ISAR image open set identification method based on generative discriminant

Abstract

The invention relates to an ISAR image open set identification method based on a generative discriminant. The method comprises the following steps: selecting a target type and generating an image as a data set; inputting the known class training set into a classifier, and updating parameters of the classifier by using the total loss of known class modules to obtain a preliminarily trained classifier; initializing parameters of a generator and a discriminator; the fixed generator inputs the known class training set and the first false sample into the discriminator, and updates parameters of the discriminator by using a discriminator loss function; the discriminator is fixed, noise vectors in the potential space are input into the generator, and parameters of the generator are updated through a generator loss function; inputting a known class training set and a false sample into the preliminarily trained classifier, and training the classifier based on a constraint condition; and a trained classifier is obtained through joint alternating optimization, and the trained classifier is used for recognizing the test set in combination with an extreme value distribution theory. According to the method, the accuracy and reliability of ISAR open set recognition are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of open set recognition, and particularly relates to an open set recognition method for ISAR images based on generative discriminant. Background Art

[0002] In a complex battlefield environment, an Inverse Synthetic Aperture Radar (ISAR) imaging system often faces the recognition requirement of unknown targets. Traditional classification methods based on the closed set hypothesis have serious defects because they cannot reject targets outside the training set. The core significance of ISAR open set recognition lies in: on the one hand, it can distinguish known / unknown targets in real time, providing a key decision-making basis for battlefield situation awareness; on the other hand, by controlling the open space risk, it can effectively avoid strategic misjudgments caused by misclassifying unknown targets as known targets. Open set recognition needs to solve both the closed set classification and unknown class rejection problems simultaneously. The core challenge lies in the control of open space risk. Currently, all ISAR image recognition technologies are based on the closed set hypothesis, and there are a large number of unknown targets in the actual task environment. Therefore, the research on ISAR image open set recognition technology has become an important research direction of inverse synthetic aperture radar.

[0003] The current technical exploration in the field of open set recognition mainly focuses on the fields of optical and Synthetic Aperture Radar (SAR) images, and the systematic research on open set recognition technology for ISAR imaging characteristics has not been formed. The existing open set recognition methods based on optical images and SAR images are mainly divided into two categories: discriminant-based and generative-based methods. Discriminant-based methods are represented by OpenMax, Extreme Value Theory (EVT), etc. The core mechanism is to construct a rejection boundary by combining statistical distribution modeling and threshold setting on the basis of a trained classification model, so as to effectively distinguish the known class from the open space; generative-based methods generate unknown class samples through models such as generative adversarial networks, jointly optimize the classification boundary with the known class data, and construct an extended classification system including unknown classes (i.e., "K + 1" classification, where K is the number of known classes). Typically, conditional variational autoencoders are used to generate adversarial samples that conform to known class targets, so that the classifier forms a more robust decision surface in adversarial training.

[0004] There is a significant research gap in the field of ISAR open set recognition, and there is a lack of publicly available ISAR datasets. For the existing discriminative open set recognition methods that only use known class learning and are applied to ISAR images, the reduction in the separability of known class features in ISAR images is more likely to cause aliasing of known class and unknown class features, making it difficult to set a suitable rejection threshold, thus limiting the performance of ISAR target open set recognition. In addition, due to the particularity of the imaging mechanism of ISAR images and the difficulty in ensuring the number of high-quality ISAR image samples, it is even more difficult to ensure the quality of unknown class ISAR images / features generated by using known class ISAR image data-driven. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides an ISAR image open set recognition method based on generating discriminants. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0006] The embodiment of the present invention provides a training method for an ISAR open set recognition network based on a generative discriminant, comprising the steps of:

[0007] S1, select the target type and generate images as a data set, and use some type data in the data set as a known class training set;

[0008] S2, inputting the known class training set into the classifier for classification, and updating the parameters of the classifier using the total loss of the known class module combined with the center loss and the cross entropy loss to obtain a preliminarily trained classifier;

[0009] S3, initialize the generator and discriminator parameters;

[0010] S4, fixing the generator, inputting the known class training set and the first fake sample into the discriminator for discrimination, and updating the parameters of the discriminator using the discriminator loss function, wherein the first fake sample is generated by inputting the first noise vector in the latent space into the generator;

[0011] S5, fixing the discriminator, inputting the second noise vector in the latent space into the generator to generate a second fake sample, and updating the parameters of the generator using the generator loss function;

[0012] S6, inputting the known class training set and the third false sample into the preliminarily trained classifier for classification, and training the classifier based on the constraints consisting of the classification loss and the open boundary constraint, wherein the third false sample is generated by inputting a third noise vector in the latent space into the generator;

[0013] S7. Repeat steps S2 - S6 for joint alternating optimization until the Nash equilibrium between the generated distribution and the true distribution is reached, and a trained classifier is obtained. The trained classifier is used to identify and classify a test set containing known and unknown classes in combination with the extreme value distribution theory.

