SAR target classification method combining local classification with feature generation and correction

Abstract

The present invention provides a SAR target classification method that combines local classification with feature generation and correction. This method fully considers the contribution of key local detail features of SAR image targets to classification, constructing a target classification model that combines local classification with feature generation and correction. This target classification model is used for SAR target recognition in the presence of a small number of samples. Under the guidance of a classification network (MLC-Net), the model generates omnidirectional support features to compensate for the real support features in a small number of azimuth angles, thereby eliminating prototype deviations and improving SAR target recognition accuracy. The present invention generates omnidirectional SAR target features for each class and uses them together with the features of the real labeled SAR image to calculate the prototype, further improving SAR target classification performance. Therefore, the present invention can improve SAR target classification accuracy with limited training samples and further improve the classification accuracy of test SAR images.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target recognition, and in particular relates to a SAR target classification method combining local classification with feature generation and correction. Background Art

[0002] Synthetic aperture radar (SAR), with its advantages of daytime and all-weather operation, is widely used in fields such as resource exploration and military reconnaissance. SAR target classification is a key technology for intelligent interpretation of SAR images. SAR target classification methods based on convolutional neural networks (CNNs) rely on a large amount of accessible labeled data and have received increasing attention in recent years. However, acquiring SAR images is time-consuming. Furthermore, because SAR images reflect radar backscatter from ground objects, targets in SAR images visually differ from actual targets, making manual labeling of SAR image target categories difficult. The performance of CNN-based SAR target classification methods is therefore limited by the limited number of labeled SAR images. Therefore, research on SAR target classification with a small number of labeled images is essential. For SAR target classification with limited samples, data augmentation (such as translation and rotation) and reducing the number of free parameters of the network are commonly used methods. Ding et al. used various data augmentation methods, such as random translation and pose synthesis, to expand the training data to obtain more samples for network training. Chen et al. designed a fully convolutional network to minimize the number of free parameters of the network and reduce the risk of overfitting during training. In recent years, transfer learning has also been frequently applied to SAR target classification with limited samples, aiming to transfer rich knowledge learned from the source domain to the target domain. For example, Huang et al. utilized transferable knowledge extracted from pre-trained convolutional layers to improve SAR target classification. While these methods achieve good recognition performance for SAR target recognition with limited samples, their classification performance degrades significantly in extreme cases (e.g., with only five training samples per class). Few-shot learning based on meta-learning has received increasing attention in recent years. The goal is to learn a classifier from a source domain with a large number of labeled images and generalize it to an unseen target domain with a small number of labeled images (typically one to ten training samples per class). Among these, few-shot learning methods based on metric meta-learning have been widely used in the field of SAR target classification due to their simplicity and high classification performance. They aim to learn an embedding space in which features of samples of the same type are clustered together, while features of different types are separated. The class label of the image is then predicted by comparing the correlation between the global features of the image and the prototype of each class. Gao et al. proposed a weighted distance classifier to adaptively weight the similarity scores between features, thereby increasing the interval between features of different categories and improving classification performance.

[0003] Patent document (CN111832580A, SAR target recognition method combining few-shot learning and target attribute features) introduces a SAR target recognition method that combines few-shot learning and target attribute features. This method uses few-shot learning to improve the classification performance of the learned features, and further improves the classification performance by combining the target attribute features of the samples. However, since this patent still uses a metric-based meta-learning method, the prototype obtained with a small number of labeled samples still has deviations, so the results of this method still have considerable room for improvement.

[0004] Current SAR target few-shot recognition methods based on metric meta-learning primarily compute prototypes for each class and then predict the class of each sample by querying the similarity between the sample's global features and the prototype. On the one hand, prototypes computed from supporting samples at limited azimuth angles are biased, leading to discrepancies in SAR target recognition results. On the other hand, for SAR images with high inter-class similarity, directly calculating the similarity between the sample's global features and the prototype in metric meta-learning can weaken the contribution of the target's key local features to classification. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a SAR target classification method that combines local classification with feature generation and correction. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] The present invention provides a SAR target classification method that combines local classification with feature generation and correction. This method fully considers the contribution of key local detail features of SAR image targets to classification and constructs a target classification model (LcFGC) that combines local classification with feature generation and correction. The multi-scale local classification network (MLC-Net) in the target classification model is used for SAR target recognition in the case of few samples. The feature generation and correction network (FGC-Net) in the target classification model is dedicated to generating supporting features at all angles under the guidance of the classification network (MLC-Net) to compensate for the real supporting features at a few azimuth angles, thereby eliminating prototype deviations and achieving the purpose of improving SAR target recognition accuracy. The present invention generates SAR target features at all angles for each class and calculates the prototype together with the features of the real labeled SAR image. The multi-scale local classification network is combined with the multi-scale local classification network to further improve SAR target classification performance. Therefore, the present invention can improve the classification accuracy of SAR targets with limited training samples and further improve the classification accuracy of test SAR images.

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1This is a flow chart of a SAR target classification method combining local classification with feature generation and correction provided by the present invention;

[0009] Figure 2 This is a diagram showing the overall structural framework of a SAR target classification method that combines local classification with feature generation and correction, provided by an embodiment of the present invention;

[0010] Figure 3 1 is a structural framework diagram of a multi-scale feature extraction module provided by an embodiment of the present invention;

[0011] Figure 4 This is the structural framework diagram of the conversion module provided by this example;

[0012] Figure 5 This is the structural framework diagram of the channel attention module provided by this example;

[0013] Figure 6 This is the structural framework diagram of the generated network provided by this example;

[0014] Figure 7 This is the structural framework diagram of the discriminant network provided in this example. DETAILED DESCRIPTION

[0015] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0016] like Figure 1 As shown, the present invention provides a SAR target classification method combining local classification with feature generation and correction, including:

[0017] Step 1: Obtain SAR images collected in real time from the field, a training set in the source domain, a support set in the target domain, and a test set;

[0018] In a specific embodiment, step 1 includes:

[0019] Step 1a: Obtain the moving and stationary target MSTAR dataset D of 10 types of targets with category labels and azimuths, as well as SAR images collected in real time from the field;

[0020] Step 1b, using the SAR image collected in real time on site as the SAR image to be classified;

[0021] Step 1c: crop each SAR image in the MSTAR dataset D and the SAR image to be classified into a size of 64×64 pixels;

[0022] Step 1d: perform L2 normalization on each SAR image in the cropped MSTAR dataset D and the SAR image to be classified, so as to obtain the preprocessed moving and stationary target MSTAR dataset D.p and the processed SAR image to be classified;

[0023] In step 1d, L2 normalization is performed, and the implementation process is as follows:

[0024]

[0025] Among them, ij represents the pixel value of the SAR image at the position of row i and column j, Represents the pixel value at the i-th row and j-th column of each SAR image after L2 normalization.

