Incremental classification method for SAR targets based on stability and plasticity of equilibrium model

Abstract

The present invention discloses a SAR target incremental classification method based on the stability and plasticity of a balanced model, which mainly solves the problems of weak SAR feature extraction capability, poor target incremental classification and stability in complex environments in the existing technology. Its implementation scheme is as follows: randomly dividing the given SAR data into a training set and a test set; constructing a SAR target classification network including a feature extractor and a feature classifier, and training it using the training set; generating an example set to assist incremental learning; obtaining newly added category data from the SAR data, and using it together with the example set and the test set to generate an incremental task training set and a test set; constructing a SAR target incremental classification network and incremental learning loss, and training it using the incremental task training set; inputting the incremental task test set into the SAR target incremental classification network to obtain the target classification result. The present invention significantly improves the incremental classification performance of SAR targets in complex environments and can be used for battlefield reconnaissance and situational awareness.

Description

Technical Field

[0001] The present invention belongs to the field of radar remote sensing technology, and further relates to a SAR target incremental classification method, which can be used for continuous battlefield reconnaissance and situation awareness. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging sensor that uses electromagnetic waves to collect surface data over a wide area, day and night, in all weather conditions. Compared to optical sensors, SAR offers advantages such as all-day, all-weather coverage, long range, and high resolution. Therefore, it is widely used in a variety of military and civilian applications, including agricultural surveys, disaster monitoring, and battlefield reconnaissance.

[0003] With the widespread application of SAR systems and the gradual maturity of imaging algorithms, a large number of high-resolution, high-quality SAR images have become available. Among them, the interpretation technology of key targets has gradually attracted the attention of scholars and researchers in related fields. Most existing SAR target classification methods have static model structures. When faced with a continuously increasing SAR image data stream, the model cannot effectively learn the newly added target categories. In addition, due to different imaging mechanisms, single-polarization radar images have the characteristics of blurred edges and strong anisotropy of details, which makes it more difficult for the model to extract key features of the newly added categories. At the same time, SAR image scenes are complex, and the target classification model is easily affected by coherent speckle noise and non-target foreground scattering points. As a result, in the incremental SAR target classification task, it is difficult to learn new categories and easy to forget the learned categories, which seriously affects the effectiveness of image interpretation.

[0004] Traditional SAR target classification methods primarily rely on manual feature extraction. Given a SAR image to be classified, data enhancement and speckle removal preprocessing are first performed. Based on this, geometric features of the SAR target are manually extracted based on prior knowledge and used as classification evidence. Finally, a DS evidence fusion decision method is used to classify the target. These methods primarily rely on manual target feature extraction and classifier design to achieve target classification. Their performance relies heavily on professional knowledge and experience, making them difficult to achieve ideal results in complex and changing SAR scenarios.

[0005] Object classification methods based on deep networks enable data-driven object feature extraction and category determination, resulting in high performance and efficiency. Currently, mainstream approaches use deep convolutional networks as image feature extractors and fully connected networks for object category determination. Compared to traditional methods, these approaches do not rely on specialized knowledge and expert experience, resulting in higher object classification accuracy.

[0006] While the aforementioned methods provide effective approaches for SAR target classification, compared to optical images, SAR images are more complex, have fewer target samples, exhibit smaller inter-class variance, and exhibit greater intra-class variance. Consequently, these methods suffer from issues such as instability in complex environments, difficulty distinguishing similar classes, and difficulty classifying targets under varying imaging parameters. In particular, the classification performance of existing methods rapidly degrades when faced with SAR image data streams containing multiple classes.

[0007] Patent document CN201910657720.8 discloses a "SAR target recognition method based on transfer learning and fully connected layer outputs." This method consists of three main parts: first, it reads a training set and a test set of SAR remote sensing images containing ground vehicle targets, and performs image cropping and preprocessing; then, it extracts SAR image features using a deep convolutional network; and finally, it uses a fully connected network to predict the target category in the image. However, because this method does not optimize the characteristics of SAR images and lacks relevant designs for continuous learning, it suffers from poor target feature extraction capabilities in the incremental SAR target classification task, severely degrades classification performance, and uneven classification performance for different task targets. Summary of the Invention

[0008] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a SAR target incremental classification method based on balancing model stability and plasticity to improve the ability to extract target features in complex environments, reduce the rapid performance degradation of the model caused by the increase of incremental tasks, and balance the performance of target classification in different incremental learning tasks.