[0014] In an embodiment of the present invention, the classifier includes a number of feature extraction units and an output unit; wherein, the number of feature extraction units are connected in sequence, and the output unit is connected after the last - stage feature extraction unit; each feature extraction unit includes a dropout layer and three convolutional blocks connected in sequence, and each convolutional block includes a convolutional layer, an MBN layer, and an activation function layer connected in sequence; the output unit includes an adaptive global average pooling layer, a fully - connected layer, and an activation function layer connected in sequence;

[0015] The generator includes a number of transposed convolution modules connected in sequence; wherein, the number of transposed convolution modules includes a transposed convolution layer, a batch normalization layer, and an activation function layer connected in sequence;

[0016] The discriminator includes a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, and an output layer connected in sequence; wherein, the first convolutional module and the fourth convolutional module both include a convolutional layer and an activation function layer connected in sequence; the second convolutional module and the third convolutional module both include a convolutional layer, a batch normalization layer, and an activation function layer connected in sequence.

[0017] In an embodiment of the present invention, the cross - entropy loss is:

[0018]

[0019] where loss 11 is the cross - entropy loss function for training the classification ability of the neural network, m is the batch size, label i is the ground truth, and f i is the feature of the i - th sample in the batch;

[0020] The center loss is:

[0021]

[0022] where loss 12 is the center loss function for constraining the model training, and c yi is a center vector maintained for each category respectively;

[0023] The total loss of the known - class module is:

[0024] L f = loss 11 + ω·loss 12

[0025] Among them, L f is the loss function of the known class module, and ω is the weight parameter.

[0026] In an embodiment of the present invention, step S4 includes:

[0027] Input the true data of the known class training set and the first fake sample into the discriminator to calculate the discrimination probability of each class of samples respectively:

[0028] D(x i ) → [0, 1]

[0029] D(G(z j )) → [0, 1]

[0030] Among them, D(x i ) is the discrimination probability of the true data in the known class training set, x i is the data in the true data batch in, m is the batch size; D(G(z j )) is the discrimination probability of the first fake sample, and G(z j ) is the first fake sample generated by inputting the noise vector batch into the generator;

[0031] Fix the generator, and use the discriminator loss function to update the parameters of the discriminator through gradient descent:

[0032]

[0033] Among them, θ D is the parameter of the discriminator, η D is the learning rate of the discriminator, which is used to control the step size of each parameter update, is the discriminator loss function L D with respect to the discriminator parameter θ D gradient, that is, the partial derivative of each parameter in the direction of the loss function.

[0034] In an embodiment of the present invention, step S5 includes:

[0035] Input the second noise vector in the latent space into the generator to generate the second fake sample G(z k );

[0036] Fix the discriminator, and use the generator loss function to update the parameters of the generator through gradient descent:

[0037]

[0038] Among them, θ Gis a parameter of the generator, η G is the learning rate of the generator, is the generator loss function, L G The gradient with respect to the parameter θ of the generator G , m is the batch size, D(G(z k )) is the discrimination probability of the second fake sample output by the discriminator, C(G(z k )) is the discrimination probability of the second fake sample output by the classifier, and β is a training hyperparameter.

[0039] In an embodiment of the present invention, the classification loss is:

[0040]

[0041] where L c is the classification loss function, p(y = k i |x i ) represents the probability that the sample x i belongs to the category k i , and N represents the number of known categories;

[0042] The open boundary constraint is:

[0043]

[0044] where L o is the open boundary constraint function, f(x i ) is the feature of the i-th data sample in the batch, p i is the corresponding prototype, d(·) represents the distance distribution, R is a learnable boundary parameter, ε = 0.001, G(z o ) is the third fake sample output by inputting the noise vector batch into the generator;

[0045] The constraint condition composed of the classification loss and the open boundary constraint is:

[0046] L t = L c + λL o

[0047] where λ is a weight parameter.

[0048] Another embodiment of the present invention provides a discriminative-based ISAR open set recognition method, including the steps of:

[0049] S1. Input a test set including known classes and unknown classes into a trained classifier to obtain the closed set classification probability, and the trained classifier is trained by the training method described in the above embodiment;

[0050] S2. Fit the extreme value cumulative distribution model for each category according to the Weibull cumulative distribution function and obtain the corresponding cumulative distribution function value F under the known category weibull (λ i , k i ), where λ i is the scale parameter of the i-th category, and k i is the shape parameter of the i-th category;

[0051] S3. Use the cumulative distribution function value under the known category to correct the closed-set classification probability of the test sample to obtain a corrected probability value;

[0052] S4. Calculate the probability that the sample to be tested is determined to be an out-of-distribution sample;

[0053] S5. Combine the corrected probability value and the probability of the out-of-distribution sample into a corrected score value, and use the maximum score value in the score values as the recognition result in the open-set scenario.

[0054] In an embodiment of the present invention, the closed-set classification probability is:

[0055]

[0056] where is the sample to be tested, is the feature output by the trained classifier, p i is the preliminary classification score that the sample to be tested belongs to the i-th known class, and N k is the number of known class categories;

[0057] The corrected probability value is:

[0058]

[0059] where p i ' is the corrected probability of the i-th category, is the cumulative distribution function value of the Weibull distribution;

[0060] The probability of the out-of-distribution sample is:

[0061]

[0062] where p u is the probability of the out-of-distribution sample;

[0063] The corrected score value is:

[0064] p = {p1', p'2,..., p' Nk , p u}

[0065] The recognition result in the open set scenario is as follows:

[0066] y = argmax(p).

