Cross-resolution radar image domain adaptive recognition method based on fuzzy regularization

Abstract

The application discloses a kind of cross-resolution radar image domain adaptive identification method based on fuzzy regularization, first, the idea of adversarial learning is used to design feature extractor and domain discriminator, the minimization of domain difference is realized, then the class information of source domain SAR image is learned using classifier, accurate identification of labeled area is realized, and it is adjusted to unlabeled target area, finally through fuzzy regularization, the identification error of unlabeled resolution domain is minimized to the greatest extent, and the identification error minimization on unlabeled resolution domain is realized.The method of the application based on fuzzy regularization SAR cross-resolution domain self-adaptation minimizes the identification error on unlabeled resolution domain to the greatest extent, solves the problem of difference between different resolution domains in existing method, realizes effective domain self-adaptability, and shows excellent cross-resolution identification performance, and lays a good foundation for further target identification.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to a cross-resolution radar image domain adaptive recognition method based on fuzzy regularization. Background Technology

[0002] Automatic target recognition in SAR is a research hotspot in the field of SAR applications. However, existing methods for SAR automatic target recognition rely too heavily on experience and lack flexibility. Therefore, applying deep learning to SAR automatic target recognition methods is of great significance.

[0003] The paper "Target classification using the deep convolutional networks for SAR images," IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 8, pp. 4806–4817, 2016, proposes a fully convolutional network that replaces all dense layers with convolutional layers, demonstrating excellent recognition performance. The paper "SAR atr by a combination of convolutional neural network and support vector machines," IEEE Transactions on Aerospace and Electronic Systems, vol. 52, no. 6, pp. 2861–2872, 2016, proposes a combination of convolutional neural networks and support vector machines to incorporate prior knowledge and enhance robustness to imaging errors. The paper "Ldgan: A synthetic aperture radar image generation method for automatic target recognition, IEEE Transactions on Geoscience and Remote Sensing, vol.58, no.5, pp.3495–3508, 2020" proposes a novel image-to-image generation method that provides labeled samples for training the recognition model, effectively supplementing target information for automatic target recognition when samples are insufficient. However, existing deep learning-based SAR automatic target recognition methods typically assume that test samples and training samples have the same imaging resolution, but this assumption is difficult to satisfy in reality, leading to domain shift between different resolution domains and reducing the model recognition performance of SAR samples across resolution domains. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a cross-resolution radar image domain adaptation recognition method based on fuzzy regularization, which achieves effective domain adaptability and exhibits excellent cross-resolution recognition performance.

[0005] The technical solution adopted in this invention is: a cross-resolution radar image domain adaptation and recognition method based on fuzzy regularization, the specific steps of which are as follows:

[0006] S1. Problem modeling for cross-resolution domain adaptive recognition;

[0007] In cross-resolution SAR image recognition tasks, labeled SAR data with R resolution is represented as... from The unlabeled SAR data with R′ resolution is obtained from the above;

[0008] Where, x i Represents SAR samples, y i The label is represented by n, which represents the number of labeled SAR samples in the source domain; and m represents the number of unlabeled samples in the target domain. Represents the SAR data space. Represents the label space; the distribution of labeled data is the source domain. The distribution of unlabeled data is the target domain.

[0009] If the unlabeled data from the target domain is the data that truly needs to be identified, then the identification error of the target domain is defined as:

[0010]

[0011] in, h represents the recognition error in the target domain. f Represents a classification function, Let Θ(·) represent the expectation, and Θ(·) represent the indicator function. Then the problem expression for cross-resolution domain adaptive recognition is as follows:

[0012]

[0013] in, This represents the optimal labeling function that minimizes the recognition error of the target domain. This represents a hypothetical class.

[0014] S2, Adaptive adversarial domain;

[0015] Domain adaptation between SAR images with resolutions R and R′ can be described as The corresponding image sizes are S R and SR′ Training samples {x1, x2, ..., x} N The datasets are from the source domain dataset and the target domain dataset, where N represents the number of training samples. The domain label of the i-th sample is represented as a binary variable d. i If d i If it is 0, it means x i It comes from the source domain, otherwise from the target domain.

