152 feature vectors of size 1×16 are obtained;

[0220] Using the dynamic routing algorithm, we obtain five 1×16 feature vectors;

[0221] Calculate the L2 norm of the feature vector to get the classification probability.

[0222] During training, the number of training samples per batch in the Alexnet network is 32, and a total of 2800 batches are trained. The learning rate is 0.0001 before the fully connected layer and 0.001 for the fully connected layer, and the optimization method is SGD; the number of training samples per batch in GCN is 16, and a total of 6000 batches are trained, with a learning rate of 0.01; the number of training samples per batch in CapsNet is 64, and a total of 100 batches are trained, with a learning rate of 0.001; the tree depth in XGBoost optimized by multiple learners is 20, the number of trees is set to 2000, and the number of categories is 5.

[0223] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with a single embodiment may be used in other described embodiments.

Claims

1. A space target ISAR image classification method based on XGBoost with multi-learner optimization under small sample conditions, characterized in that: For the ISAR spinning space target images to be classified, the XGBoost model based on multi-learner optimization is used to classify the space target ISAR images. The XGBoost model based on multi-learner optimization is an XGBoost network model formed by integrating a stacked network using the XGBoost algorithm; The stacked network is a stacked network based on meta-learner integration, and the parallel basic learners in the stacked network include a graph convolutional network, a capsule network, and an Alexnet network with an attention mechanism module based on rotation invariance added; the Alexnet network with an attention mechanism module based on rotation invariance added is a network model formed by adding the attention mechanism module based on rotation invariance to the back of the second convolutional layer of the Alexnet network; the capsule network includes a basic feature extraction module, a vector feature extraction module, and a dynamic routing layer, and the capsule network does not contain a pooling layer; The process of using the XGBoost algorithm to integrate the XGBoost network model formed by the stacked network The following steps are involved: 2.4.1 Calculate the weight of the base learner: The input of the XGBoost network is D{x i,α ,l i ,e i,α }, where i = 1, 2, ..., n represents the number of meta-features, x i,α is a meta-feature, l i It is a label, i,α is the weight of the meta-feature, α=1,2,3 is the serial number of the basic learner, and the weight q corresponding to the learner is introduced α , the weight expression of the meta-feature is as follows: Among them, x i,α,1 ,……,x i,α,N They are meta-features x i,α Elements in is the mean of the meta-feature; n is the total number of meta-features, and N is the number of elements in the meta-feature; 2.4.2 Calculation of the partial derivative of the weight to the loss function: First define the loss function L, including the predicted value and label l i The cross entropy loss l, and the tree structure f k The corresponding regularization loss Ω(); The loss function L is defined as: l is defined as follows: Among them, γ is the regularization parameter of the number of leaves, T is the number of leaf nodes, λ is the regularization parameter of the leaf weight, and w represents the weight of the leaf; About q α Partial derivative of: 2.4.3 Adaptive update weights: Use the partial derivative and update rate η calculated above to update the weights of the learner: The weights corresponding to the meta-features are updated by the changes in the learner weights to achieve adaptive weight updates; 2.4.4 Calculation of network gain: The gain of the network is calculated by the first and second derivatives of the predicted value of l, g i,α and h i,α is the first-order derivative and second-order derivative of the meta-feature, L and R are the left and right nodes separated by the decision tree, and the gain is defined as: The tree is split according to the maximum gain to obtain the final classification result.

2. According to claim 1, the space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners is characterized in that: Before using the XGBoost model based on multi-learner optimization to classify space target ISAR images, the ISAR spin space target images to be classified need to be preprocessed. The preprocessing process includes the following steps: Firstly, enhance the Lee filter on the ISAR image of the space target; Then contrast enhancement and energy normalization are performed.

3. The space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners according to claim 1 or 2, characterized in that: The specific structure of the capsule network is as follows: (A) Basic feature extraction module: The basic feature extraction module consists of three convolution units, each of which consists of a convolution layer, a batch normalization layer, and a ReLU activation layer; (B) Vector feature extraction module: The vector feature extraction module contains 8 convolutional layers, which are in parallel relationship and do not use activation functions. Each convolutional layer has the same size and performs convolution on the input feature map respectively. Then the obtained results are organized into tensor form; (C) Dynamic Routing Layer The dynamic routing layer is implemented using a dynamic routing method.