[0026] Step 1e, from the preprocessed MSTAR dataset D p Randomly select n types of SAR images with category labels and azimuth angles at a pitch angle of 15° as the test set D of the target domain test ; D p The remaining (10-n) types of SAR images with 17° elevation angles with category labels and azimuth angles are used as the training set D in the source domain. train ; Randomly select S SAR images for each category from the selected n categories of SAR images with category labels and azimuth angles at a pitch angle of 17° as the support set D of the target domain s ;

[0027] Among them, n=3, S≤20.

[0028] Step 2: constructing a target classification model LcFGC that combines local classification, feature generation, and correction networks; wherein the target classification model is composed of a multi-scale feature extraction module, a generation network, a discriminant network, an adaptive feature correction module, and a local feature classification module stacked in sequence;

[0029] refer to Figure 2 The multi-scale feature extraction module and local feature classification module form a multi-scale local classification network MLC-Net; the generation network, discriminant network, and adaptive feature correction module form a feature generation and correction network FGC-Net. The multi-scale feature extraction module is used to extract features from the input support image and query image, respectively obtaining true support features and query features. The generation network is used to generate support features under specified category labels and azimuth angles. The discriminant network D is used as the discriminant network D of the feature generation adversarial network FGAN. The adaptive feature correction module AFCM is used to perform distribution correction on the input generated support features under all azimuth angles. The local feature classification module LC is used to predict the category label of the query feature.

[0030] refer to Figure 3The multi-scale feature extraction module consists of four identical convolution modules Conv, three identical downsampling modules DS and a transformation module Tran; the four identical convolution modules Conv are respectively the 0_1 convolution module, the 0_2 convolution module, the 0_3 convolution module, and the 0_4 convolution module; the three identical downsampling modules are respectively the 0_1 downsampling module, the 0_2 downsampling module and the 0_3 downsampling module;

[0031] The input of the 0_1th sampling module is the output of the 0_1th convolution module. The input of the remaining 0_1th downsampling modules is the output of the 0_(l-1)th downsampling module and the output of the 0_1th convolution module, where l = 2, 3. The output of the 0_3th downsampling module is spliced with the output of the 0_4th convolution module according to the channel and input to the transformation module Tran.

[0032] Among them, the four identical convolution modules Conv include the 0_1th convolution layer, the 0_1th batch normalization layer, the 0_1th ReLU activation layer and the 0_1th maximum pooling layer stacked in sequence; the parameters of the convolution module Conv are specifically set as follows: the convolution kernel size of the 0_1th convolution layer is 3×3, and the convolution step is 1; the pooling area size of the 0_1th maximum pooling layer is set to 2×2, and the pooling step is 2; the downsampling module DS has the same structure as the convolution module Conv;

[0033] refer to Figure 4 ,The transformation module Tran is composed of the 0_2th batch normalization layer, the channel attention module CA, the convolution module Conv0 and the reshape layer stacked in sequence;

[0034] refer to Figure 5 The channel attention module CA is composed of the 0_1th parallel layer, the 0_1th fully connected layer, the 0_2th fully connected layer, the 0_1th addition module, the 0_1th sigmoid activation layer and the 0_1th multiplication module stacked in sequence; wherein the 0_1th parallel layer is composed of the average pooling layer and the 0_2th maximum pooling layer in parallel; the average pooling layer and the 0_2th maximum pooling layer of the 0_1th parallel layer are the average and maximum values of the features of each channel of the multi-channel features of the input CA, and the average pooling is obtained. Features and maximum pooling features; the 0_1th fully connected layer and the 0_2th fully connected layer transform the average pooling features and maximum pooling features output by the 0_1th parallel layer; the 0_1th addition module adds the transformed average pooling features and the transformed maximum pooling features output by the 0_2th fully connected layer element by element, and outputs the channel attention features; the 0_1th multiplication module multiplies the channel attention features activated by the 0_1th sigmoid activation layer by the features of the original input CA element by element;

[0035] The convolution module Conv0 consists of the 1_1th convolution layer, the 1_1th batch normalization layer and the 1_1th ReLU activation layer stacked in sequence;

[0036] The parameters of each layer in the channel attention module CA are specifically set as follows: the pooling area size of the average pooling layer of the 0_1th parallel layer is set to 4×4, and the pooling step is 1; the pooling area size of the 0_2th maximum pooling layer of the 0_1th parallel layer is set to 4×4, and the pooling step is 1; the number of nodes in the output of the 0_1th fully connected layer and the 0_2th fully connected layer are 64 and 256 respectively;

[0037] The parameter settings of the convolution module Conv0 are the same as those of the first convolution module Conv.

[0038] refer to Figure 6 , the generating network G is composed of the first stacking layer, the first fully connected layer, the first spectral normalization layer, the first ReLU activation layer, the second stacking layer, the second fully connected layer, the second spectral normalization layer, the first layer normalization layer, the second ReLU activation layer, the third stacking layer, the third fully connected layer, the third spectral normalization layer and the third ReLU activation layer;

[0039] The first overlay layer is used to concatenate the random vector, feature category label, and feature orientation information of the input generative network G. The second and third overlay layers are used to concatenate the output of the upper network with the feature category label and feature orientation information. The parameters of each layer of the generative network are set as follows:

[0040] The number of output nodes of the first fully connected layer is set to 256; the number of output nodes of the second fully connected layer is set to 512; and the number of output nodes of the third fully connected layer is set to 1024.

