Incomplete data ship identification device based on generative adversarial network and depth measurement

Abstract

The invention discloses an incomplete data ship identification device based on a generative adversarial network and depth measurement, which is used for identifying and classifying an incomplete ship target in a synthetic aperture radar image. The defects that an existing ship target recognition model is not good enough in recognition effect, difficult in feature selection and not suitable for incomplete data conditions are overcome. The preprocessing module realizes supplementation of incomplete data by introducing an adversarial generative network, the feature extraction and selection module selects optimal features by using a fusion evaluation strategy, and the classification and identification module designs a weighted triple loss function based on a depth measurement method, so that the classification and identification accuracy is improved. The introduction of the adaptive weight enables the model to pay more attention to the important triple in the training process, thereby obtaining a ship target recognition device under the condition of incomplete data. The non-complete ship target recognition device in the synthetic aperture radar image is good in recognition effect, feature selection is easy, and the target recognition training process is more stable.

Description

Technical Field

[0001] The present invention relates to the field of ship target recognition, and particularly to a ship recognition device based on a generative adversarial network and deep metric for incomplete data. Background Art

[0002] Ship target recognition refers to accurately identifying and locating real ship targets at sea by analyzing marine image data using various technologies and algorithms in the actual marine environment, so as to safeguard marine rights and interests and ensure maritime navigation safety. In recent years, synthetic aperture radar images have become one of the important methods for marine ship target monitoring and recognition due to their imaging characteristics such as all-weather, all-day, and unaffected by weather. Currently, many methods for monitoring and recognizing ship targets have been proposed, such as ship target detection based on background clutter statistical distribution, ship target detection based on polarization decomposition, ship target detection based on polarization characteristics, etc. Recently, deep learning technology has also been applied to real ship target detection. For example, the improved YOLOv5 algorithm combines the characteristics of synthetic aperture radar images and extracts feature information through parallel multi-level pyramid modules, thereby improving the detection accuracy of ship targets.

[0003] However, the recognition effect of the models proposed by the above algorithms still has deficiencies. The feature selection work is relatively complicated and is greatly affected by individual differences of human factors. In addition, in actual data, there are some samples that are difficult to distinguish (such as the features between positive and negative samples are very close). These key difficult samples are particularly important for improving the model performance, but traditional methods cannot dynamically adjust the training process, resulting in insufficient attention of the model to these samples.

[0004] In addition, in the ship target detection task, there are often situations where the dataset is incomplete or missing due to various reasons, which increases the difficulty of establishing a ship target recognition model. For example, since ship targets are relatively complex in the high-frequency ground wave radar environment, labeling all target points may consume a large amount of time and labor costs, resulting in a lack of sufficient labeled data in the dataset; in the high-frequency ground wave radar environment, the energy of target points may be weak, making it difficult to accurately detect target points in radar images. In this case, there may be problems of missed detection or false detection of target points in the dataset, resulting in insufficient completeness of the dataset; in the ship target detection task, complex detection environments may be encountered, such as sea waves, clouds, etc. These environmental factors may affect the quality of radar images and the visibility of targets, resulting in a certain degree of incompleteness in the dataset.

[0005] In view of the deficiencies of the current ship target recognition models, the research on ship target recognition devices in the case of incomplete data has become a frontier and hot topic in the academic and industrial fields. Summary of the Invention

[0006] The object of the present invention is to provide a ship recognition device based on a generative adversarial network and deep metric for incomplete data, which has good recognition effect and is easy to perform feature selection, aiming at the deficiencies of the existing ship target recognition models, such as poor recognition effect, difficulty in feature selection, and inability to be applied to the case of incomplete data.

[0007] To achieve the above object, the technical solution adopted by the present invention is: a ship recognition device based on a generative adversarial network and deep metric for incomplete data, which is used to identify and classify incomplete ship targets in synthetic aperture radar images, and includes a preprocessing module, a feature extraction and selection module, and a classification and recognition module; the preprocessing module is used to fill in the data of the slices of incomplete ship targets in the synthetic aperture radar images based on the generative adversarial network; the feature extraction and selection module is used to extract and select features from the data output by the preprocessing module, and fuse and evaluate the features based on the separability criterion and the stability criterion to select the optimal features; the classification and recognition module is used to train a deep convolutional neural network model based on the data set processed by the feature extraction and selection module, perform classification and recognition, and obtain the ship target recognition result.