[0016] Adversarial domain adaptation is achieved through three modules: a feature extractor, a domain discriminator, and a classifier, as detailed below:

[0017] (1) Feature extractor;

[0018] Before feature extraction, the SAR samples {x1,x2,...,x} are... N} Size adjusted to S * :S * =max{S R ,S R′ The feature extractor is then equivalent to a mapping function that transforms the SAR image x into a D-dimensional feature vector. Right now:

[0019] f = G f (x;θ f (3)

[0020] in, Represents the real number field. θ represents the mapping function of the feature extractor. f This represents the parameters of the feature extractor.

[0021] (2) Domain discriminator;

[0022] Based on the extracted feature vector f, the domain discriminator determines whether the SAR image x comes from the source domain or the target domain, and provides the corresponding binary domain label d. i ∈{0,1}, its expression is as follows:

[0023] d = G d (f;θ d (4)

[0024] in, The mapping function θ represents the domain discriminant. d The parameters of the domain discriminator.

[0025] (3) Classifier;

[0026] The classifier uses extracted SAR image features to predict target features and outputs a C-dimensional prediction vector. The recognition process expression is as follows:

[0027]

[0028] Where C represents the number of categories, The mapping function θ represents the classifier. y This represents the parameters of the classifier.

[0029] S3, Fuzzy Regularization;

[0030] Class fuzziness is represented by the correlation between the predicted probabilities of each category, and the prediction results are then... Combined into a matrix, its expression is as follows:

[0031]

[0032] in, M represents the number of SAR samples in a batch, f M The vector represents the feature of the Mth sample, and the j-th column vector reflects the probability of the j-th class of all M input samples. express The M-th row vector. The probability correlation C between classes. jj′ The expression is as follows:

[0033]

[0034] in, express The j-th column vector, j,j′∈(1,...,C) and j≠j′, with the superscript T indicating transpose.

[0035] Then, the class normalization given by equation (8) is applied:

[0036]

[0037] in, Let j″ represent the normalized probability correlation between classes, where j″∈(1,...,C). Then the expression for the class fuzzy regularization function is as follows:

[0038]

[0039] S4. Loss function construction;

[0040] The loss function consists of a classifier loss term, a domain discriminator loss term, and a class fuzzy regularization term. The classifier loss term is represented by cross-entropy, and its expression is as follows:

[0041]

[0042] in, n represents the number of SAR images from the source domain, y iThis represents the label vector of the i-th SAR image. Let L represent the prediction vector for the i-th SAR image. The domain discriminator loss term L... d i This can be represented by the binary cross-entropy function, as shown in the following expression:

[0043] L d i =-d i logG d (G f (x i ))-(1-d i log(1-G) d (G f (x i (11)

[0044] The domain discriminant loss function expression for all SAR images from the source and target domains is as follows:

[0045]

[0046] Where m represents the number of unlabeled SAR images from the target domain.

[0047] The total loss function L is expressed as follows:

[0048] L = L y (θ f ,θ y )-L d (θ f ,θ d )+γL c (θ f ,θ y (13)

[0049] Where γ represents the coefficient of the fuzzy regularization term. The parameter update process is expressed as follows:

[0050]

[0051] Where μ represents the learning rate, Represent θ f ,θ d ,θ y The gradient of the loss function is used to obtain the optimal classifier parameters at the minimum of the loss function. Feature extractor parameters And domain discriminant parameters The expression is as follows:

[0052]

[0053] The beneficial effects of this invention are as follows: First, the method employs adversarial learning to design a feature extractor and a domain discriminator, minimizing domain differences. Then, a classifier learns the category information of the source domain SAR image to accurately identify labeled regions and adjusts them to unlabeled target regions. Finally, through class-based fuzzy regularization, the method minimizes recognition errors in the unlabeled resolution domain. This invention's method, based on class-based fuzzy regularization for SAR cross-resolution domain adaptation, minimizes recognition errors in the unlabeled resolution domain, solves the problem of differences between different resolution domains in existing methods, achieves effective domain adaptability, and demonstrates excellent cross-resolution recognition performance, laying a solid foundation for further target recognition. Attached Figure Description

[0054] Figure 1 This is a flowchart of a cross-resolution radar image domain adaptation recognition method based on fuzzy regularization according to the present invention.