[0041] refer to Figure 7 , the discriminant network D is composed of the 1_1th overlay layer, the 1_1th fully connected layer, the 1_1th spectral normalization layer, the 1_1th Leaky ReLU activation layer, the 2_1th overlay layer, the 2_1th fully connected layer, the 2_1th spectral normalization layer, the 1_1th layer normalization layer, the 2_1th Leaky ReLU activation layer, and the 3rd parallel branch;

[0042] The 1_1 stacking layer splices together the features of the input discriminant network D, the category labels of the features, and the azimuth angle information of the features; the 2_1 stacking layer splices together the output of the upper network with the category labels and azimuth angle information of the features; the third parallel branch is composed of the 3_1 branch, the 3_2 branch, and the 3_3 branch in parallel; wherein the 3_1 branch is composed of the 3_1 fully connected layer, the 3_1 spectral normalization layer, and the 3_1 sigmoid activation layer stacked in sequence, which is used to determine the probability that the features of the input discriminant network are from the true feature distribution; the 3_2 branch is composed of the 3_2 fully connected layer, the 3_2 spectral normalization layer, and the 3_2 softmax activation layer stacked in sequence, which is used to predict the category labels of the features of the input discriminant network; the 3_3 branch is composed of the 3_3 fully connected layer and the 3_3 spectral normalization layer stacked in sequence, which is used to predict the azimuth angle of the features of the input discriminant network.

[0043] The parameters of each layer of the discriminant network are set as follows:

[0044] The number of output nodes of the 1_1 fully connected layer is set to 512; the number of output nodes of the 2_1 fully connected layer is set to 256; the number of output nodes of the 3_1 fully connected layer in the 3_1 branch is set to 1, the number of output nodes of the 3_2 fully connected layer in the 3_2 branch is set to 10, and the number of output nodes of the 3_3 fully connected layer in the 3_3 branch is set to 1.

[0045] The adaptive feature correction module AFCM is composed of the first parallel multiplication branch, the second multiplication layer, the second division layer, the second softmax layer and the second weighted sum layer stacked in sequence; wherein, the first parallel multiplication branch is composed of the third multiplication module, the third multiplication module and the third multiplication module in parallel;

[0046] In the first parallel multiplication branch, the 3_1 multiplication module is to combine the input AFCM features with the weight W K Perform inner product operation; the 3_2 multiplication module is to input the features of AFCM and the weight W Q Perform inner product operation; the 3_3 multiplication module is to input the features of AFCM and the weight W V Perform inner product operation; the 2_1st multiplication layer performs inner product operation on the output of the 3_2nd multiplication module and the transpose of the output of the 3_1st multiplication module; the 2_1st division layer divides the output of the 2_1st multiplication layer by the coefficient Where d represents the dimension d of the feature of size 1×d of the input AFCM; the 2_1th softmax layer is implemented by the softmax function; the 2_1th weighted sum layer is a weighted sum of the output of the 2_1th softmax layer and the output of the 3_3th multiplication module.

[0047] The parameters of each layer of the adaptive feature correction module AFCM are set as follows:

[0048] The weight W in the 3_1st multiplication module K The size is 1024*1024; the weight W in the 3_2 multiplication module Q The size is 1024*1024; the weight W in the 3_3 multiplication module V The size of is 1024*1024; the value of d in the 2_1th division layer is 1024.

[0049] The local feature classification module LC consists of the 4_1st parallel branch, the 4_1st Euclidean distance calculation layer, the 4_1st convolutional layer, the 4_1st sigmoid activation layer, the 4_1st reshape layer, the 4_1st fully connected layer, the 4_2nd fully connected layer and the 4_1st softmax layer, which are stacked in sequence;

[0050] The 4_1th parallel branch consists of an upper branch and a lower branch, wherein the upper branch consists of the 5_1th embedding layer, the 5_1th reshape layer, and the 5_1th local area feature partitioning layer stacked in sequence; the lower branch consists of the 6_1th mean calculation layer, the 6_1th reshape layer, and the 6_1th local area feature partitioning layer stacked in sequence;

[0051] The 5_1th embedding layer in the upper branch is to integrate the query features of the input local feature classification module LC with the weight W V Perform inner product operation; the 6_1th mean calculation layer in the lower branch takes the mean of the supporting features of the input local feature classification module LC by category; the role of the 5_1st reshape layer and the 6_1st reshape layer is to reshape the size of the input feature vector to restore the original spatial position relationship; the role of the 5_1st and 6_1st local area feature partitioning layers is to refine the input features into many local area features; the role of the 4_1st Euclidean distance calculation layer is to calculate the Euclidean distance between the local area features output by the upper branch and the lower branch in the 4_1st parallel branch.

[0052] The parameters of each layer of the local feature classification module LC are set as follows:

[0053] The weight of the 5_1 embedding layer in the upper branch of the 4_1 parallel branch is W of size 1024*1024 VThe convolution kernel size of the 4_1th convolutional layer is 7×7, the number of convolution kernels is 1, the convolution stride is 1, and the parameter "Padding" is set to "SAME". The number of output nodes of the 4_1th fully connected layer is 64; the number of output nodes of the 4_2th fully connected layer is 1.

[0054] Step 3, using the training set and the support set to form a total training set, and using the total training set to pre-train the multi-scale feature extraction module and the local feature classification module to obtain a pre-trained multi-scale local classification network MLC-Net;

[0055] In a specific embodiment, step 3 includes:

[0056] Step 3a: The support set D of the target domain s Each SAR image contained in is randomly rotated three times, and the rotated SAR image is compared with the support set D s All SAR images in are merged to obtain the expanded support set D sr ;

[0057] Step 3b: Expand the support set D sr and the training set D of the source domain train Merge to get a total training dataset D of 10 categories of targets with category labels and azimuths train_all ;

[0058] Step 3c, using the multi-scale feature extraction module and the local feature classification module as a multi-scale local classification network MLC-Net;

[0059] Step 3e, the total training data set D train_all K iterations of supervised training are performed as input to the multi-scale local classification network MLC-Net to obtain a pre-trained multi-scale local classification network MLC-Net, where K ≥ 2000.