[0067] In one embodiment of the present invention, step S2 includes:

[0068] S21. Input the training samples of the i-th class in the known class training set into the trained classifier to obtain the activation vectors of the training samples of the i-th class, and retain the activation vectors correctly classified as the i-th class samples by the classifier to obtain an activation vector set, and calculate the mean of the activation vector set as the centroid of the target class;

[0069] S22. Calculate the distance from each activation vector in the activation vector set to the centroid to obtain a distance set;

[0070] S23. Fit the maximum value distribution in the distance set, and input the maximum value distribution into the Weibull distribution model for fitting to obtain the cumulative distribution function value under the known class.

[0071] In one embodiment of the present invention, the cumulative distribution function of the Weibull distribution model is:

[0072]

[0073] where k is the shape parameter, λ is the scale parameter, and x is the maximum value distribution to be fitted.

[0074] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0075] 1. In the training method of the present invention, in the process of training the classifier with the known class training set, through the joint optimization of the center loss and the cross-entropy loss, the feature distribution of the known classes is made more compact and separated, realizing intra-class compactness and inter-class separation. Compared with the method that only uses the cross-entropy loss, it performs better on the benchmark dataset; when training with the unknown class combined with the known class distribution, the generator network generates unknown class features and jointly trains them with the known class features, enabling the classifier to learn the feature distribution of a wider range of ISAR data classes and improving the generalization ability for unknown classes; further, constraints corresponding to the real data distribution and the false data distribution are respectively defined for the classifier and combined, that is, the classifier is trained based on the constraint conditions composed of the classification loss and the open boundary constraint. This design makes the classifier more targeted and effective when dealing with unknown classes, enables the classifier to learn a wider feature distribution, and improves the generalization ability for unknown classes;

[0076] 2. In the recognition method of the present invention, the classifier trained with the samples obtained by the generative model is used to calculate the score of the test sample, and the cumulative distribution function value is further used for correction, effectively adjusting the judgment of the classifier on the score of the test sample and improving the accuracy and reliability of ISAR open set recognition. Description of the Drawings

[0077] Figure 1 Schematic diagram of an ISAR open set recognition method based on generative discriminant provided by an embodiment of the present invention;

[0078] Figure 2 Flow schematic diagram of a training method for an ISAR open set recognition network based on generative discriminant provided by an embodiment of the present invention;

[0079] Figure 3 Classifier structure diagram provided by an embodiment of the present invention;

[0080] Figure 4 Generator structure diagram provided by an embodiment of the present invention;

[0081] Figure 5 Discriminator structure diagram provided by an embodiment of the present invention;

[0082] Figure 6 Flow schematic diagram of an ISAR open set recognition method based on generative discriminant provided by an embodiment of the present invention;

[0083] Figure 7 Three-dimensional models of nine space targets provided by an embodiment of the present invention;

[0084] Figure 8 Schematic diagram of a set of imaging results corresponding to nine space targets. Detailed Embodiments

[0085] The present invention will be further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0086] Embodiment 1

[0087] In this embodiment, in combination with the actual requirements of the ISAR image open set recognition method, an open set recognition algorithm that can accurately recognize known class targets and reject unknown class targets is studied, so that the recognition results better meet the actual engineering requirements. Please refer to Figure 1 , Figure 1 Schematic diagram of an ISAR open set recognition method based on generative discriminant provided by an embodiment of the present invention, Figure 1 where (a) closed set training and (b) unknown class generation and model update are the training stages, Figure 1In (c), it is the open-set test phase. The closed-set training module constructs constraints by combining the center loss and the cross-entropy loss during the closed-set modeling phase, ensuring inter-class dispersion while enhancing the intra-class feature compactness of known classes and reserving more feature space for unknown class features. After the initial joint learning, the unknown class feature generation and model update module combines the adversarial feature generation mechanism to generate ISAR unknown class images / features, and jointly trains with known class features to further enhance the discriminability of known and unknown classes in the feature space. The open-set test module combines the extreme value theory, obtains the preliminary feature distribution through the classifier, calculates the distances between the known class features and the corresponding class prototypes, and determines the boundaries of the known class distributions in combination with the extreme value theory, thereby correcting the scores of each sample in the final test set to distinguish between known and unknown ISAR samples to be tested.

[0088] Please combine Figure 1 in (a), (b) and Figure 2 , Figure 2 is a schematic flow chart of a training method for a discriminative generative-based ISAR open-set recognition network provided by an embodiment of the present invention. The training method includes the steps:

[0089] S1. Select the target type and generate images as the data set, and use part of the data of the data set as the known class training set;

[0090] S2. Input the known class training set into the classifier for classification, and update the parameters of the classifier using the total loss of the known class module jointly composed of the center loss and the cross-entropy loss to obtain the preliminarily trained classifier;

[0091] S3. Initialize the parameters of the generator and the discriminator;

[0092] S4. Fix the generator, input the known class training set and the first fake sample into the discriminator for discrimination, and update the parameters of the discriminator using the discriminator loss function, where the first fake sample is generated by inputting the first noise vector in the latent space into the generator;

[0093] S5. Fix the discriminator, input the second noise vector in the latent space into the generator to generate the second fake sample, and update the parameters of the generator using the generator loss function;

[0094] S6. Input the known class training set and the third fake sample into the preliminarily trained classifier for classification, and train the classifier based on the constraint conditions composed of the classification loss and the open boundary constraint, where the third fake sample is generated by inputting the third noise vector in the latent space into the generator;

[0095] S7. Repeat steps S2 - S6 for joint alternating optimization until the Nash equilibrium between the generated distribution and the true distribution is reached, and then the trained classifier is obtained. The trained classifier is used to identify and classify the test set containing known classes and unknown classes in combination with the extreme value distribution theory.