[0055] Figure 2 This is a general framework diagram of the method of the present invention in an embodiment of the present invention.

[0056] Figure 3 This is a visualization diagram of the SAR features extracted in an embodiment of the present invention. Detailed Implementation

[0057] This invention primarily employs simulation experiments for verification. All steps and conclusions have been verified as correct using the PyTorch deep learning framework on the Windows 10 operating system platform. The method of this invention will be further described below with reference to the accompanying drawings and embodiments.

[0058] like Figure 1 The flowchart of a cross-resolution radar image domain adaptation and recognition method based on fuzzy regularization according to the present invention is shown below. The specific steps are as follows:

[0059] S1. Problem modeling for cross-resolution domain adaptive recognition;

[0060] like Figure 2 The diagram shown illustrates the overall framework of the method in this embodiment. Figure 2 In the cross-resolution SAR image recognition task shown, the labeled SAR data with R resolution is represented as from The unlabeled SAR data with R′ resolution is obtained from the above;

[0061] Where, x i Represents SAR samples, y iThe label is represented by n, which represents the number of labeled SAR samples in the source domain; and m represents the number of unlabeled samples in the target domain. Represents the SAR data space. Represents the label space; the distribution of labeled data is the source domain. The distribution of unlabeled data is the target domain.

[0062] If the unlabeled data from the target domain is the data that truly needs to be identified, then the identification error of the target domain is defined as:

[0063]

[0064] in, h represents the recognition error in the target domain. f Represents a classification function, Let Θ(·) represent the expectation, and Θ(·) represent the indicator function. Then the problem expression for cross-resolution domain adaptive recognition is as follows:

[0065]

[0066] in, This represents the optimal labeling function that minimizes the recognition error of the target domain. This represents a hypothetical class.

[0067] S2, Adaptive adversarial domain;

[0068] Domain adaptation between SAR images with resolutions R and R′ can be described as The corresponding image sizes are S R and S R′ Training samples {x1, x2, ..., x} N The datasets are from the source domain dataset and the target domain dataset, where N represents the number of training samples. The domain label of the i-th sample is represented as a binary variable d. i If d i If it is 0, it means x i It comes from the source domain, otherwise from the target domain.

[0069] like Figure 2 As shown, adversarial domain adaptation is implemented through three modules: a feature extractor, a domain discriminator, and a classifier, as detailed below:

[0070] (1) Feature extractor;

[0071] Its structure is based on the design of residual networks, and the structure and parameters of the classifier and domain discriminator are shown in Table 1.

[0072] Table 1

[0073]

[0074] Before feature extraction, the SAR samples {x1,x2,...,x} are... N} Size adjusted to S * :S * =max{S R ,S R′ The feature extractor is then equivalent to a mapping function that transforms the SAR image x into a D-dimensional feature vector. Right now:

[0075] f = G f (x;θ f (3)

[0076] in, Represents the real number field. θ represents the mapping function of the feature extractor. f This represents the parameters of the feature extractor.

[0077] (2) Domain discriminator;

[0078] Based on the extracted feature vector f, the domain discriminator determines whether the SAR image x comes from the source domain or the target domain, and provides the corresponding binary domain label d. i ∈{0,1}, its expression is as follows:

[0079] d = G d (f;θ d (4)

[0080] in, The mapping function θ represents the domain discriminant. d The parameters of the domain discriminator.

[0081] (3) Classifier;

[0082] The classifier uses extracted SAR image features to predict target features and outputs a C-dimensional prediction vector. The recognition process expression is as follows:

[0083]

[0084] Where C represents the number of categories, The mapping function θ represents the classifier. y This represents the parameters of the classifier.