[0060] Step 4: Select the support set and query set for each iteration from the total training set, and use the support set and query set to jointly train the pre-trained multi-scale local classification network MLC-Net and feature generation and correction network FGC-Net in each iteration to obtain the trained target classification model LcFGC;

[0061] In a specific embodiment, step 4 includes:

[0062] Step 4a, set the number of iterations to The maximum number of iterations is And order

[0063] Step 4b, randomly select the total training data set Dtrain_all The n categories in the SAR image are randomly selected and the S SAR images with category labels and azimuth angles of each category are randomly selected as the current The support set of the iteration Randomly select 15 SAR images from the remaining SAR images in each of the selected n categories as the current The query set of iterations

[0064] in, represents the support set of the b-th class, xs bh represents the h-th support sample of the b-th class, b and as bh Represents xs respectively bh Category label and orientation angle of xq i represents the i-th query sample, yq i and aq i Represent xq respectively i The category label and orientation, yq i ∈{1,...,n}; n=3, S≤20, as bh ∈[0°,360°), aq i ∈[0°,360°);

[0065] Step 4c: Set the support set The supporting image in is sent to the multi-scale feature extraction module FX to obtain the current The support feature set of the iteration The query set The query image in is sent to the multi-scale feature extraction module FX to obtain the current The query feature set of the iteration

[0066] Among them, fq i Represents the query sample xq i Query features; represents the support feature set of class b, fs bh Represents the support sample xs bh The real supporting features;

[0067] Step 4d, The support sample xs selected in the iteration bhThe category label, azimuth angle and random noise vector are sent to the generative network G, and T support features with azimuth angles uniformly distributed on [0°, 360°) with an interval of 5° are generated for each real support feature; the category labels and azimuth angles corresponding to the n×S×T generated support features and the n×S×T generated support features are input into the discriminant network D, and the probability value of the generated support features predicted by the discriminant network D is output as the real support feature, the probability value predicted as different categories and the predicted azimuth angle size, and the loss function L of the generative network is calculated. G ;

[0068] The hth true support feature (fs bh ,b,as bh ) as an example, the generated network G is f sbh The generated T-axis support features with azimuth angles uniformly distributed on [0°, 360°) with an interval of 5° are represented as Input the n×S×T generated support features, the category labels and azimuths corresponding to the n×S×T generated support features into the discriminant network D, output the probability value of the support features predicted by the discriminant network D as the real support features, the probability values predicted as different categories and the predicted azimuth angle, and calculate the loss function L of the generated network G The loss function L of the generated network G The calculation is as follows:

[0069]

[0070] Among them, P g (f) represents the distribution of the generated support features; t' represents the probability value of the discriminant network D predicting that the feature f is the true support feature; y' represents the probability value of the discriminant network D predicting that the feature f is the true category label; a' represents the azimuth angle of the feature f predicted by the discriminant network D, and a is the true azimuth angle of the feature f.

[0071] Step 4e, The n×S×T generated support features and their corresponding category labels and azimuth angles in the iteration are fed into the discriminant network D, and at the same time, the n×S real support features and their corresponding category labels and azimuth angles are fed into the discriminant network D. The discriminant network D evaluates the input real and generated support features and outputs the probability value predicted as the real support feature, the category prediction probability value and the azimuth prediction value respectively, and calculates the loss function L of the discriminant network D D ;

[0072] The loss function L of the discriminant network D The calculation is as follows:

[0073]

[0074] Among them, P r (f) represents the distribution of the true supporting features; t' represents the probability value of the discriminant network D predicting that the feature f is the true supporting feature; y' represents the probability value of the discriminant network D predicting that the feature f is the true category label; a' represents the azimuth angle of the feature f predicted by the discriminant network D, and a is the true azimuth angle of the feature f.

[0075] Step 4f, The n×S×T generated support features and their category labels in the iteration, and the n×S real support features and their category labels are sent to the adaptive feature correction module AFCM to obtain the corrected generated support features;

[0076] Generated support features is the real support feature of all the same class {fs bh ,h=1,...,S}, the support features generated by the correction are is calculated as:

[0077]

[0078]

[0079] K=FW K

[0080] V=FW V

[0081] Among them, W Q 、W K and W V Represents the weights of the adaptive feature correction module with a size of 1024*1024; F∈R (S+1)×1024 Indicates the support features that will be generated The matrix obtained by concatenating all real support features of the same type by row, The size of d is 1024; softmax represents the softmax function.

[0082] In order to ensure that the real support features and the corrected support features are in the same feature space, the weighted W V The embedded real support features participate in the classification, that is, the features from the second row to the S+1th row in the matrix V; the embedded real support features are represented as {fsc bh ,h=1,...,S}.

[0083] Step 4g, using The true support features {fsc bh,b=1,...,n,h=1,...,S} and the generated support features of n×S×T corrections Calculate the prototype mean square error loss and center constraint loss to ensure the consistency of the distribution of the corrected generated support features and the embedded real support features;

[0084] Prototype mean square error loss L PMSE and the center constraint loss L CC The calculation is as follows:

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] Among them, MSE represents the mean square error loss function, The prototype representing the true supporting feature of the embedding of class b, The prototype representing the corrected generated support features of class b; c bi represents the true support feature fsc of the i-th embedding of the b-th class bi And the center of the cluster of support features generated by T-correction with azimuth angles uniformly distributed from 0° to 360° and an interval of 5°; similarly, c bj represents the true support feature fsc of the j-th embedding of the b-th class bj and the centers of the clusters of T-corrected generated support features for which the azimuth angles are uniformly distributed from 0° to 360° with an interval of 5°.

[0092] Step 4h, The true support features {fsc bh ,b=1,...,n,h=1,...,S}, n×S×T support features generated by the correction And query features {(fq i ,yq i ,aq i ),i=1,...,n×15} Use the local feature classification network LC to predict the category label of each query sample; the process is: the query feature fq i ,i∈{1,...,n×15}with weight W VEmbed the embedded query feature fqc i , i∈{1,...,n×15} and the embedded real support features and the corrected generated support features are in the same feature space; n×S embedded real support features and n×S×T corrected generated support features are merged and averaged by category to obtain a hybrid prototype for each category

[0093] Embedded query feature fqc i ,i∈{1,...,n×15}, mixed prototype The calculation is as follows:

[0094]

[0095]

[0096] in, Represents a reshape operation, which reshapes the feature of size (1,1024) to size (4,4,64) to restore the original spatial relationship.