[0096] Please refer to Figure 3 , Figure 3 which is the classifier structure diagram provided by the embodiment of the present invention. The classifier Net - C adopts a multi - layer convolution structure and a multi - domain normalization strategy (Multi - Batch Normalization, MBN) to enhance the performance of open - set recognition. The classifier Net - C includes several feature extraction units and an output unit; among them, several feature extraction units are connected in sequence, and the output unit is connected after the last - stage feature extraction unit; each feature extraction unit includes a dropout layer and three convolution blocks connected in sequence, and each convolution block includes a convolution layer Conv2D, an MBN layer, and an activation function layer LeakyReLU connected in sequence; the output unit includes an adaptive global average pooling layer GAP, a fully - connected layer FC, and an activation function layer sigmoid. Exemplarily, the number of several feature extraction units is 3 groups.

[0097] Specifically, the classifier Net - C network first alleviates the over - fitting risk through a dropout layer, and then sequentially passes through two 3×3 convolution layers. After each convolution, domain - adaptive normalization is performed through MBN, and then processed by the LeakyReLU activation function; subsequently, downsampling is achieved through the third convolution block to extract preliminary features. Next, the network respectively passes through the second and third groups of feature extraction units with a similar structure to capture higher - level features. After the spatial information is reduced to a single value through the adaptive global average pooling layer (GlobalAverage Pooling, GAP), it is flattened into a one - dimensional vector and the final classification result is output through the fully - connected layer FC. In the whole structure, the MBN module effectively alleviates the cross - domain differences in data distribution by providing independent normalization parameters for different domains (or different data distributions), thereby enhancing the ability of the network to distinguish between known and unknown class data in the open - set recognition task.

[0098] Please refer to Figure 4 , Figure 4 which is the generator structure diagram provided by the embodiment of the present invention. The generator Net - G includes several transposed convolution modules connected in sequence; among them, several transposed convolution modules include a transposed convolution layer DeConv2D, a batch normalization layer BN, and an activation function layer ReLU connected in sequence. Exemplarily, the number of several transposed convolution modules is 4.

[0099] Specifically, the generator consists of a series of transposed convolution layers (i.e., deconvolution layers), batch normalization layers (BatchNormalization, BN), and activation function layers, gradually mapping the low-dimensional noise vector or the feature map of the autoencoder to the high-dimensional image space. Among them, the transposed convolution layer is responsible for gradually increasing the spatial dimension of the feature map, the activation function is used to introduce non-linearity, and the batch normalization layer can effectively stabilize the training process and accelerate the convergence rate of the model.

[0100] It should be noted that in this embodiment, a variational autoencoder (Variational Autoencoder, VAE) or its variant (such as a conditional variational autoencoder Conditional VAE, CVAE) can be used as the generator network to generate unknown class features.

[0101] Please refer to Figure 5 , Figure 5 which is the discriminator structure diagram provided by the embodiment of the present invention. The discriminator Net-D includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, and an output layer connected in sequence; among them, both the first convolution module and the fourth convolution module include a convolution layer Conv2D and an activation function layer connected in sequence. The activation function of the first convolution module uses LeakyReLU, and the activation function of the fourth convolution module uses sigmoid; both the second convolution module and the third convolution module include a convolution layer Conv2D, a batch normalization layer BN, and an activation function layer LeakyReLU connected in sequence.

[0102] Specifically, the discriminator designed in the generative adversarial network adopts a fully convolutional network architecture, and constructs a deep feature extraction structure by stacking multiple convolutional layers, regularization layers and non-linear activation functions. The main body of the network consists of multiple convolutional modules connected in series. Each module includes a convolutional layer Conv2D, a batch normalization layer BN, and a LeakyReLU activation function (the first convolutional module and the fourth convolutional module are adjusted). The spatial dimension is compressed by gradually downsampling through convolutional layers with a stride of 2. At the same time, the number of channels is gradually doubled from the initial 3 layers to 512 layers. Finally, the feature map is converted into a scalar discriminant value through a global average pooling layer. This architecture uses convolutional layers instead of fully connected layers to retain the spatial topological characteristics of the input image, enabling the network to adaptively process spatial target ISAR images of different sizes. By utilizing the local receptive field characteristics of convolutional operations, it can capture both global structural features such as target contours and scatterer distributions, and local detail features such as edge textures and noises in a multi-level feature space. The network suppresses the random noise interference that may be introduced during the imaging process through the batch normalization layer, and combines the LeakyReLU activation function to maintain the gradient flow during backpropagation, enabling the discriminator to effectively distinguish the subtle differences between the generated image and the real ISAR image in terms of the consistency of scattering characteristics and the rationality of target geometric structures, thereby guiding the generator to output synthetic images that conform to the physical imaging law. This discriminator realizes the end-to-end adaptive feature learning and authenticity discrimination functions for spatial target ISAR images with variable data dimensions and significant noise interference.