[0085] S3, Fuzzy Regularization;

[0086] Classifiers often produce similar predicted probabilities due to easily confused categories. Based on this phenomenon, this embodiment uses the correlation between the predicted probabilities of each category to represent class ambiguity. To evaluate class ambiguity in statistics, this embodiment uses the prediction results... Combined into a matrix, its expression is as follows:

[0087]

[0088] in, M represents the number of SAR samples in the mini-batch, f M The vector represents the feature of the Mth sample, and the j-th column vector reflects the probability of the j-th class of all M input samples. express The M-th row vector. The probability correlation C between classes. jj′ The expression is as follows:

[0089]

[0090] in, express The j-th column vector, j,j′∈(1,...,C) and j≠j′, with the superscript T denoteing transpose. To avoid class imbalance, this embodiment uses the class normalization given by the following formula:

[0091]

[0092] in, Let j″ represent the normalized probability correlation between classes, where j″∈(1,...,C). Then the expression for the class fuzzy regularization function is as follows:

[0093]

[0094] S4. Loss function construction;

[0095] The loss function consists of a classifier loss term, a domain discriminator loss term, and a class fuzzy regularization term. The classifier loss term is represented by cross-entropy, and its expression is as follows:

[0096]

[0097] in, n represents the number of SAR images from the source domain, y i This represents the label vector of the i-th SAR image. Let L represent the prediction vector for the i-th SAR image. The domain discriminator loss term L... d i This can be represented by the binary cross-entropy function, as shown in the following expression:

[0098] L d i =-d i logG d (G f (x i ))-(1-di log(1-G) d (G f (x i (11)

[0099] The domain discriminant loss function expression for all SAR images from the source and target domains is as follows:

[0100]

[0101] Where m represents the number of unlabeled SAR images from the target domain.

[0102] The total loss function L is expressed as follows:

[0103] L = L y (θ f ,θ y )-L d (θ f ,θ d )+γL c (θ f ,θ y (13)

[0104] Where γ represents the coefficient of the fuzzy regularization term. The parameter update process is expressed as follows:

[0105]

[0106] Where μ represents the learning rate, Represent θ f ,θ d ,θ y The gradient of the loss function is used to obtain the optimal classifier parameters at the minimum of the loss function. Feature extractor parameters And domain discriminant parameters The expression is as follows:

[0107]

[0108] In this embodiment, to achieve better recognition performance of the method of the present invention, the regularization coefficient γ was set to 0.6 during the adaptive process from 128 SAR images to 64 SAR images. To verify the superiority of the method of the present invention, it was labeled with three other methods: residual network, interval difference, and domain adversarial neural network. The results are shown in Table 2, which represents the different cross-resolution SAR image recognition tasks performed in this embodiment on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset.

[0109] Table 2

[0110]

[0111] As can be seen from Table 2, the method of the present invention exhibits superior overall performance. Finally, the adaptability of the method of the present invention across different resolution domains was further verified through SAR feature visualization experiments, such as... Figure 3 As shown, Figure 3 (a) shows the feature visualization results for domain-free adaptive features. Figure 3 (b) is the feature visualization result after processing by the method of the present invention. It can be seen that the method of the present invention maps the extracted features to a new feature space and avoids domain shifting between SAR features of different resolution domains.

[0112] In summary, the method of this invention first employs adversarial learning to design a feature extractor and a domain discriminator to eliminate domain differences. Then, it utilizes a classifier to learn the category information of labeled SAR images, enabling the classifier to accurately identify targets in the labeled domain before adapting to the unlabeled target domain. Finally, it introduces class-fuzzy regularization to minimize recognition errors in the unlabeled resolution domain. This method, based on class-fuzzy regularization, adaptively minimizes recognition errors in the unlabeled resolution domain across SAR resolutions, resolving the issue of differences between different resolution domains in existing methods and laying a solid foundation for further target recognition.