[0097] Step 4i, in the local feature classification network LC, the mixed prototype of each class is And each embedded query feature fqc i ,i∈{1,...,n×15} is split into 16 64-dimensional local region features, which are represented as and

[0098] in, Represents a mixed prototype The j-th local region feature, LRF i o Represents the embedded query feature fqc i The oth local region feature;

[0099] Step 4j, calculate the embedded query feature fqc i , all local region features of i∈{1,...,n×15} and the hybrid prototype The Euclidean distance between all local region features is used to obtain the embedded query feature fqc i With hybrid prototype The relationship matrix RM ib =(ro j ) 16×16 ;

[0100] Among them, r oj Represents the intersection element between row o and column j in the relationship matrix, which represents the embedded query feature fqc iThe oth local region feature LRF i o With hybrid prototype The jth local region feature The negative of the Euclidean distance between;

[0101] r oj The calculation is as follows:

[0102]

[0103] Step 4k, the relationship matrix RM ib =(r oj ) 16×16 ,i∈{1,...,n×15},b∈{1,...,n} are transformed to obtain the embedded query feature fqc i ,i∈{1,...,n×15} and mixed prototypes The similarity score s between ib , and then get the query sample xq i ,i∈{1,...,n×15} is predicted to be the probability p(b|xq i ),b∈{1,...,n}; using probability p(bxq i ),b∈{1,...,n} calculates the classification loss L of the local feature classification network LC C ;

[0104] Similarity score ib , probability p(b|xq i ) and classification loss L C The calculation is as follows:

[0105]

[0106]

[0107]

[0108] Among them, MLPs represents two layers of fully connected layers, where the output nodes of the first and second fully connected layers are 64 and 1 respectively; the first fully connected layer is followed by a ReLU activation function; Indicates reshaping the feature from size (16,16) to size (1,256); σ represents the sigmoid activation function; θ represents the convolution operation; Indicates element-wise multiplication; I indicates the indicator function. When querying sample xq i The true category label yq i When it is equal to b, I outputs 1, otherwise it outputs 0.

[0109] Step 4l, use the gradient back propagation algorithm to generate loss L G , classification loss L C , prototype mean square error loss L PMSE and the center constraint loss L CC The total generation loss is obtained by summing Calculate the gradient of the parameters of the generated network G and use the Adam optimizer with a learning rate of 0.0001 to update the parameters of the generated network G. The rest of the parameters remain unchanged.

[0110] Total generation loss The calculation is as follows:

[0111]

[0112] Among them, λ1, λ2 and λ3 represent hyperparameters, and their values are all 5.

[0113] Step 4m, repeat steps 4b to 41 three times;

[0114] Step 4n, repeat steps 4b to 4e;

[0115] Step 4o, using the gradient back propagation algorithm, using the discriminant loss L D To calculate the gradient of the parameters of the discriminant network D, the Adam optimizer with a learning rate of 0.00001 is used to update the parameters of the discriminant network D, and the rest of the parameters remain unchanged;

[0116] Step 4p, repeat steps 4b to 4k;

[0117] Step 4q, use the gradient back propagation algorithm and use the classification loss L C , prototype mean square error loss L PMSE and the center constraint loss L CC The total classification loss obtained by summing To update the gradients of the parameters of MLC-Net and the adaptive feature correction module AFCM, the Adam optimizer with a learning rate of 0.001 is used to update the above parameters. The learning rate is multiplied by 0.5 every 5000 iterations, and the rest of the parameters remain unchanged;

[0118] Total classification loss The calculation is as follows:

[0119]

[0120] Among them, β1 and β2 are hyperparameters whose values are both 1.

[0121] Step 4r, judge Is it true? If it is true, end the iteration and get the trained target classification model LcFGC; otherwise, let Repeat steps 4b to 4r.

[0122] Before step 5, the support set D of the target domain can be s , test set D test Send it into the trained target classification model for classification to obtain the test set D test The target recognition accuracy of each SAR image is used to evaluate the quality of the target classification model. This allows the trained model to be adaptively adjusted based on the evaluation results, providing a reference for subsequent classification of images collected on site.

[0123] Step 5: Input the real-time collected SAR image into the trained target classification model to obtain the category of the target contained in the real-time collected SAR image.

[0124] The present invention can display the location and category of the target by marking the target in the image.

[0125] The effect of the present invention can be further illustrated by the following experimental data:

[0126] 1. Experimental conditions:

[0127] 1) Experimental data:

[0128] The experiment uses the publicly available MSTAR dataset of moving and stationary targets. The MSTAR dataset contains 10 types of ground vehicle targets with an image resolution of 0.3m × 0.3m. The target domain dataset used in this experiment includes three types of targets at 17° and 15° elevation: the BMP2, BTR70, and T72. The source domain dataset contains the remaining seven types of targets at a 17° elevation: the BTR60, 2S1, BRDM2, D7, T62, ZIL131, and ZSU23 / 4.

[0129] The MSTAR dataset used in the experiment consists of 2049 SAR images in the source domain, 698 SAR images at a 17° elevation angle in the target domain, and 1365 SAR images at a 15° elevation angle in the target domain. All SAR images in both the source and target domains were originally 128×128 pixels and were uniformly cropped to 64×64 pixels during preprocessing. All SAR images used in the experiment were L2-normalized.

[0130] The support set D of the target domain s It is obtained by randomly selecting S SAR images from each of the three categories at a pitch angle of 17° in the target domain. s Each SAR image is rotated three times at random angles and then compared with D s Merge to get the rotated support set Dsr 2049 SAR images at a 17° elevation angle in the source domain are used as the training set D train , 1365 SAR images at a 15° pitch angle in the target domain are used as the test set D test .

[0131] In the experiment, the training set D of the source domain is train , the support set D of the rotated target domain sr As the input of the joint local classification and feature generation and correction network LcFGC and train LcFGC, the trained LcFGC is obtained. Then the target domain support set D s and the test set D of the target domain test As the input of the trained LcFGC, the test set D of the target domain is output test The target recognition accuracy of each SAR image is compared with the results of other classification models.

[0132] 2) Experimental content: Comparative experiment on the above experimental data using the method of the present invention and the existing method

[0133] Existing methods include: SAR target recognition method based on convolutional neural network (A-ConvNet, from the paper "Target classification using the deep convolutional networks for SAR images", IEEE Transactions on Geoscience and Remote Sensing, Chen et al., 2016), SAR target recognition method based on data augmentation (DA-Net, from the paper "Convolutional neural network with data augmentation for SAR target recognition", IEEE Geoscience and Remote Sensing Letters, Ding et al., 2016), SAR target recognition method based on angular rotation generative network (ARGN, from the paper "SAR target recognition with limited training data based on angular rotation generative network", IEEE Geoscience and Remote Sensing Letters, Sun et al., 2020), SAR target recognition network based on few-shot learning (MSAR, from the paper "Few-shot SAR target classification via metalearning", IEEE Transactions on Geoscience and Remote Sensing, Fu et al., 2021), and prototype network based on few-shot learning (PN, from the paper "Prototypical networks for few-shot learning",Neural Information Processing Systems, Snell et al., 2017), Relational Network for Few-Shot Learning (RN, from the paper “Learning to compare: Relation network for few-shot learning”, IEEE Conference on Computer Vision and Pattern Recognition, Sung et al., 2018), Meta-Baseline for Few-Shot Learning (Meta-Baseline, from the paper “A new meta-baseline for few-shot learning”, arXiv preprint arXiv:2003.04390, Chen et al., 2020), and Attribute-Guided Multi-Scale Prototypical Network (AG-MsPN, from the paper “Attribute-guided multi-scale prototypical network for few-shot SAR target classification”, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Wang et al., 2021).