[0103] In a specific embodiment, the target models of the dataset in step S1 are nine satellite models: Hubble Space Telescope, Tiangong-1 satellite, WorldView satellite, Aqua satellite, Envisat satellite, ACE satellite, AIM satellite, Aquarius satellite, and Aura satellite. Electromagnetic simulation data of each satellite target at different azimuth and elevation angles are obtained using electromagnetic simulation technology, and electromagnetic simulation images are obtained by imaging using the Polar Format Algorithm (PFA). And a part of the electromagnetic simulation images is used as the known class training set, and the other part is used as the unknown test set.

[0104] In a specific embodiment, step S2 is a closed-set training process.

[0105] Specifically, for the training of the known class module, first select the classifier Net-C as the feature extractor, extract features from the input data of the known class training set through the feature extractor, then input these features into the classification layer for classification, and update the parameters of the classifier using the total loss of the known class module to obtain a preliminarily trained classifier.

[0106] To improve the classification performance of the model, first, the cross-entropy loss is used to train the classification ability of the neural network, denoted as loss1:

[0107]

[0108] where loss 11 is the cross-entropy loss function for training the classification ability of the neural network, m is the batch size, label i is the ground truth, and f i is the feature of the i-th sample in the batch.

[0109] It should be noted that the contrastive loss or the graph-based loss function can also be used to train the classification ability of the neural network to optimize the feature distribution of known classes

[0110] Furthermore, in order to obtain compact intra-class features and reserve sufficient space for the unknown class distribution, the center loss is adopted to constrain the training of the model. The center loss is defined as follows:

[0111]

[0112] where loss 12 is the center loss function for constraining the model training, is a center vector maintained separately for each category.

[0113] Finally, the total loss of the known class module is designed as follows:

[0114] L f = loss 11 + ω·loss 12

[0115] where L f is the total loss function of the known class module, and ω is the weight parameter used to balance the contributions between intra-class compactness and inter-class separation.

[0116] In a specific embodiment, steps S3 - S6 are the training stage for generating unknown class features and jointly optimizing the classifier distribution with known classes. The specific implementation steps are as follows:

[0117] S3. The parameters of the generator Net-G and the discriminator Net-D are randomly initialized. The input of the generator is a random noise vector z ~ p z (z) in the latent space, and the output is the generated sample G(z); the input of the discriminator is the real data x ~ p(x) and the generated sample G(z), and the output is a scalar representing the probability that the input sample comes from the real distribution.

[0118] S4. Train the discriminator Net-D, which specifically includes:

[0119] First, input the true data of the known class training set and the first fake samples into the discriminator to calculate the discrimination probabilities of each class of samples respectively.

[0120] Specifically, input the noise vector batch into the generator to generate fake samples G(z j ), and then input the true data batch and the fake samples G(z j ) into the discriminator. The discriminator calculates the probabilities for the two classes of samples respectively:

[0121] D(x i ) → [0, 1]

[0122] D(G(z j )) → [0, 1]

[0123] Among them, D(x i ) is the discrimination probability of the true data in the known class training set, x i is the data in the true data batch , and m is the batch size; D(G(z j )) is the discrimination probability of the first fake sample, and G(z j ) is the first fake sample generated by inputting the noise vector batch into the generator.

[0124] To maximize the discrimination score of the discriminator for the known class samples and minimize the discrimination score of the discriminator for the generated samples, design the discriminator loss function L D as follows:

[0125]

[0126] Then, fix the generator Net-G and use the discriminator loss function to update the parameters of the discriminator Net-D through gradient descent (such as the Adam optimizer):

[0127]

[0128] Among them, θ D is the parameter of the discriminator, η D is the learning rate of the discriminator, which is used to control the step size of each parameter update, is the gradient of the discriminator loss function L D with respect to the discriminator parameter θ D , that is, the partial derivative of each parameter in the direction of the loss function.

[0129] Thus, the discriminator Net-D has the ability to distinguish between known-class samples and generated samples. That is, when known-class samples are fed into the discriminator, the discriminator outputs a relatively high discrimination score, and when generated samples are fed into the discriminator, the discriminator outputs a relatively low discrimination score.

[0130] S5. Training of the generator Net-G. Specifically, it includes:

[0131] First, a second batch of noise vectors in the latent space is input into the generator to generate a second set of fake samples G(z k ).