[0009] The technical idea of ​​the present invention is to improve the incremental classification performance of SAR targets in complex environments by designing a SAR target incremental classification model based on the stability and plasticity of the balanced model. The implementation steps include the following:

[0010] (1) Obtain SAR images with multiple target classes and corresponding labels from a given dataset and randomly divide them into training and test sets;

[0011] (2) Constructing SAR target classification network:

[0012] (2a) Establish a local feature extraction module A consisting of a cascade of convolution modules;

[0013] (2b) Establish a global feature extraction module B composed of self-attention;

[0014] (2c) connecting the local feature extraction module A and the global feature extraction module B in parallel to form a feature extractor F;

[0015] (2d) Establish a fully connected layer that matches the feature dimension and output category as the feature classifier C;

[0016] (2e) cascading the feature extractor and the feature classifier to form a SAR target classification network;

[0017] (3) Randomly sample a set of SAR images from the training set and input them into the SAR target classification network to calculate the cross entropy loss Based on this loss, the network parameters are updated by the stochastic gradient descent algorithm until the network converges to obtain a trained SAR target classification network;

[0018] (4) Generate an example set to assist incremental learning:

[0019] (4a) Using the Herding algorithm, select at most 10% representative samples from the training set generated in step (1);

[0020] (4b) The representative samples obtained by screening are stored as an example set, and the same number of samples are screened again after each incremental task to keep the number of samples in the example set unchanged;

[0021] (5) Generate incremental task training and test sets:

[0022] (5a) Obtain multiple SAR images containing newly added target categories from a given dataset and randomly divide them into incremental training sets and incremental test sets containing newly added target categories;

[0023] (5b) adding the example set samples generated in step (4) to the incremental training set to obtain the incremental task training set;

[0024] (5c) adding the test set samples generated in step (1) to the incremental test set to obtain the incremental task test set;

[0025] (6) Constructing the SAR target incremental classification network and incremental learning loss function:

[0026] (6a) Copying the SAR target classification network obtained in step (3), and performing dimension expansion on its feature classifier to obtain a SAR target incremental classification network;

[0027] (6b) Based on the SAR target classification network of step (3) and the feature extractor of the SAR target incremental classification network obtained in step (6a), a feature-level distillation loss function is constructed.

[0028] (6c) Based on the SAR target classification network of step (3) and the feature classifier of the SAR target incremental classification network obtained in step (6a), a decision-level distillation loss function is constructed.

[0029] (6d) Construct a feature separation loss function based on the feature extractor of the SAR target incremental classification network

[0030] (6e) According to the cross entropy loss function Feature-level distillation loss function Decision-level distillation loss function Feature separation loss function Get the incremental learning loss function of the SAR target incremental classification network

[0031]

[0032] Among them, N cur Indicates the number of newly added target categories in the current incremental task; N C It represents the number of target categories that have been trained by the SAR target incremental classification network. α is a dynamic adaptive weight that decreases as the number of trained categories increases.

[0033] (7) Randomly sample a set of SAR images from the incremental task training set generated in step (5) and input them into the SAR target incremental classification network to calculate the incremental learning loss Based on this loss, the network parameters are updated by the stochastic gradient descent algorithm until the network converges, and a trained SAR target incremental classification network is obtained;

[0034] (8) Input the SAR images in the incremental task test set into the trained SAR target incremental classification network to obtain the classification results.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] First, the present invention designs a new SAR target classification network structure, which uses convolutional networks and self-attention structures to extract local features and global features from input SAR images, and merges features at two different levels, so that the network can adapt to the interference of background clutter and noise in complex environments, thereby improving the classification network's feature extraction ability for SAR image targets.

[0037] Second, this paper designs a new incremental learning loss function. This function maintains previously learned classification knowledge through distillation learning, without storing all previous task samples, and enables learning of newly learned classification knowledge through feature separation. This improves the model's ability to classify newly added target categories while reducing catastrophic forgetting of old categories, thereby increasing the classification accuracy of each incremental task.