[0134] In order to verify the recognition effect of the joint local classification and feature generation and correction network LcFGC in the case of limited samples, the number S of SAR images of each category in the support set of the target domain is selected and compared with other methods. The selection ratio and comparison results are shown in Table 1.

[0135] Table 1 Performance comparison results of the method of the present invention and the existing method

[0136]

[0137] For example, if the number S is 5, in the target domain dataset, 5 SAR images are randomly selected from each category at a pitch angle of 17° as the support set D of the target domain. s , and all SAR images at a pitch angle of 15° are selected as the target domain test set D test In order to eliminate the randomness of the experimental results, each set of experiments was performed 20 times, and each time a new support set was randomly selected from the three types of data at a pitch angle of 17° in the target domain, and the average of all results was taken as the final recognition result in Table 1.

[0138] As can be seen from Table 1, when each class has only 5 samples, 10 samples and 20 samples, the recognition rate of the experiment of the present invention is better than the results of other comparison methods, and has a better target recognition rate in the case of few samples.

[0139] Although the present application is described herein with reference to various embodiments, those skilled in the art will be able to understand and implement other variations of the disclosed embodiments in practicing the claimed application by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0140] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A SAR target classification method combining local classification with feature generation and correction, characterized in that: include: Step 1: Obtain SAR images collected in real time from the field, as well as the training set of the source domain and the support set of the target domain; Step 2: constructing a target classification model LcFGC that combines local classification, feature generation, and correction networks; wherein the target classification model is composed of a multi-scale feature extraction module, a generation network, a discriminant network, an adaptive feature correction module, and a local feature classification module stacked in sequence; The multi-scale feature extraction module and the local feature classification module constitute a multi-scale local classification network MLC-Net; the generation network, the discrimination network and the adaptive feature correction module constitute a feature generation and correction network FGC-Net; Step 3, using the training set and the support set to form a total training set, and using the total training set to pre-train the multi-scale feature extraction module and the local feature classification module to obtain a pre-trained multi-scale local classification network MLC-Net; Step 4: Select the support set and query set for each iteration from the total training set, and use the support set and query set to jointly train the pre-trained multi-scale local classification network MLC-Net and feature generation and correction network FGC-Net in each iteration to obtain the trained target classification model LcFGC; Step 5: Input the real-time collected SAR image into the trained target classification model to obtain the category of the target contained in the real-time collected SAR image.

2. The SAR target classification method combining local classification with feature generation and correction according to claim 1, characterized in that: Step 1 includes: Step 1a: Obtain a moving and stationary target MSTAR dataset of 10 types of targets with category labels and azimuths, as well as SAR images collected in real time from the field; Step 1b, using the SAR image collected in real time on site as the SAR image to be classified; Step 1c: crop each SAR image in the MSTAR dataset D and the SAR image to be classified into a size of 64×64 pixels; Step 1d: perform L2 normalization on each SAR image in the cropped MSTAR dataset D and the SAR image to be classified, so as to obtain the preprocessed moving and stationary target MSTAR dataset D. p and the processed SAR image to be classified; Step 1e, from the preprocessed MSTAR dataset D p Randomly select n types of SAR images with category labels and azimuth angles at a pitch angle of 15° as the test set D of the target domain test ; D p The remaining (10-n) types of SAR images with 17° elevation angles with category labels and azimuth angles are used as the training set D in the source domain. train ; Randomly select S SAR images for each category from the selected n categories of SAR images with category labels and azimuth angles at a pitch angle of 17° as the support set D of the target domain s ; Among them, n=3, S≤20.

3. The SAR target classification method combining local classification with feature generation and correction according to claim 2, characterized in that: In step 1d, L2 normalization is performed, and the implementation process is as follows: Among them, ij represents the pixel value of the SAR image at the position of row i and column j, Represents the pixel value at the i-th row and j-th column of each SAR image after L2 normalization.

4. The SAR target classification method combining local classification with feature generation and correction according to claim 2, characterized in that: Before step 5, the SAR target classification method combining local classification with feature generation and correction further includes: The support set D of the target domain s , test set D test Send it into the trained target classification model for classification to obtain the test set D test The target recognition accuracy of each SAR image is used to evaluate the quality of the target classification model.