4. According to claim 3, the space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners is characterized in that: The Alexnet network with the attention mechanism module based on rotation invariance is as follows: The attention mechanism module based on rotation invariance includes a rotation processing module and a spatial attention mechanism module; Rotation processing module: Based on the Alexnet network, the output feature map of the ReLU activation function of the second convolutional layer is rotated to obtain the rotated feature map; The spatial attention mechanism module consists of three convolutional units, namely Conv1, Conv2 and Conv3; Conv1: The rotated feature map first passes through the convolution layer and then through the ReLU activation; Conv2 includes Conv2-1 and Conv2-2 subunits. The feature maps processed by Conv1 are sent to Conv2-1 and Conv2-2 subunits respectively. In Conv2-1 and Conv2-2 subunits, they are first passed through the convolution layer and then deactivated using Softmax to obtain the output feature map F 2-1 and F 2-2 ; The Conv3 unit includes Conv3-1 and Conv3-2 subunits. The rotation feature map obtained by the rotation processing module is directly sent to the Conv3-1 and Conv3-2 subunits respectively. It first passes through the convolutional layers in the Conv3-1 and Conv3-2 subunits, and then undergoes batch normalization to obtain G 3-j ,j=1,2; For the feature map F 2-j Transpose to get F2′ -j , put F2′ -j And feature map G 3-j With randomly initialized coefficients α j Multiply and add to get the output feature map Finally, the output feature maps obtained by rotating at each angle are averaged to obtain the output feature map of the RIAM module. K represents the number of rotation operations performed by the rotation processing module.

5. According to claim 4, the space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners is characterized in that: The graph convolutional network is as follows: The graph convolutional network includes a feature extraction module and a graph neural network processing module; (a) Feature extraction module: The feature extraction module contains three convolution units, each of which contains: convolution layer, batch normalization layer, maximum pooling layer and ReLU activation function layer; The last convolution unit outputs the feature expression of the image in vector form and combines it with the one-hot encoding formed by the image category as the input of the graph neural network module; (b) Graph neural network module: The graph neural network module includes two neighbor matrix calculation units and two graph convolution units; the first neighbor matrix calculation unit is connected to the first graph convolution unit, and then connected to the second neighbor matrix calculation unit and the second graph convolution unit; The input of the first calculation neighbor matrix unit is the two-dimensional feature map obtained by subtracting the node representation vector of the feature extraction module; the input of the second calculation neighbor matrix unit is the feature map obtained by subtracting the node representation vector output by the first graph convolution unit; both calculation neighbor matrix units contain three convolution modules, the first two convolution modules contain convolution layers, batch normalization layers, and ReLU activation layers, and the last convolution module has only one convolution layer; The first graph convolution unit contains a fully connected layer and a batch normalization layer; the second graph convolution unit contains two fully connected layers and a batch normalization layer, and the output of the second graph convolution unit is equal to the classification probability of the five categories of targets.

6. The space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners according to claim 1 is characterized in that: The training process of the XGBoost network model includes the following steps: Meta-feature output of graph convolutional network: the classification probability of the data in the validation set is used as the training meta-feature; The meta-feature output of the Alexnet network with the rotation-invariant attention mechanism module: The classification confidence of the validation sample set is combined to become the training meta-feature; Meta-feature output of capsule network: Take the L2 norm of the output vector to get the classification probability, and use the classification probability of the data in the validation sample set as the training meta-feature; The multi-learner optimized XGBoost network is trained using the training meta-features.

7. The space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners according to claim 6 is characterized in that: The training process of graph convolutional network, capsule network and Alexnet network is as follows: The Alexnet network model training process includes the following steps: During the training process, each category of the training sample library is divided into K equal parts, of which K-1 parts are training sample sets and the remaining part is the validation sample set. K-fold cross-validation is performed to ensure that all K parts of data have been used as validation samples. The Alexnet network model is trained using the training sample set and the validation sample set, and the classification confidence of the validation sample set is combined to become a training meta-feature; The graph convolutional network model training process includes the following steps: The training sample library is divided into a training set and a validation set. The training set has qshot images for each category, and the validation set has K-qshot images for each category. During the training process, 5×qshot+1 images are input into the network for each iteration, where 1 means taking one image from the validation set as validation in turn, and its label is known, which is used for supervised learning. The graph convolutional network model is trained using the training set and validation set, and the classification probability of the data in the validation set is used as the training meta-feature; The capsule network model training process includes the following steps: The training sample library is divided into a training set and a validation set. The capsule network model is trained using the training set and the validation set. The L2 norm of the output vector is taken to obtain the classification probability. The classification probability of the data in the validation sample set is used as the training meta-feature. The training sample library stores ISAR spin space target samples.

8. The space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners according to claim 7 is characterized in that: The ISAR spinning space target samples are obtained based on the three-dimensional space target model and the range Doppler algorithm simulation.

9. The space target ISAR image classification method under small sample conditions based on XGBoost optimized by multiple learners according to claim 8 is characterized in that: The ISAR spin space target samples stored in the training sample library are preprocessed samples, and the preprocessing process includes the following steps: Firstly, enhance the Lee filter on the ISAR image of the space target; Then contrast enhancement, energy normalization and data expansion are performed, where data expansion includes left-right flipping and up-down flipping.

Images