[0038] Third, the present invention designs an adaptive dynamic loss function structure. By adjusting the weight coefficients of each part of the loss function in different tasks, the model's maintenance of the learned category classification knowledge and the learning of new category classification knowledge are enhanced, thereby balancing the classification performance of different target categories in incremental learning tasks and improving the average classification accuracy of all incremental tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Flowchart for the implementation of the present invention;

[0040] Figure 2 Schematic diagram of the SAR target classification network structure constructed in the present invention;

[0041] Figure 3 Schematic diagram of the incremental classification method for SAR targets constructed in the present invention;

[0042] Figure 4 A comparison chart of the target incremental classification results achieved by the present invention and existing methods under different task settings. DETAILED DESCRIPTION

[0043] The examples and effects of the present invention are further described in detail below with reference to the accompanying drawings.

[0044] refer to Figure 1 , the implementation steps of this example are as follows:

[0045] Step 1: Generate training and test sets from the given dataset.

[0046] SAR images with multiple categories of targets and their corresponding labels are obtained from a given dataset and randomly divided into training and test sets.

[0047] In this example, the SAR image dataset is from the MSTAR radar dataset, which provides data on the acquisition and recognition of moving and stationary targets provided by the U.S. Defense Advanced Research Projects Agency (DARPA) and the U.S. Air Force Research Laboratory (AFRL). The SAR images have a resolution of 128 × 128 pixels and contain 10 types of ground military vehicles. The training set contains 3,671 images, and the test set contains 3,203 images.

[0048] Step 2: Construct a SAR target classification network.

[0049] refer to Figure 2 , the implementation of this step is:

[0050] 2.1) Build feature extractor F:

[0051] 2.1.1) Establish a local feature extraction module A consisting of five cascaded convolution modules a1, a2, a3, a4, and a5, where:

[0052] The first convolution module a1 consists of a 7×7 standard convolution layer, a batch normalization layer, a ReLU activation layer, and a maximum pooling downsampling layer cascade. The number of channels of the output feature map of this convolution module is 8;

[0053] The second convolution module a2 consists of a residual block with an input dimension of 8 and an output dimension of 16. The number of channels of the output feature map of this convolution module is 16;

[0054] The third convolution module a3 consists of a residual block with an input dimension of 16 and an output dimension of 32. The number of channels of the output feature map of this convolution module is 32;

[0055] The fourth convolution module a4 consists of a residual block with an input dimension of 32 and an output dimension of 64. The number of channels of the output feature map of this convolution module is 64;

[0056] The fifth convolution module a5 consists of a residual block with an input dimension of 64 and an output dimension of 128. The number of channels of the output feature map of this convolution module is 128.

[0057] 2.1.2) Establish a global feature extraction module B consisting of 5 cascaded self-attention modules b1, b2, b3, b4, b5, where: each self-attention module b i Calculation module by query value Key value calculation module Value calculation module Position encoding module Attention calculation module It consists of five parts.

[0058] In an embodiment of the present invention, the query value calculation module Key value calculation module Value calculation module Both consist of a 1×1 standard convolutional layer and a position encoding module It consists of a 1×1 standard convolution layer with 2 input channels and an attention calculation module. It consists of 1 sliding window unit and 1 Softmax calculation unit.

[0059] 2.1.3) Connect the local feature extraction module A and the global feature extraction module B in parallel. That is, connect the five convolution modules of the local feature extraction module A and the five self-attention modules of the global feature extraction module B in parallel to form the feature extractor F of the SAR target classification network.