5. The SAR target classification method combining local classification with feature generation and correction according to claim 1, characterized in that: The multi-scale feature extraction module consists of four identical convolution modules Conv, three identical downsampling modules DS and a transformation module Tran; the four identical convolution modules Conv are respectively the 0_1 convolution module, the 0_2 convolution module, the 0_3 convolution module, and the 0_4 convolution module; The three identical downsampling modules are the 0_1 downsampling module, the 0_2 downsampling module, and the 0_3 downsampling module; The input of the 0_1th sampling module is the output of the 0_1th convolution module. The input of the remaining 0_1th downsampling modules is the output of the 0_(l-1)th downsampling module and the output of the 0_1th convolution module, where l = 2, 3. The output of the 0_3th downsampling module is spliced with the output of the 0_4th convolution module according to the channel and input to the transformation module Tran. Among them, the four identical convolution modules Conv include the 0_1th convolution layer, the 0_1th batch normalization layer, the 0_1th ReLU activation layer and the 0_1th maximum pooling layer stacked in sequence; the parameters of the convolution module Conv are specifically set as follows: the convolution kernel size of the 0_1th convolution layer is 3×3, and the convolution step is 1; the pooling area size of the 0_1th maximum pooling layer is set to 2×2, and the pooling step is 2; the downsampling module DS has the same structure as the convolution module Conv; The transformation module Tran consists of the 0_2th batch normalization layer, the channel attention module CA, the convolution module Conv0 and the reshape layer stacked in sequence; The channel attention module CA is composed of a 0_1th parallel layer, a c_1th fully connected layer, a c_2th fully connected layer, a 0_1th addition module, a 0_1th sigmoid activation layer and a 0_1th multiplication module stacked in sequence; wherein the 0_1th parallel layer is composed of a parallel average pooling layer and a 0_2th maximum pooling layer; the average pooling layer and the 0_2th maximum pooling layer of the 0_1th parallel layer take the average and maximum values of the features of each channel of the multi-channel features of the input CA to obtain the average pooling features and the maximum pooling features; the c_1th fully connected layer and the c_2th fully connected layer transform the average pooling features and the maximum pooling features output by the 0_1th parallel layer; the 0_1th addition module adds the transformed average pooling features and the transformed maximum pooling features output by the 0_2th fully connected layer element by element and outputs the channel attention features; the 0_1th multiplication module multiplies the channel attention features activated by the 0_1th sigmoid activation layer by the features of the original input CA element by element; The convolution module Conv0 consists of the 1_1th convolution layer, the 1_1th batch normalization layer and the 1_1th ReLU activation layer stacked in sequence; The parameters of each layer in the channel attention module CA are specifically set as follows: the pooling area size of the average pooling layer of the 0_1th parallel layer is set to 4×4, and the pooling step is 1; the pooling area size of the maximum pooling layer of the 0_1th parallel layer is set to 4×4, and the pooling step is 1; the number of nodes in the output of the 0_1th fully connected layer and the 0_2th fully connected layer are 64 and 256 respectively; The parameter settings of the second convolution module Conv0 are the same as those of the first convolution module Conv.

6. The SAR target classification method combining local classification with feature generation and correction according to claim 1, characterized in that: The generating network G is composed of the first stacking layer, the first fully connected layer, the first spectral normalization layer, the first ReLU activation layer, the second stacking layer, the second fully connected layer, the second spectral normalization layer, the first layer normalization layer, the second ReLU activation layer, the third stacking layer, the third fully connected layer, the third spectral normalization layer and the third ReLU activation layer. The first overlay layer is used to splice together the random vector of the input generation network G, the category label of the feature, and the azimuth information of the feature; the second and third overlay layers are both used to splice together the output of the upper network with the category label of the feature and the azimuth information of the feature; The discriminant network D is composed of the 1_1th overlay layer, the 1_1th fully connected layer, the 1_1th spectral normalization layer, the 1_1th Leaky ReLU activation layer, the 2_1th overlay layer, the 2_1th fully connected layer, the 2_1th spectral normalization layer, the 1_1th layer normalization layer, the 2_1th Leaky ReLU activation layer, and the 3rd parallel branch. The 1_1 stacking layer splices together the features of the input discriminant network D, the category labels of the features, and the azimuth angle information of the features; the 2_1 stacking layer splices together the output of the upper network with the category labels and azimuth angle information of the features; the third parallel branch is composed of the 3_1 branch, the 3_2 branch, and the 3_3 branch in parallel; wherein the 3_1 branch is composed of the 3_1 fully connected layer, the 3_1 spectral normalization layer, and the 3_1 sigmoid activation layer stacked in sequence, which is used to determine the probability that the features of the input discriminant network are from the true feature distribution; the 3_2 branch is composed of the 3_2 fully connected layer, the 3_2 spectral normalization layer, and the 3_2 softmax activation layer stacked in sequence, which is used to predict the category labels of the features of the input discriminant network; the 3_3 branch is composed of the 3_3 fully connected layer and the 3_3 spectral normalization layer stacked in sequence, which is used to predict the azimuth angle of the features of the input discriminant network.

7. The SAR target classification method combining local classification with feature generation and correction according to claim 1, It is characterized by: The adaptive feature correction module AFCM is composed of the first parallel multiplication branch, the second multiplication layer, the second division layer, the second softmax layer and the second weighted sum layer stacked in sequence; wherein, the first parallel multiplication branch is composed of the third multiplication module, the third multiplication module and the third multiplication module in parallel; In the first parallel multiplication branch, the 3_1 multiplication module is to combine the input AFCM features with the weight W K Perform inner product operation; the 3_2 multiplication module is to input the features of AFCM and the weight W Q Perform inner product operation; the 3_3 multiplication module is to input the features of AFCM and the weight W V Perform inner product operation; the 2_1st multiplication layer performs inner product operation on the output of the 3_2nd multiplication module and the transpose of the output of the 3_1st multiplication module; the 2_1st division layer divides the output of the 2_1st multiplication layer by the coefficient Where d represents the dimension d of the feature of size 1×d of the input AFCM; the 2_1th softmax layer is implemented by the softmax function; the 2_1th weighted sum layer is a weighted sum of the output of the 2_1th softmax layer and the output of the 3_3th multiplication module.

8. The SAR target classification method combining local classification with feature generation and correction according to claim 1, characterized in that: The local feature classification module LC consists of the 4_1st parallel branch, the 4_1st Euclidean distance calculation layer, the 4_1st convolutional layer, the 4_1st sigmoid activation layer, the 4_1st reshape layer, the 4_1st fully connected layer, the 4_2nd fully connected layer and the 4_1st softmax layer, which are stacked in sequence; The 4_1th parallel branch consists of an upper branch and a lower branch, wherein the upper branch consists of the 5_1th embedding layer, the 5_1th reshape layer, and the 5_1th local area feature partitioning layer stacked in sequence; the lower branch consists of the 6_1th mean calculation layer, the 6_1th reshape layer, and the 6_1th local area feature partitioning layer stacked in sequence; The 5_1th embedding layer in the upper branch is to integrate the query features of the input local feature classification module LC with the weight W V Perform inner product operation; the 6_1th mean calculation layer in the lower branch takes the mean of the supporting features of the input local feature classification module LC by category; the role of the 5_1st reshape layer and the 6_1st reshape layer is to reshape the size of the input feature vector to restore the original spatial position relationship; the role of the 5_1st and 6_1st local area feature partitioning layers is to refine the input features into many local area features; the role of the 4_1st Euclidean distance calculation layer is to calculate the Euclidean distance between the local area features output by the upper branch and the lower branch in the 4_1st parallel branch.