[0008] Further, the input of the preprocessing module is the slices of incomplete ship targets in the synthetic aperture radar images, and data filling is performed based on the generative adversarial network. The specific steps are as follows:

[0009] (1) For the incomplete ship target slice data matrix X = (X 1 ,..., X d ), let the mask matrix M = (M 1 ,..., M d ) with elements 0 or 1, where if the data of a certain element in matrix X is not clear, the corresponding element value of M is 0, otherwise it is 1; define a random matrix Z = (Z 1 ,..., Z d ), which is independent of all other variables;

[0010] (2) The generator is used to observe the real data, supplement the data, and output a complete data matrix;

[0011] (3) The discriminator distinguishes the vectors from the real data and the supplemented vectors, and outputs an estimated matrix of the mask matrix M;

[0012] (4) Train the generator and the discriminator multiple times to output the best supplemented data matrix. Among them, the training process is as follows:

[0013] (4.1) The generator G takes the data matrix X, the mask matrix M, and the random matrix Z as inputs, samples t D samples from the data matrix X and the mask matrix M samples t from the random matrix ZD t independent and identically distributed samples Sample t from the random variable B D t independent and identically distributed samples where the random variable B = (B 1 ,..., B d ) has elements that are either 0 or 1. First, randomly and uniformly sample k from {1,..., d}, and then let

[0014] (4.2) For i = 1,..., t D , the generator G generates the elements y i of the augmented data matrix:

[0015] y i = m i × x i + (1 - m i ) × G(X i , m i , z i )(12);

[0016] Then, update the elements h i of the hint matrix;

[0017] h i = b i × m i + 0.5(1 - b i )(13);

[0018] where × denotes element-wise multiplication, 1 is a vector of 1s, and b i is an indicator scalar. When its value is 0, it means the data point is augmented data; when its value is 1, it means non-augmented data;

[0019] (4.3) The discriminator D takes the augmented data matrix Y = (Y 1 ,..., Y d ) and the hint matrix H = (H 1 ,..., H d ) as inputs and updates the discriminator D using stochastic gradient descent. The gradient is:

[0020]

[0021] where L D () represents the loss function of the discriminator D. Its goal is to distinguish the difference between real data and supplementary data by minimizing the loss. The loss formula is essentially a variant of the cross-entropy loss function and is defined as follows:

[0022]

[0023] where, m'i = D(y i , h i ) is the estimated value in the mask matrix M predicted by the model, corresponding to the predicted value of the true mask matrix m i . And L D () is used to quantify the error between the model prediction m'i and the true label m i ;

[0024] (4.4) Sample t G samples from the data matrix X and the mask matrix M Sample t G independent and identically distributed samples from the random matrix Z Sample t G independent and identically distributed samples from the random variable B For i = 1,..., t G , update the elements of the hint matrix:

[0025] h i = b i × m i + 0.5(1 - b i )(16);

[0026] (4.5) Update the generator G using stochastic gradient descent, and the gradient is:

[0027]

[0028] where L G () is the adversarial loss of the generator, which encourages the distribution of the mask matrix m'i predicted by the model to be similar to that of the true mask matrix m i . L M () is the reconstruction loss function, which ensures that the data matrix x'i generated by the generator can be as close as possible to the true data x i in characteristics. a is a hyperparameter for balancing the two losses. The specific definitions of the two loss functions are:

[0029]

[0030] The meaning of the above formula is to generate the mask matrix estimate m'i according to the samples where b i = 0, so that it can match the true mask m i in the augmented data part. By optimizing this loss, the generator learns how to generate a mask matrix that can reconstruct the true data distribution in the part where data needs to be augmented.

[0031] Next, the definition of the reconstruction loss function L M () is given:

[0032]

[0033] where l M () is the unidirectional loss, which measures the difference between the generated data x′i and the real data x i and is defined as follows:

[0034]

[0035] If x i is a continuous variable, the mean square error is used to evaluate the generation result; otherwise, the cross entropy is used to calculate the probability distribution of the generation result. By minimizing this loss function, the generator makes the completed data matrix x′ i consistent with the real data x i in different characteristics.