[0060] The local feature extraction module A is used to extract the local features of the target in the SAR image. For an input SAR image with a width and height of W and H respectively The output of the local feature extraction module can be expressed as where Y i is the i-th convolution module a i Output features of Z i-1 is the i-1th feature extraction module f i-1 Output features of

[0061] The global feature extraction module B is used to extract the global features of the target in the SAR image. The output of the global feature extraction module B can be expressed as:

[0062]

[0063] Among them, U i is the i-th self-attention module b i The output feature map of ; UnFold(·) represents a sliding window unit in the same way as the convolutional neuron sliding window; Softmax(·) represents the unit that calculates Softmax along the first dimension; d represents the normalization scale, whose value is the number of output channels;

[0064] Q i ,K i ,V i are the i-th self-attention module b i Query value calculation module Key value calculation module and value calculation module Output; P i is the i-th self-attention module b i Position encoding module Output;

[0065] Q i ,K i ,V i ,P i The output expressions are as follows:

[0066]

[0067] Among them, Pos(·) represents the position encoding operation, which outputs a position information matrix containing 2 channels and the same dimension as the channel matrix of the input feature, and the values ​​of the position information matrix are uniformly distributed on (-1, 1); Conv1×1(·) represents a 1×1 standard convolution;

[0068] The output Z of the feature extractor F i Expressed as:

[0069]

[0070] Among them, f i represents the i-th module of the feature extractor, and (·||·) means concatenating the two feature maps along the channel dimension.

[0071] 2.2) Establish feature classifier C:

[0072] In an embodiment of the present invention, the feature classifier C includes a fully connected network layer FCN, whose parameters and output are expressed as follows:

[0073] The bias of the fully connected network layer FCN is 0, and the dimension of the weight matrix is ​​H×N C , where H is the dimension of the feature vector output by the feature extractor F, and N C Indicates the number of all categories that the target classification model has learned;

[0074] The output of the feature classifier C is expressed as:

[0075] O=C(Z5)=FCN(GAP(Z5)),

[0076] Among them, O represents the output of the feature classifier; FCN(·) is the fully connected network layer; GAP(·) represents global average pooling; Z5 is the output feature of the feature extractor f5 module.

[0077] 2.3) The established feature extractor F and feature classifier C are cascaded to form a SAR target classification network.

[0078] Step three: train the SAR target classification network.

[0079] 3.1) Randomly sample a set of SAR images from the training set generated in step 1 and input them into the SAR target classification network to calculate the cross entropy classification loss Based on this loss, the network parameters are updated using the stochastic gradient descent algorithm:

[0080] 3.1.1) Calculate the cross entropy classification loss of the SAR target classification network

[0081]

[0082] Among them, N B is the number of samples in each batch, c n and are the output labels of the SAR target classification network and the true labels corresponding to the input images;

[0083] 3.1.2) Solve the cross entropy classification loss in 3.1.1) Gradient of the SAR target classification network parameter θ:

[0084]

[0085] in, is the loss of the SAR target classification network, θ is the learnable parameter of the SAR target classification network;

[0086] 3.1.3) According to the gradient solved in 3.1.2) Update the SAR target classification network parameters, expressed as:

[0087]

[0088] Wherein θ′ is the currently updated network parameter, θ is the network parameter before the update; lr is the learning rate, which is set according to the input image batch size. In the embodiment of the present invention, lr=0.01.

[0089] 3.2) Repeat the process of 3.1) until the network loss tends to be stable and all network parameters converge. Stop updating and obtain the trained SAR target classification network.

[0090] Step 4: Generate an example set to assist incremental learning.

[0091] 4.1) Use the Herding algorithm to select representative samples from the training set generated in step 1:

[0092] 4.1.1) Using the SAR target classification network trained in step 3, calculate the feature mean μ of each class sample in the training set:

[0093]

[0094] Where n represents the number of training samples in a certain category; Z5| x Represents the output of the feature extractor f5 module for the input sample x.

[0095] 4.1.2) Select representative samples p of each category that meet the conditions in the training set based on the feature average μ. In the embodiment of the present invention, the screening conditions for representative samples are as follows:

[0096]

[0097] Where m represents the number of representative samples screened for each category.

[0098] 4.2) The representative samples obtained by screening are stored as an example set, and the same number of samples are re-screened after each incremental task to keep the number of samples in the example set unchanged.

[0099] Step 5: Generate incremental task training and test sets.