[0132] The optimization objective of the generator is designed as a dynamic balance process under dual constraints: on the one hand, it is necessary to generate samples highly similar to known classes through adversarial games to deceive the discriminator, and on the other hand, it is necessary to avoid falling into the following two extreme situations to ensure the performance of Open Set Recognition (OSR). First, the generator needs to maintain the effective discrimination gradient of the discriminator: if the generated samples are too realistic (i.e., the discriminator cannot distinguish the distribution difference between the generated samples and the real samples of known classes), the adversarial loss gradient disappears, resulting in the generator stopping optimization and thus being unable to further explore the potential distribution boundary of unknown classes. Second, the generator needs to avoid the uniform distribution trap: if the generated samples only cover simple unknown class regions that are significantly different from known classes (i.e., the generated distribution is overly uniform), it is difficult to generate "difficult negative samples" adjacent to the decision boundary of known classes, thereby weakening the discriminative ability of the model for complex unknown classes. To achieve this goal, the generator loss function L G is extended to:

[0133]

[0134] where m is the batch size, D(G(z k )) is the discrimination probability of the second set of fake samples output by the discriminator, C(G(z k )) is the discrimination probability of the second set of fake samples output by the classifier, and β is a training hyperparameter. Then, fix the discriminator and update the parameters of the generator Net-G using the generator loss function through gradient descent:

[0135]

[0136] where θ G is the parameter of the generator, η G is the learning rate of the generator, is the gradient of the generator loss function L G with respect to the parameter θ G of the generator.

[0137] S6. Training of classifier Net-C. The third batch of noise vectors in the latent space is input into the generator to generate the third fake samples G(z o ). The real data batch and the third fake samples G(z o ) are input into the preliminarily trained classifier Net-C. Through the classifier C, the latent feature z c and the output y c of the classification layer are obtained.

[0138] To enable the classifier to simultaneously learn the distribution boundaries of known and unknown classes, first design the classification loss to optimize the classification accuracy of known classes, and optimize the model by maximizing the correct classification probability of known class samples. The classification loss is:

[0139]

[0140] where L c is the classification loss function, p(y = k i |x i ) represents the probability that the sample x i belongs to the class k i , and N represents the number of known classes.

[0141] It should be noted that a loss function based on metric learning (such as Metric Contrastive Loss) or a loss function based on clustering (such as Clustering Loss) can also be used as the classification loss function.

[0142] Meanwhile, to reduce the risk that unknown class samples are misclassified as known classes, further optimize the boundary between known and unknown classes, and improve the discrimination ability of the model, design the open boundary constraint:

[0143]

[0144] where L o is the open boundary constraint function, f(x i ) is the feature of the i-th data sample in the batch, p i is the corresponding prototype, d(·) represents the distance distribution, R is a learnable boundary parameter, ε = 0.001 is to avoid division by zero, and G(z o ) is the third fake sample output by inputting the batch of noise vectors into the generator. This design makes it more likely that unknown class samples are restricted in the internal space, thus reducing the open space risk.

[0145] Finally, the constraint conditions composed of the classification loss and the open boundary constraint are:

[0146] L t = L c + λL o

[0147] Among them, λ is a weight parameter used to balance the contributions between the classification loss and the open boundary constraint.

[0148] S7. Repeat steps S2 - S6 for joint alternating optimization, that is, in the first training, S2 - S6 are carried out in sequence. In subsequent trainings, after S2 is completed, S4 - S6 are trained until the Nash equilibrium (loss convergence or reaching the number of iterations) between the generated distribution and the true distribution is achieved, then terminate the training to obtain the trained classifier.

[0149] After the training is completed, perform an open set test and introduce the extreme value distribution theory in the detection stage. Please refer to Figure 6 , Figure 6 which is a schematic flow chart of an ISAR open set recognition method based on generative discriminant provided by an embodiment of the present invention. The recognition method includes the steps:

[0150] S1. Input a test set including known classes and unknown classes into the trained classifier to obtain the closed set classification probability. The trained classifier is trained by the training method of the above - mentioned embodiment.

[0151] Specifically, input the image to be measured, denote the sample to be measured as obtain its feature representation through Net - C after softmax normalization, obtain the closed set classification probability p:

[0152]

[0153] Among them, is the sample to be measured, is the feature output by the trained classifier, and p i is the preliminary classification score that the sample to be measured belongs to the i - th known class, and N k is the number of known class categories.

[0154] S2. Fit the extreme value cumulative distribution model for each class according to the Weibull cumulative distribution function and obtain the corresponding cumulative distribution function value F weibull (λ i , k i ), among which, λ i is the scale parameter of the i - th class, and k i is the shape parameter of the i - th class. Specifically, it includes the steps:

[0155] S21. Input the training samples of the i-th class in the known class training set into the trained classifier to obtain the activation vectors of the training samples of the i-th class, and retain the activation vectors correctly classified as the i-th class samples by the classifier to obtain an activation vector set, and calculate the mean value of the activation vector set as the centroid of the target class.

[0156] Specifically, input the training samples and load the optimal classifier model obtained from training. Taking the i-th known class as an example, input all the training samples of the i-th class into the classifier to obtain the activation vectors (Activation Vector, AV) of the training samples of the i-th class, and retain the AVs correctly classified as the i-th class samples by the classifier. Denote the set of retained AVs as AV i ={AV1, AV2, …, AV m}, and calculate its mean value MAV i , where m refers to that m samples are recognized as the i-th class by the DCNN, and MAV i is the centroid of the i-th class samples.

[0157] S22. Calculate the distance from each activation vector in the activation vector set to the centroid to obtain a distance set.

[0158] Specifically, calculate the distance sets of each known class. Calculate the distance from AV i ={AV1, AV2, …, AV m} to its centroid MAV i , and denote the distance set as D i ={D1, D2, …, D m}.