[0036] Furthermore, the feature extraction and selection module includes a feature extraction module, a fusion evaluation module, a correlation coefficient screening module, and a K-nearest neighbor screening module, which are used to extract and select features from the preprocessed data, fuse and evaluate the features based on the separability criterion and the stability criterion, and select the optimal features. The specific implementation steps are as follows:

[0037] (1) Extract an initial feature set through the feature extraction module, including the basic features and electromagnetic scattering features in the research on ship target classification and recognition of high-resolution synthetic aperture radar images, where the basic features include geometric structure features and gray-scale statistical features;

[0038] (2) In the fusion evaluation module, fuse and evaluate the initial feature set, and perform mean fusion on the evaluation results obtained for each feature according to the separability criterion and the stability criterion; according to the fused evaluation values, eliminate the features with evaluation values less than 0.25 to obtain the remaining feature set;

[0039] (3) In the correlation coefficient screening module, calculate the correlation coefficient r ij between any two features:

[0040]

[0041] where s i , s j represent the means of features p i and p j , and σ i , σ j represent the standard deviations of features p i and p j ; use the correlation coefficient between features to measure the correlation between features and screen out redundant features;

[0042] (4) In the K-nearest neighbor screening module, for the candidate feature subsets, use the K-nearest neighbor classifier to classify the candidate feature subsets, take the classification accuracy rate as the evaluation index for feature screening, select the feature subset with the highest accuracy rate, and obtain the optimal feature subset.

[0043] Furthermore, the classification and recognition module includes a batch triplet sampling module and a convolutional neural network module. The specific implementation steps are as follows:

[0044] (1) In the batch triplet sampling module, perform batch sampling on the dataset after feature extraction and selection. Select three samples as a batch of data triplets for each training. The first sample x a is used as the anchor sample, the second positive sample x p has the same label as the anchor sample, and the third negative sample x n has a different label from the anchor sample;

[0045] (2) In the convolutional neural network module, input the batch of data triplets into the convolutional neural network F and train it using the adaptive triplet loss function. After introducing the adaptive weights, the model can dynamically assign higher weights to important triplets, making the training more efficient and accurate, and at the same time avoiding excessive attention to invalid or extreme samples. The loss function is as follows:

[0046]

[0047] where w i is the weight, and different adaptive weights are assigned according to different classes; b is a constant, called the margin;

[0048] (3) In the testing stage, use the trained convolutional neural network F to classify and recognize ship targets for the dataset after feature extraction and selection, and obtain the ship target recognition results.

[0049] The beneficial effects of the present invention are mainly manifested as follows: The present invention identifies and classifies incomplete ship targets in synthetic aperture radar images, overcoming the deficiencies of existing ship target recognition models, such as poor recognition effects, difficulty in feature selection, and inability to apply to incomplete data situations. The preprocessing module realizes the supplementation of incomplete data by introducing a generative adversarial network. The feature extraction and selection module uses a fusion evaluation strategy to select the optimal features. The classification and recognition module designs a weighted triplet loss function based on a deep metric method. The introduction of adaptive weights enables the model to pay more attention to important triplets during the training process, thereby obtaining a ship target recognition device under incomplete data conditions. After introducing adaptive weights, the model can dynamically assign higher weights to important triplets, making the training more efficient and accurate, while avoiding excessive attention to invalid or extreme samples and solving the drawbacks of traditional methods. The present invention has a good recognition effect for the incomplete ship target recognition device in synthetic aperture radar images, is easy to perform feature selection, and the target recognition training process is more stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a schematic diagram of the basic structure of a ship target recognition device under incomplete data conditions;

[0051] Figure 2 is a schematic diagram of the basic structure of a data preprocessing module based on a generative adversarial network;

[0052] Figure 3 is a schematic diagram of the basic structure of a feature extraction and selection module based on a fusion evaluation strategy;

[0053] Figure 4 is a schematic diagram of the basic structure of a classification and recognition module based on a deep metric method. DETAILED DESCRIPTION OF THE INVENTION

[0054] The present invention will be specifically described below with reference to the accompanying drawings.