9. The SAR target classification method combining local classification with feature generation and correction according to claim 4, characterized in that: Step 3 includes: Step 3a: The support set D of the target domain s Each SAR image contained in is randomly rotated three times, and the rotated SAR image is compared with the support set D s All SAR images in are merged to obtain the expanded support set D sr ; Step 3b: Expand the support set D sr and the training set D of the source domain train Merge to get a total training dataset D of 10 categories of targets with category labels and azimuths train_all ; Step 3c, using the multi-scale feature extraction module and the local feature classification module as a multi-scale local classification network MLC-Net; Step 3e, the total training data set D train_all As the input of the multi-scale local classification network MLC-Net, b iterative supervised training is performed to obtain the pre-trained multi-scale local classification network MLC-Net, where K ≥ 2000.

10. The SAR target classification method combining local classification with feature generation and correction according to claim 9, characterized in that: Step 4 includes: Step 4a, set the number of iterations to The maximum number of iterations is And order Step 4b, randomly select the total training data set D train_all The n categories in the SAR image are randomly selected and the S SAR images with category labels and azimuth angles of each category are randomly selected as the current The support set of the iteration Randomly select 15 SAR images from the remaining SAR images in each of the selected n categories as the current The query set of iterations in, represents the support set of the b-th class, xs bh represents the h-th support sample of the b-th class, b and as bh Represents xs respectively bh Category label and orientation angle of xq i represents the i-th query sample, yq i and aq i Represent xq respectively i The category label and orientation, yq i ∈{1,...,n}; n=3, S≤20, as bh ∈[0°,360°), aq i ∈[0°,360°); Step 4c: Set the support set The supporting image in is sent to the multi-scale feature extraction module FX to obtain the current The support feature set of the iteration The query set The query image in is sent to the multi-scale feature extraction module FX to obtain the current The query feature set of the iteration Among them, fq i Represents the query sample xq i Query features; in, represents the support feature set of class b, fs bh Represents the support sample xs bh The real supporting features; Step 4d, The support sample xs selected in the iteration bh The category label, azimuth angle and random noise vector are sent to the generative network G, and T support features with azimuth angles uniformly distributed on [0°, 360°) with an interval of 5° are generated for each real support feature; the category labels and azimuth angles corresponding to the n×S×T generated support features and the n×S×T generated support features are input into the discriminant network D, and the probability value of the generated support features predicted by the discriminant network D is output as the real support feature, the probability value predicted as different categories and the predicted azimuth angle size, and the loss function L of the generative network is calculated. G ; Step 4e, The n×S×T generated support features and their corresponding category labels and azimuth angles in the iteration are fed into the discriminant network D, and at the same time, the n×S real support features and their corresponding category labels and azimuth angles are fed into the discriminant network D. The discriminant network D evaluates the input real and generated support features and outputs the probability value predicted as the real support feature, the category prediction probability value and the azimuth prediction value respectively, and calculates the loss function L of the discriminant network D D ; Step 4f, The n×S×T generated support features and their category labels in the iteration, and the n×S real support features and their category labels are sent to the adaptive feature correction module AFCM to obtain the corrected generated support features; Step 4g, using The true support features {fsc bh ,b=1,...,n,h=1,...,S} and the generated support features of n×S×T corrections Calculate the prototype mean square error loss and center constraint loss to ensure the consistency of the distribution of the corrected generated support features and the embedded real support features; Step 4h, The true support features {fsc bh ,b=1,...,n,h=1,...,S}, n×S×T support features generated by the correction And query features {(fq i ,yq i ,aq i ),i=1,...,n×15} Use the local feature classification network LC to predict the category label of each query sample; the process is: the query feature fq i ,i∈{1,...,n×15}with weight W V Embed the embedded query feature fqc i , i∈{1,...,n×15} and the embedded real support features and the corrected generated support features are in the same feature space; n×S embedded real support features and n×S×T corrected generated support features are merged and averaged by category to obtain a hybrid prototype for each category Step 4i, in the local feature classification network LC, the mixed prototype of each class is And each embedded query feature fqc i ,i∈{1,...,n×15} is split into 16 64-dimensional local region features, which are represented as for peace in, Represents a mixed prototype The j-th local region feature, LRF i o Represents the embedded query feature fqc i The oth local region feature; Step 4j, calculate the embedded query feature fqc i , all local region features of i∈{1,...,n×15} and the hybrid prototype The Euclidean distance between all local region features is used to obtain the embedded query feature fqc i With hybrid prototype The relationship matrix RM ib =(ro j ) 16×16 ; Among them, r oj Represents the intersection element between row o and column j in the relationship matrix, which represents the embedded query feature fqc i The oth local region feature LRF i o With hybrid prototype The jth local region feature The negative of the Euclidean distance between; Step 4k, the relationship matrix RM ib =(r oj ) 16×16 ,i∈{1,...,n×15},b∈{1,...,n} are filtered and transformed to obtain the embedded query feature fqc i ,i∈{1,...,n×15} and mixed prototypes The similarity score s between ib , and then get the query sample xq i ,i∈{1,...,n×15} is predicted to be the probability p(b|xq i ),b∈{1,...,n}; using the probability p(b|xq i ),b∈{1,...,n} calculates the classification loss L of the local feature classification network LC C ; Step 4l, use the gradient back propagation algorithm to generate loss L G , classification loss L C , prototype mean square error loss L PMSE and the center constraint loss L CC The total generation loss is obtained by summing Calculate the gradient of the parameters of the generated network G and use the Adam optimizer with a learning rate of 0.0001 to update the parameters of the generated network G. The rest of the parameters remain unchanged. Step 4m, repeat steps 4b to 41 three times; Step 4n, repeat steps 4b to 4e; Step 4o, using the gradient back propagation algorithm, using the discriminant loss L D To calculate the gradient of the parameters of the discriminant network D, the Adam optimizer with a learning rate of 0.00001 is used to update the parameters of the discriminant network D, and the rest of the parameters remain unchanged; Step 4p, repeat steps 4b to 4k; Step 4q, use the gradient back propagation algorithm and use the classification loss L C , prototype mean square error loss L PMSE and the center constraint loss L CC The total classification loss obtained by summing To update the gradients of the parameters of MLC-Net and the adaptive feature correction module AFCM, the Adam optimizer with a learning rate of 0.001 is used to update the above parameters. The learning rate is multiplied by 0.5 every 5000 iterations, and the rest of the parameters remain unchanged; Step 4r, judge Is it true? If it is true, end the iteration and get the trained target classification model LcFGC; otherwise, let Repeat steps 4b to 4r.

Images