[0159] S23. Fit the maximum value distribution in the distance set, and input the maximum value distribution into the Weibull distribution model for fitting to obtain the cumulative distribution function value under the known class.

[0160] Specifically, use the libMR algorithm library to fit the maximum value distribution in the distance set, and further input the fitted maximum value distribution into the Weibull distribution model to obtain the calibration probability value. The cumulative distribution function of the Weibull distribution is:

[0161]

[0162] where x is the fitted maximum value distribution, k is the shape parameter, and λ is the scale parameter. By fitting the shape parameter and the scale parameter, the extreme value cumulative distribution model F i corresponding to each known class can be obtained.

[0163] It should be noted that in this embodiment, a method based on probability distribution (such as the Bayesian method) or a method based on a machine learning model (such as the ensemble learning method) can also be used to correct the classifier score.

[0164] S3. Correct the closed-set classification probability of the test sample using the cumulative distribution function values in the known classes to obtain the corrected probability value.

[0165] Specifically, by modeling the extreme value cumulative distribution model F corresponding to each class i , the cumulative distribution function values in each known class can be obtained Given that the cumulative distribution function has the property of monotonically increasing, when the distance between the test sample and the prototype of a certain class increases, its will also increase accordingly, indicating that the possibility of this sample belonging to this class is relatively low. Therefore, when correcting the score value of the test sample, this negative correlation should be considered, and the specific correction implementation follows the following formula:

[0166]

[0167] where p i ' is the corrected probability of the i-th class, is the cumulative distribution function value of the Weibull distribution, λ i is the scale parameter of the i-th class, and k i is the shape parameter of the i-th class.

[0168] S4. Calculate the probability that the test sample is determined to be an out-of-distribution sample.

[0169] Specifically, after correction, when the test sample has a significant geometric deviation from the prototype center of a specific known class in the feature space, the model will assign it a very low confidence score for the corresponding class. Based on this, the probability p u that the sample is determined to be an out-of-distribution sample can be analytically expressed as:

[0170]

[0171] S5. Combine the corrected probability value and the probability of the out-of-distribution sample into the corrected score value, and take the maximum score value in the score value as the recognition result in the open-set scenario.

[0172] Specifically, the corrected score value is:

[0173] p = {p1', p'2, …, p' Nk , p u}

[0174] The recognition result in the open-set scenario is:

[0175] y = argmax(p).

[0176] This embodiment is further verified through the following simulation.

[0177] (1) Simulation conditions

[0178] The target models of the dataset selected nine spatial target models, such as Figure 7 shown Figure 7 are the three-dimensional models of the nine spatial targets provided by the embodiments of the present invention. Each target generates 1600 ideal images, and the total number of images is 14400, Figure 8 which is a schematic diagram of a set of imaging results corresponding to the nine spatial targets.

[0179] (2) Simulation experiment results

[0180] The openness of the ISAR recognition task is expressed as:

[0181]

[0182] where C TA , C TR and C TE respectively represent the set of classes to be recognized, the set of classes used in training, and the set of classes used during testing, and |·| represents the number of classes in the corresponding set. The greater the openness, the greater the corresponding problem openness, and when the openness is equal to 0, the problem is completely closed.

[0183] The metric Accuracy is used to detect the closed-set recognition accuracy:

[0184]

[0185] where represents the number of known class samples correctly classified in the test set, represents the total number of known class samples in the test set.

[0186] The Area Under the Receiver Operating Characteristic Curve (AUROC) is selected to detect the open-set recognition situation. AUROC measures the ability of the model to correctly identify the positive class and the negative class at all possible classification thresholds. The AUROC value ranges from 0 to 1, and the closer the value is to 1, the stronger the ability of the model to distinguish between known and unknown classes in the open-set recognition task. Please refer to Figure 7 , Figure 7 which is Figure 7Schematic diagram of ROC curve and AUROC score. The ROC curve is a two-dimensional graph, where the horizontal axis represents the False Positive Rate (FPR), and the vertical axis represents the True Positive Rate (TPR). TPR represents the proportion of samples correctly identified as known classes among all actual known samples, while FPR represents the proportion of unknown samples misidentified as known classes.

[0187] When inputting the test set samples and in the scenario with an openness of 6.25%, the closed-set recognition accuracy of the above method on the ISAR dataset can reach 98.03%, and the AUROC index can reach 89.4%. The results are shown in Table 1.

[0188] Table 1 ISAR open-set recognition results of nine targets

[0189]

[0190] In the training method of this embodiment, in the link of training the classifier with the known class training set, through the joint optimization of the center loss and the cross-entropy loss, the known class feature distribution becomes more compact and separated, realizing intra-class compactness and inter-class separation, and improving the performance of the ISAR known class classifier. Compared with the method that only uses the cross-entropy loss, it performs better on the benchmark dataset; when training with the unknown class combined with the known class distribution, the generator network is used to generate unknown class features and jointly train them with the known class features, enabling the classifier to learn the feature distribution of a wider range of ISAR data classes and improving the generalization ability for unknown classes; further, constraints corresponding to the real data distribution and the fake data distribution are respectively defined for the classifier and combined, that is, the classifier is trained based on the constraint conditions composed of the classification loss and the open boundary constraint. This design makes the classifier more targeted and effective when dealing with ISAR unknown classes, enables the classifier to learn a wider feature distribution, and improves the generalization ability for unknown classes.