[0055] Refer to Figure 1 , a ship recognition device for incomplete data based on a generative adversarial network and a deep metric, includes a ship target slice 1 of incomplete data, a preprocessing module 2, a generative adversarial network module 3, a feature extraction and selection module 4, a fusion evaluation strategy module 5, a classification and recognition module 6, and a deep metric learning module 7. Connect the ship target slice 1 of the incomplete data to the preprocessing module 2, connect the generative adversarial network module 3 to the preprocessing module 2, connect the output end of the preprocessing module 2 to the input end of the feature extraction and selection module 4, connect the fusion evaluation strategy module 5 to the feature extraction and selection module 4, connect the output end of the feature extraction and selection module 4 to the input end of the classification and recognition module 6, and connect the deep metric learning module 7 to the classification and recognition module 6.

[0056] Based on Figure 2 、 Figure 3 and Figure 4 , the incomplete data ship recognition device based on generative adversarial network and deep metric includes:

[0057] (1) Preprocessing module 2, which is used to preprocess the incomplete data ship target slices 1 in the synthetic aperture radar image and fill in the data based on the generative adversarial network module 3;

[0058] (1.1) For the incomplete ship target slice data matrix X = (X 1 ,..., X d ), let the mask matrix M = (M 1 ,..., M d ) with elements 0 or 1, where if the data of an element in matrix X is not clear, the corresponding element value of M is 0, otherwise it is 1. Define a random matrix Z = (Z 1 ,..., Z d ), which is independent of all other variables;

[0059] (1.2) The generator module 8 is used to observe the real data, supplement the data, and output the complete data matrix;

[0060] (1.3) The discriminator module 9 is used to judge which is the vector from the real data and which is the supplemented vector, and output the probability matrix that the matrix generated by the generator module 8 comes from the real data, that is, output an estimated matrix of the mask matrix M;

[0061] (1.4) By continuously training the generator module 8 and the discriminator module 9, the optimal generator module and discriminator module are obtained, so as to output the best supplemented data matrix. The training process is as follows:

[0062] (1.4.1) The generator G takes the data matrix X, the mask matrix M, and the random matrix Z as inputs, samples t D samples from the data matrix X and the mask matrix M samples t D independent and identically distributed samples from the random matrix Z samples t D independent and identically distributed samples from the random variable B where the elements of the random variable B = (B 1 ,..., B d ) are 0 or 1. First, randomly and uniformly sample k from {1,..., d}, and then let

[0063] (1.4.2) For i = 1,..., t D , the generator G generates the elements of the supplemented data matrix

[0064] y i = m i × x i + (1 - m i ) × G(X i , m i , z i )(23);

[0065] Then update the elements of the hint matrix

[0066] h i = b i × m i + 0.5(1 - b i )(24);

[0067] where × represents element - by - element multiplication, 1 is a vector of all 1s, and b i is an indicator scalar. When its value is 0, it means the data point is supplementary data; when its value is 1, it means non - supplementary data;

[0068] (1.4.3) The discriminator D takes the supplementary data matrix Y = (Y 1 ,..., Y d ) and the hint matrix H = (H 1 ,..., H d ) as inputs, and updates the discriminator D using stochastic gradient descent. The gradient is:

[0069]

[0070] where L D () represents the loss function of the discriminator D. Its goal is to distinguish the difference between real data and supplementary data by minimizing the loss. The loss formula is essentially a variant of the cross - entropy loss function, defined as follows:

[0071]

[0072] where, m' i = D(y i , h i ) is the estimated value in the mask matrix M predicted by the model, corresponding to the predicted value of the real mask matrix m i . And L D () is used to quantify the error between the model prediction m'i and the real label m i ;

[0073] (1.4.4) Sample t G samples from the data matrix X and the mask matrix M Sample t G independent and identically - distributed samples from the random matrix Z Sample \(t\) G independent and identically distributed samples from the random variable \(B\). For \(i = 1,\ldots,t\), G update the hint matrix elements