[0100] 5.1) Obtain multiple SAR images and corresponding labels containing the newly added target categories from the given dataset and randomly partition them to obtain incremental training sets and incremental test sets containing the newly added target categories;

[0101] 5.2) Add the example set samples generated in step 4 to the incremental training set to obtain the incremental task training set;

[0102] 5.3) Add the test set samples generated in step 1 to the incremental test set to obtain the incremental task test set.

[0103] Step 6: Construct the SAR target incremental classification network and incremental learning loss function.

[0104] Reference Figure 3 , this step is implemented as follows:

[0105] 6.1) Copy the SAR target classification network obtained in step 3 and expand the dimension of its feature classifier to obtain a SAR target incremental classification network;

[0106] 6.2) Based on the feature extractor of the SAR target classification network obtained in step 3 and the SAR target incremental classification network obtained in step (6.1), construct a feature-level distillation loss function

[0107]

[0108] Among them, Z i represents the output of the feature extractor of the classification network obtained in step (6.1), Z i ′ represents the output of the feature extractor of the classification network trained in step 3; Pool avg (·) indicates that the feature map is pooled into a one-dimensional feature vector; Dis cos (·,·) means calculating the cosine distance between two vectors; C pear (·,·) means calculating the Pearson coefficient of two vectors;

[0109] 6.3) Based on the feature classifier of the SAR target classification network obtained in step 3 and the SAR target incremental classification network obtained in step (6.1), a decision-level distillation loss function is constructed.

[0110]

[0111] Where O represents the output of the feature classifier of the classification network obtained in step (6.1), and O′ represents the output of the feature classifier of the classification network trained in step 3; Indicates the calculation of the Kullback-Leibler divergence of O and O′; It means to perform softmax distillation on the output distribution of the feature classifier of the classification network trained in step 3. The formula is as follows:

[0112]

[0113] Among them, softmax(O′)| cur Indicates the softmax calculation of the output of the classification head corresponding to the newly added categories of the incremental task in O′; Represents the true category label of the input sample; represents the category set of the current incremental task learning; g i is an identifier. When the input training sample category does not belong to the current incremental task category, the value is 1, otherwise it is 0;

[0114] 6.4) Based on the output of the feature extractor of the SAR target classification network obtained in step 3 and the SAR target incremental classification network obtained in step (6.1), construct the feature separation loss function

[0115]

[0116] Among them, Z5| x represents the output of the feature extractor f5 module of the incremental classification network for the input sample x, Z′5| x represents the output of the feature extractor f5 module of the classification network trained in step 3; N B is the number of samples in a batch; k i,j Is an identifier, when the input sample x i And the input sample x j The value is 1 if they belong to the same category, otherwise it is 0;

[0117] 6.5) Construct an adaptive dynamic loss function structure and convert the cross entropy loss Characteristic distillation loss Decision-level distillation loss Feature separation loss Integrate to get the incremental learning loss function It is expressed as follows:

[0118]

[0119] Among them, N cur Indicates the number of target categories for the current incremental task learning; N C Represents the number of all categories that the target classification model has learned; α is a dynamic adaptive weight that gradually decreases as the number of categories that have completed training increases.

[0120] Step 7: Train the SAR target incremental classification network.

[0121] 7.1) Randomly sample a set of SAR images from the incremental task training set generated in step 5 and input them into the SAR target incremental classification network to calculate the incremental learning loss. Based on this loss, the network parameters are updated using the stochastic gradient descent algorithm;

[0122] 7.1.1) Calculate the incremental loss function of the SAR target incremental classification network constructed in step 6 Solve the loss for the network parameters θ inc Gradient:

[0123]

[0124] in, is the incremental learning loss function, θ inc Learnable parameters for the SAR target incremental classification network;

[0125] 7.1.2) According to the gradient solved in 7.1.1) Update the SAR target incremental classification network parameters:

[0126]

[0127] where θ i ' nc is the currently updated network parameter, θ inc is the network parameter before updating; lr is the learning rate, which is set according to the input image batch size. In this embodiment, lr = 0.01;

[0128] 7.2) Repeat the process of 7.1) until the network loss tends to be stable and all network parameters converge, stop updating, and obtain the trained SAR target incremental classification network.

[0129] Step 8: Obtain SAR target classification results.