[0113] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A cross-resolution radar image domain adaptation and recognition method based on fuzzy regularization, the specific steps of which are as follows: S1. Problem modeling for cross-resolution domain adaptive recognition; In cross-resolution SAR image recognition tasks, labeled SAR data with R resolution is represented as from The unlabeled SAR data with R′ resolution is obtained from the above; in, x i Represents SAR samples, y i The label is represented by n, which represents the number of labeled SAR samples in the source domain; and m represents the number of unlabeled samples in the target domain. Represents the SAR data space. Represents the label space; the distribution of labeled data is the source domain. The distribution of unlabeled data is the target domain. If the unlabeled data from the target domain is the data that truly needs to be identified, then the identification error of the target domain is defined as: in, h represents the recognition error in the target domain. f Represents a classification function, Let Θ(·) represent the expectation, and Θ(·) represent the indicator function; then the problem expression for cross-resolution domain adaptive recognition is as follows: in, This represents the optimal labeling function that minimizes the recognition error of the target domain. Represents a hypothetical class; S2, Adaptive adversarial domain; Domain adaptation between SAR images with resolutions R and R′ can be described as The corresponding image sizes are S R and S R′ Training samples {x1, x2, ..., x} N The datasets are from the source domain dataset and the target domain dataset, where N represents the number of training samples. The domain label of the i-th sample is represented as a binary variable d. i If d i If it is 0, it means x i If it comes from the source domain, otherwise it comes from the target domain; Adversarial domain adaptation is achieved through three modules: a feature extractor, a domain discriminator, and a classifier, as detailed below: (1) Feature extractor; Before feature extraction, the SAR samples {x1,x2,...,x} are... N } Size adjusted to S * :S * =max{S R ,S R′ The feature extractor is equivalent to a mapping function that transforms the SAR image x into a D-dimensional feature vector. Right now: f=G f (x;θ f ) (3) in, Represents the real number field. θ represents the mapping function of the feature extractor. f Indicates the parameters of the feature extractor; (2) Domain discriminator; Based on the extracted feature vector f, the domain discriminator determines whether the SAR image x comes from the source domain or the target domain, and provides the corresponding binary domain label d. i ∈{0,1}, its expression is as follows: d=G d (f;θ d ) (4) in, The mapping function θ represents the domain discriminant. d The parameters of the domain discriminator; (3) Classifier; The classifier uses extracted SAR image features to predict target features and outputs a C-dimensional prediction vector. The recognition process expression is as follows: Where C represents the number of categories, The mapping function θ represents the classifier. y Indicates the parameters of the classifier; S3, Fuzzy Regularization; Class fuzziness is represented by the correlation between the predicted probabilities of each category, and the prediction results are... Combined into a matrix, its expression is as follows: in, M represents the number of SAR samples in a batch, f M The vector represents the feature of the Mth sample, and the j-th column vector reflects the probability of the j-th class of all M input samples. express The M-th row vector; the probability correlation C between classes jj′ The expression is as follows: in, express The j-th column vector, j,j′∈(1,...,C) and j≠j′, with the superscript T indicating transpose; Then, the class normalization given by equation (8) is applied: in, Let j″∈(1,...,C) represent the normalized probability correlation between classes; then the expression for the class fuzzy regularization function is as follows: S4. Loss function construction; The loss function consists of a classifier loss term, a domain discriminator loss term, and a class fuzzy regularization term. The classifier loss term is represented by cross-entropy, and its expression is as follows: in, n represents the number of SAR images from the source domain, y i Let represent the label vector of the i-th SAR image. L represents the prediction vector of the i-th SAR image; the domain discriminator loss term L d i This can be represented by the binary cross-entropy function, as shown in the following expression: L d i =-d i logG d (G f (x i ))-(1-d i )log(1-G d (G f (x i ))) (11) The domain discriminant loss function expression for all SAR images from the source and target domains is as follows: Where m represents the number of unlabeled SAR images from the target domain; The total loss function L is then expressed as follows: L=L y (i f ,i y )-L d (i f ,i d )+γL c (i f ,i y ) (13) Where γ represents the coefficient of the fuzzy regularization term; the parameter update process expression is as follows: Where μ represents the learning rate, Represent θ f ,θ d ,θ y The gradient; to obtain the optimal classifier parameters at the minimum of the loss function. Feature extractor parameters And domain discriminant parameters The expression is as follows:

Images