[0074] \(h\) i \(= b\) i \(\times m\) i \(+ 0.5(1 - b\) i \()(27)\);

[0075] (1.4.5) Update the generator \(G\) using stochastic gradient descent, with the gradient being

[0076]

[0077] where \(L\) G \(()\) is the adversarial loss of the generator, which encourages the predicted mask matrix \(m'\) by the model to be similar to the distribution of the true mask matrix \(m\). i \(L\) M \(()\) is the reconstruction loss function, which ensures that the generated data matrix \(x'\) by the generator can be as close as possible to the true data \(x\) i . \(a\) is a hyperparameter that balances the two losses. The specific definitions of the two loss functions are:

[0078]

[0079] The above formula means that according to the samples where \(b\) i \(= 0\), generate the estimated mask matrix \(m'\) so that it can match the true mask \(m\) i in the augmented data part. By optimizing this loss, the generator learns how to generate a mask matrix that can reconstruct the true data distribution in the part where data needs to be augmented.

[0080] Next, the definition of the reconstruction loss function \(L\) M \(()\) is given:

[0081]

[0082] where \(l\) M \(()\) is the unidirectional loss, which measures the difference between the generated data \(x'\) i and the true data \(x\) i , and is defined as follows:

[0083]

[0084] If \(x\) i is a continuous variable, the mean squared error is used to evaluate the generation result; otherwise, the cross - entropy is used to calculate the probability distribution of the generation result. By minimizing this loss function, the generator makes the completed data matrix \(x'\) iBe consistent with the real data x in different characteristics i Keep consistent.

[0085] (2) Feature extraction and selection module 4 is used to extract and select features from the preprocessed data, fuse and evaluate the features based on the separability criterion and stability criterion, and accurately and effectively select the optimal features;

[0086] (2.1) In the research on ship target classification and recognition based on high-resolution synthetic aperture radar images, the most basic and commonly used features include geometric structure features and gray-scale statistical features, and the electromagnetic scattering feature is a property that can better reflect the essence of the target in the synthetic aperture radar image. Therefore, the feature extraction module 10 first extracted 27 feature sets from the incomplete data ship target slices in the synthetic aperture radar image as shown in Table 1;

[0087] Table 1: 27 features extracted from ship target slices

[0088]

[0089]

[0090] (2.2) Then, the feature set containing the above 27 features is passed through the fusion evaluation module 11, and the evaluation results obtained for each feature according to the separability criterion and stability criterion are used for mean fusion, so as to simultaneously consider the prediction ability of individual features and the robustness of features. According to the fused evaluation values, the features with evaluation values less than 0.25 are removed to obtain the remaining features p 1 , p 2 , p 4 , p 6 , p 7 , p 17 , p 18 , p 20 , p 22 , p 23 , p 24 , p 25 , p 26 , p 27 ;

[0091] (2.3) In the correlation coefficient screening module 12, calculate the correlation coefficient between any two features

[0092]

[0093] where s i 、s j represent the means of features p i and p j , and σ i 、σ j represent the standard deviations of features pi and p j The standard deviation of. The correlation coefficient between features is used to measure the correlation between features, and redundant feature p is screened out 4 ,p 18 ,p 20 ,p 22 ;

[0094] (2.4) In the K-nearest neighbor screening module 13, for the candidate feature subset S 1 ={p 1 ,p 2 ,p 6 ,p 7 ,p 17 ,p 23}, S 2 ={p 1 ,p 2 ,p 6 ,p 7 ,p 17 ,p 24}, S 3 ={p 1 ,p 2 ,p 6 ,p 7 ,p 17 ,p 25}, S 4 ={p 1 ,p 2 ,p 6 ,p 7 ,p 17 ,p 26}, S 5 ={p 1 ,p 2 ,p 6 ,p 7 ,p 17 ,p 27} Finally, the K-nearest neighbor classifier is used to classify the candidate feature subsets, and the classification accuracy rate is used as the evaluation index for feature screening. The feature subset with the highest accuracy rate is selected to obtain the optimal feature subset S = {p 1 ,p 2 ,p 6 ,p 7 ,p 17 ,p 26}.