[0130] The SAR images in the incremental task test set generated in step 5 are input into the SAR target incremental classification network trained in step 7 to obtain the SAR target classification results.

[0131] The effect of the present invention can be further illustrated by the following simulation experiments:

[0132] 1. Simulation experiment conditions:

[0133] The software platform of the simulation experiment of the present invention is: Ubuntu18.04 operating system and Pytorch 1.8.0, and the hardware configuration is: Core i9-10980XE CPU and NVIDIA GeForce RTX 3090 GPU.

[0134] The simulation experiment of the present invention uses SAR measured data from the MSTAR dataset, where the SAR image scale is 128×128, contains 10 types of ground military vehicle targets, the number of training set images is 3671, and the number of test set images is 3203.

[0135] During the experiment, two task settings, S1 and S2, were used. Both S1 and S2 consisted of an initial learning phase and four incremental learning phases. S1 initially learned two categories of objectives, then evenly divided the remaining eight categories into four incremental tasks. S2 initially learned six categories of objectives, then evenly divided the remaining four categories into four incremental tasks.

[0136] 2. Simulation content and result analysis:

[0137] Under the above simulation conditions, the present invention and the existing "SAR target recognition method based on transfer learning and fully connected layer output" were used to conduct experiments on the two task settings S1 and S2 respectively. The results are as follows Figure 4 .in:

[0138] Figure 4 (a) is a graph showing the accuracy change of the SAR target incremental classification experiment under the S1 task setting, wherein the solid line represents the target classification accuracy of the present invention in the classification experiment, and the dotted line represents the target classification accuracy of the prior art in the classification experiment;

[0139] Figure 4 (b) is a curve chart showing the accuracy change of the SAR target incremental classification experiment under the S2 task setting, wherein the solid line represents the target classification accuracy of the present invention in the classification experiment, and the dotted line represents the target classification accuracy of the prior art in the classification experiment.

[0140] according to Figure 4 ,The classification indicators of the present invention and the prior art on the test set images are sorted out, including the classification accuracy of SAR targets in the five incremental tasks, and the average classification accuracy of all incremental tasks. The results are shown in Table 1.

[0141] Table 1 SAR target incremental classification experimental performance indicators

[0142]

[0143] from Figure 4As can be seen from Table 1, the classification accuracy of each incremental task and the average classification accuracy of the present invention are better than those of the existing technology solutions, indicating that the target incremental classification performance of the present invention is significantly better than that of the existing technology.

Claims

1. A SAR target incremental classification method based on the stability and plasticity of the equilibrium model, characterized by: The following steps are involved: (1) Obtain SAR images with multiple target classes and corresponding labels from a given dataset and randomly divide them into training and test sets; (2) Constructing SAR target classification network: (2a) Establish a local feature extraction module A consisting of a cascade of convolution modules; (2b) Establish a global feature extraction module B composed of self-attention; (2c) connecting the local feature extraction module A and the global feature extraction module B in parallel to form a feature extractor F; (2d) Establish a fully connected layer that matches the feature dimension and output category as the feature classifier C; (2e) cascading the feature extractor and the feature classifier to form a SAR target classification network; (3) Randomly sample a set of SAR images from the training set and input them into the SAR target classification network to calculate the cross entropy loss Based on this loss, the network parameters are updated by the stochastic gradient descent algorithm until the network converges to obtain a trained SAR target classification network; (4) Generate an example set to assist incremental learning: (4a) Using the Herding algorithm, select at most 10% representative samples from the training set generated in step (1); (4b) The representative samples obtained by screening are stored as an example set, and the same number of samples are screened again after each incremental task to keep the number of samples in the example set unchanged; (5) Generate incremental task training and test sets: (5a) Obtain multiple SAR images and corresponding labels containing the newly added target categories from the given dataset, and randomly divide them into incremental training sets and incremental test sets containing the newly added target categories; (5b) adding the example set samples generated in step (4) to the incremental training set to obtain the incremental task training set; (5c) adding the test set samples generated in step (1) to the incremental test set to obtain the incremental task test set; (6) Constructing the SAR target incremental classification network and incremental learning loss function: (6a) Copying the SAR target classification network obtained in step (3), and performing dimension expansion on its feature classifier to obtain a SAR target incremental classification network; (6b) Based on the SAR target classification network of step (3) and the feature extractor of the SAR target incremental classification network obtained in step (6a), a feature-level distillation loss function is constructed. (6c) Based on the SAR target classification network of step (3) and the feature classifier of the SAR target incremental classification network obtained in step (6a), a decision-level distillation loss function is constructed. (6d) Construct a feature separation loss function based on the feature extractor of the SAR target incremental classification network (6e) According to the cross entropy loss function Feature-level distillation loss function Decision-level distillation loss function Feature separation loss function Get the incremental learning loss function of the SAR target incremental classification network Among them, N cur Indicates the number of new target categories added in the current incremental task; N C It represents the number of target categories that have been trained by the SAR target incremental classification network. α is a dynamic adaptive weight that decreases as the number of trained categories increases. (7) Randomly sample a set of SAR images from the incremental task training set generated in step (5) and input them into the SAR target incremental classification network to calculate the incremental learning loss Based on this loss, the network parameters are updated by the stochastic gradient descent algorithm until the network converges, and a trained SAR target incremental classification network is obtained; (8) Input the SAR images in the incremental task test set into the trained SAR target incremental classification network to obtain the classification results.