[0095] (3) Classification and recognition module 6 is used to classify and recognize the dataset after feature extraction and selection. Based on deep metric learning, an effective loss function is designed to train the deep convolutional neural network model, optimize the feature space distribution, increase the distance between features of different classes, and ensure the discriminability of the distribution; continuously compress the distance between features of the same class to improve the discriminability of the distribution.

[0096] (3.1) In the batch triplet sampling module 14, the dataset after feature extraction and selection is sampled in batches. Each time during training, three samples are selected as a batch of data triplets. The first sample x a is used as the anchor sample, the second positive sample x p has the same label as the anchor sample, and the third negative sample x n has a different label from the anchor sample.

[0097] (3.2) In the convolutional neural network module 15, the batch of data triplets is input into the convolutional neural network F and trained using the adaptive triplet loss function. After introducing the adaptive weights, the model can dynamically assign higher weights to important triplets, making the training more efficient and accurate, and at the same time avoiding excessive attention to invalid or extreme samples. The loss function is:

[0098]

[0099] where w i is the weight, and different adaptive weights are assigned according to different classes; b is a constant, called the margin.

[0100] (3.3) In the test phase, the trained convolutional neural network F is used to classify and recognize the ship targets in the dataset after feature extraction and selection, and the ship target recognition results are obtained.

[0101] The above embodiments are used to explain the present invention, rather than limit the present invention. Any modifications and changes made within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A ship recognition device for incomplete data based on generative adversarial networks and deep metrics, characterized in that: It includes a preprocessing module, a feature extraction and selection module, and a classification and recognition module; the preprocessing module is used to fill data with incomplete ship target slices in synthetic aperture radar images based on a generative adversarial network; the feature extraction and selection module is used to extract and select features from the data output by the preprocessing module, and to perform feature fusion evaluation and select the optimal features based on the separability criterion and the stability criterion; the classification and recognition module is used to train a deep convolutional neural network model based on the data set processed by the feature extraction and selection module, perform classification and recognition, and obtain a ship target recognition result.

2. The incomplete data ship identification device based on generative adversarial network and deep metric according to claim 1 is characterized by: The input of the preprocessing module is the incomplete ship target slice in the synthetic aperture radar image, and data filling is performed based on the generative adversarial network. The specific steps are as follows: (1) For the incomplete ship target slice data matrix X = (X1,...,X d ), let the mask matrix M with 0 or 1 as the element = (M1,...,M d ), where if the data of an element of matrix X is unclear, the corresponding element value of M is 0, otherwise it is 1; define a random matrix Z = (Z1,...,Z d ), which is independent of all other variables; (2) The generator is used to observe real data, supplement the data, and output a complete data matrix; (3) The discriminator distinguishes between the vectors derived from the real data and the supplemented vectors, and outputs an estimated matrix of the mask matrix M; (4) Train the generator and discriminator multiple times to output the best supplementary data matrix, where the training process is as follows: (4.1) The generator G takes the data matrix X, the mask matrix M and the random matrix Z as input and samples t from the data matrix X and the mask matrix M. D Samples Sample t from a random matrix Z D Independent and identically distributed samples Sample t from the random variable B D Independent and identically distributed samples The random variable B=(B1,...,B d ) elements are 0 or 1, first randomly sample k from {1,...,d}, and then let (4.2) for i=1,...,t D , the generator G generates the supplementary data matrix element y i : y i =m i ×x i +(1-m i )×G(x i ,m i ,z i ) (1); Then, update the hint matrix element h i ; h i =b i ×m i +0.5(1-b i ) (2); Where × represents the multiplication of corresponding elements, 1 is a vector with 1 as its element, and b i It is a scalar indicator. When the value is 0, it means that the data point is supplementary data, and when the value is 1, it means non-supplementary data; (4.3) The discriminator D is supplemented with the data matrix Y = (Y1, ..., Y d ) and the prompt matrix H = (H1,...,H d ) is input, and the discriminator D is updated using stochastic gradient descent, and the gradient is: Where L D () represents the loss function of the discriminator D, whose goal is to distinguish the difference between real data and supplementary data by minimizing the loss; the loss formula is essentially a variant of the cross entropy loss function, defined as follows: Among them, m′ i =D(y i ,h i ) is the estimated value in the mask matrix M predicted by the model, corresponding to the true mask matrix m i The predicted value of L D () is used to quantify the model prediction m′ i and the true label m i The error between (4.4) Sample t from the data matrix X and the mask matrix M G Samples Sample t from a random matrix Z G Independent and identically distributed samples Sample t from the random variable B G Independent and identically distributed samples For i=1,...,t G , update the prompt matrix elements: h i =b i ×m i +0.5(1-b i ) (5); (4.5) Use stochastic gradient descent to update the generator G, and the gradient is: Among them, L G () is the adversarial loss of the generator, which encourages the model to predict the mask matrix m′ i With the real mask matrix m i The distribution of L M () is the reconstruction loss function, ensuring that the data matrix x′ generated by the generator i Can be as close as possible to the real data x in terms of characteristics i ; a is a hyperparameter that balances the two losses; the specific definitions of the two loss functions are: The above formula means that according to b i = 0 samples, generate the mask matrix to estimate m′ i , so that it can be combined with the real mask m in the supplementary data part i The generator optimizes this loss to learn how to generate a mask matrix that can reconstruct the true data distribution in the part where the data needs to be supplemented. Next, we give the reconstruction loss function L M Definition of (): Among them, l M () is a one-way loss, measuring the generated data x′ i With the real data x i The difference is defined as follows: If x i If it is a continuous variable, the mean square error is used to evaluate the generated result; otherwise, the cross entropy is used to calculate the probability distribution of the generated result; the generator minimizes this loss function to make the completed data matrix x′ i Compared with the real data x in different features i Be consistent.