2. The method according to claim 1, characterized in that The convolution module that constitutes the local feature extraction module A in step (2a) includes five cascaded convolution modules a1, a2, a3, a4, and a5. The structural parameters of each convolution module and the output features of the entire local feature extraction module are as follows: The first convolution module a1 is composed of a 7×7 standard convolution layer, a batch normalization layer, a ReLU activation layer, and a maximum pooling downsampling layer in cascade. The number of channels of the output feature map of this convolution module is 8; The second convolution module a2 consists of a residual block with an input dimension of 8 and an output dimension of 16. The number of channels of the output feature map of this convolution module is 16; The third convolution module a3 consists of a residual block with an input dimension of 16 and an output dimension of 32. The number of channels of the output feature map of this convolution module is 32; The fourth convolution module a4 is composed of a residual block with an input dimension of 32 and an output dimension of 64. The number of channels of the output feature map of this convolution module is 64; The fifth convolution module a5 consists of a residual block with an input dimension of 64 and an output dimension of 128. The number of channels of the output feature map of this convolution module is 128; The output feature of the entire local feature extraction module is expressed as where Y i is the i-th convolution module a i The output features of X are the input SAR images with width and height of W and H respectively, and Z i-1 is the i-1th feature extraction module f in the SAR target classification network i-1 The output features of .

3. The method according to claim 1, characterized in that The self-attention module that constitutes the global feature extraction module B in step (2b) includes five cascaded self-attention modules b1, b2, b3, b4, and b5. The structural parameters of each self-attention module and the output feature map of the entire global feature extraction module B are as follows: Each self-attention module b i , by the query value calculation module Key value calculation module Value calculation module Position encoding module Attention calculation module The query value calculation module consists of five parts. Key value calculation module Value calculation module Each consists of a 1×1 standard convolutional layer. The position encoding module It consists of a 1×1 standard convolution layer with 2 input channels. It consists of 1 sliding window unit and 1 Softmax calculation unit; The output feature map of the entire global feature extraction module B is expressed as: Among them, U i is the i-th self-attention module b i The output feature map of X is the input SAR image with width and height of W and H respectively; Z i-1 is the i-1th feature extraction module f in the SAR target classification network i-1 The output features of ; UnFold(·) represents a sliding window unit in the same way as the convolutional neuron sliding window; Softmax(·) represents a unit that calculates Softmax along the first dimension; d represents the normalized scale, which is the number of output channels; Q i ,K i ,V i are the i-th self-attention module b i Query value calculation module Key value calculation module and value calculation module Output; P i is the i-th self-attention module b i Position encoding module The output, Q i ,K i ,V i ,P i is expressed as: Among them, Pos(·) represents the position encoding operation, which outputs a position information matrix containing 2 channels and with the same dimension as each channel matrix of the input feature, and the values ​​of the position information matrix are uniformly distributed on (-1, 1); Conv1×1(·) represents a 1×1 standard convolution.

4. The method according to claim 1, wherein The feature extractor F in step (2c) consists of five cascaded feature extraction modules f1, f2, f3, f4, and f5, which are composed of the corresponding modules in the convolution modules a1, a2, a3, a4, and a5 and the self-attention modules b1, b2, b3, b4, and b5 in parallel; The output feature map of the entire feature extractor F is expressed as: Where (·||·) represents the parallel concatenation of two feature maps along the channel dimension.

5. The method according to claim 1, wherein The fully connected network layer FCN in the feature classifier C in step (2d) has the following parameters and outputs: The bias of the fully connected network layer FCN is 0, and the dimension of the weight matrix is ​​H×N C , where H is the dimension of the feature vector output by the feature extractor F, and N C Indicates the number of all categories that the target classification model has learned; The output feature map of the entire feature classifier C is expressed as: O=C(Z5)=FCN(GAP(Z5)), Among them, O represents the output of the feature classifier; FCN(·) is the fully connected network layer; GAP(·) represents global average pooling; Z5 is the output feature of the feature extraction module f5.

6. The method according to claim 1, characterized in that Calculate the cross entropy loss in step (3) The formula is as follows: Among them, N B is the number of samples in each batch, c n and are the output labels of the SAR target classification network and the true labels corresponding to the input images, respectively.

7. The method according to claim 1, characterized in that In step (3), the network parameters are updated using stochastic gradient descent method, which is implemented as follows: (3a) The gradient of the SAR target classification network parameters is solved and expressed as: in is the loss of the SAR target classification network, θ is the learnable parameter of the SAR target classification network; (3b) According to the gradient of the solution Update the parameters of the SAR target classification network, expressed as: Where θ′ is the current network parameter, θ is the network parameter of the previous iteration; lr is the learning rate, which is set according to the number of samples in each batch of input; (3c) Repeat the above operations (3a) and (3b) until the network loss tends to be stable and all network parameters converge, and then stop updating.

8. The method according to claim 1, characterized in that In step (4a), the Herding algorithm is used to select representative samples, which is implemented as follows: (4a1) Using the SAR target classification network trained in step (3), calculate the feature mean μ of each category sample in the training set: Among them, n represents the number of training samples of a certain category, Z5| x Represents the output features of the input sample x in the feature extraction module f5; (4a2) Select representative samples p of each category in the training set that meet the following conditions based on the feature mean μ, Where m represents the number of representative samples screened for each category.

9. The method according to claim 1, characterized in that The feature-level distillation loss function constructed in step (6b) It is expressed as follows: Among them, Z i represents the output of the feature extractor of the classification network obtained in step (6a), Z′ i represents the output of the feature extractor of the classification network trained in step (3); Pool avg (·) indicates that the feature map is pooled into a one-dimensional feature vector; Dis cos (·,·) means calculating the cosine distance between two vectors; C pear (·,·) means calculating the Pearson coefficient of two vectors.

10. The method according to claim 1, characterized in that The decision-level distillation loss function constructed in step (6c) It is expressed as follows: Wherein, O represents the output of the feature classifier in the SAR target incremental classification network obtained in step (6a), and O′ represents the output of the feature classifier in the SAR target classification network trained in step (3); Indicates the calculation of the Kullback-Leibler divergence of two distributions; Indicates that softmax distillation is performed on the distribution, which is expressed as: Among them, soft max(·)| cur Indicates that the softmax calculation is performed on the output distribution of the classifier weight corresponding to the current incremental task learning category in the feature classifier C; Represents the true category label of the input sample; represents the category set of the current incremental task learning; g i is an identifier, which is 1 when the input training sample category does not belong to the current incremental task category, otherwise it is 0.

11. The method according to claim 1, wherein The feature separation loss function constructed in step (6d) It is expressed as follows: Among them, Z5| x and Z′5| x Respectively represent the output features of the feature extraction module f5 of the input sample x in the model trained for the current classification model and the previous task; N B is the number of samples in a batch; k i,j Is an identifier, when the input sample x i And the input sample x j The value is 1 if they belong to the same category, and 0 otherwise.

Images