3. The incomplete data ship identification device based on generative adversarial network and deep metric according to claim 1 is characterized by: The feature extraction and selection module includes a feature extraction module, a fusion evaluation module, a correlation coefficient screening module, and a K nearest neighbor screening module, which are used to extract and select features from the preprocessed data, and to perform fusion evaluation on the features based on the separability criterion and the stability criterion to select the optimal features. The specific implementation steps are as follows: (1) extracting an initial feature set by a feature extraction module, including basic features and electromagnetic scattering features in the ship target classification and recognition research of high-resolution synthetic aperture radar images, wherein the basic features include geometric structure features and grayscale statistical features; (2) In the fusion evaluation module, the initial feature set is subjected to fusion evaluation, and the evaluation results of each feature obtained according to the separability criterion and the stability criterion are averaged and fused; based on the fused evaluation value, the features with an evaluation value less than 0.25 are eliminated to obtain the remaining feature set; (3) In the correlation coefficient screening module, calculate the correlation coefficient r between any two features ij : where s i 、s j Represents feature p i and p j The mean value, σ i , σ j Represents feature p i and p j The standard deviation of the feature; the correlation coefficient between features is used to measure the correlation between features and filter out redundant features; (4) In the K-nearest neighbor screening module, for the feature subset to be selected, the K-nearest neighbor classifier is used to classify the feature subset to be selected, and the classification accuracy is used as the evaluation index of feature screening. The feature subset with the highest accuracy is selected to obtain the optimal feature subset.

4. The incomplete data ship identification device based on generative adversarial network and deep metric according to claim 1 is characterized by: The classification and recognition module includes a batch triple sampling module and a convolutional neural network module, and the specific implementation steps are as follows: (1) In the batch triplet sampling module, batch sampling is performed on the data set after feature extraction and selection. Three samples are selected as batch data triplet for each training. The first sample x a As an anchor sample, the second positive sample x p Consistent with the anchor sample label, the third negative sample x n Inconsistent with the anchor sample label; (2) In the convolutional neural network module, batch data triplets are input into the convolutional neural network F and trained using the adaptive weighted triplet loss function. After introducing adaptive weights, the model can dynamically assign higher weights to important triplets. The loss function is as follows: where w i is the weight, and different adaptive weights are assigned according to different categories; b is a constant, called the boundary margin; (3) In the testing phase, the trained convolutional neural network F is used to classify and identify ship targets on the feature-extracted and selected data sets to obtain the ship target recognition results.

Images