[0191] In the recognition method of this embodiment, the classifier trained by expanding the samples obtained by the generative model is used to calculate the scores of the test samples, and further combined with the EVT extreme value theory, the cumulative distribution function value is used for calibration, effectively adjusting the classifier's judgment of the scores of the test samples and improving the accuracy and reliability of ISAR open-set recognition.

[0192] The method of this embodiment significantly improves the open set recognition rate and the generalization ability of the classifier by integrating the distribution modeling ability of the generative model and explicitly constraining the risk boundary of the open space. Compared with the methods that only rely on a single loss function, generative model, or do not use the EVT theory, it has obvious advantages. It solves the problem that the existing discriminative-based methods overly rely on the closed set classification boundary of known classes and are difficult to effectively identify unknown targets beyond the training distribution, resulting in a high risk in the open space. At the same time, it makes up for the problem that the unknown class samples generated by the existing generative adversarial networks may be too similar to the features of the known classes and cannot simulate the unknown class distribution.

[0193] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A training method for an ISAR open set recognition network based on generative discriminant, characterized in that Including the steps: S1. Select a target type and generate images as a data set, and use part of the type data in the data set as a known class training set; S2. Input the known class training set into a classifier for classification, and update the parameters of the classifier using the total loss of the known class module jointly composed of center loss and cross-entropy loss to obtain a preliminarily trained classifier; S3. Initialize the parameters of the generator and discriminator; S4. Fix the generator, input the known class training set and the first fake sample into the discriminator for discrimination, and update the parameters of the discriminator using the discriminator loss function, where the first fake sample is generated by inputting the first noise vector in the latent space into the generator; S5. Fix the discriminator, input the second noise vector in the latent space into the generator to generate a second fake sample, and update the parameters of the generator using the generator loss function; S6. Input the known class training set and the third fake sample into the preliminarily trained classifier for classification, and train the classifier based on the constraint conditions composed of classification loss and open boundary constraints, where the third fake sample is generated by inputting the third noise vector in the latent space into the generator; S7. Repeat steps S2 to S6 for joint alternating optimization until the Nash equilibrium of the generated distribution and the real distribution is reached, and obtain a trained classifier, and the trained classifier is used to combine the extreme value distribution theory to identify and classify a test set including known classes and unknown classes.

2. The training method of the ISAR open set recognition network based on generative discriminative according to claim 1, characterized in that The classifier includes a plurality of feature extraction units and an output unit; wherein, the plurality of feature extraction units are connected in sequence, and the output unit is connected after the last-level feature extraction unit; each feature extraction unit includes a dropout layer and three convolutional blocks connected in sequence, and each convolutional block includes a convolutional layer, an MBN layer and an activation function layer connected in sequence; the output unit includes an adaptive global average pooling layer, a fully connected layer and an activation function layer connected in sequence; The generator includes a plurality of transposed convolution modules connected in sequence; wherein, the plurality of transposed convolution modules include a transposed convolution layer, a batch normalization layer and an activation function layer connected in sequence; The discriminator includes a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module and an output layer connected in sequence; wherein, both the first convolutional module and the fourth convolutional module include a convolutional layer and an activation function layer connected in sequence; both the second convolutional module and the third convolutional module include a convolutional layer, a batch normalization layer and an activation function layer connected in sequence.

3. The training method of the ISAR open set recognition network based on the generative discriminative model according to claim 1, characterized in that The cross-entropy loss is: Among them, loss 11 is the cross-entropy loss function for training the classification ability of the neural network, m is the batch size, and label i is the ground truth, and f i is the feature of the i-th sample in the batch; The center loss is: Among them, loss 12 is the central loss function that constrains the model training, and c yi is a center vector maintained separately for each category; The total loss of the known class module is: L f = loss 11 + ω·loss 12 Among them, L f is the loss function of the known class module, and ω is the weight parameter.

4. The training method of the ISAR open set recognition network based on the generative discriminative model according to claim 1, characterized in that Step S4 includes: Input the real data of the known class training set and the first fake sample into the discriminator to calculate the discrimination probability of each class of samples respectively: D(x i ) → [0, 1] D(G(z j )) → [0, 1] Among them, D(x i ) is the discrimination probability of the real data in the known class training set, x i is the data in the real data batch , and m is the batch size; D(G(z j )) is the discrimination probability of the first fake sample, and G(z j ) is the first fake sample generated by inputting the noise vector batch into the generator; Fix the generator, and update the parameters of the discriminator by gradient descent using the discriminator loss function: Among them, θ D is the parameter of the discriminator, and η D is the learning rate of the discriminator, which is used to control the step size of each parameter update. is the loss function L D of the discriminator with respect to the discriminator parameter θ D gradient, that is, the partial derivative of each parameter in the direction of the loss function.

5. The training method of the ISAR open set recognition network based on the generative discriminative model according to claim 1, wherein Step S5 includes: Input the second noise vector in the latent space into the generator to generate a second fake sample G(z k ); Fix the discriminator, and update the parameters of the generator by gradient descent using the generator loss function: Among them, θ G is the parameter of the generator, η G is the learning rate of the generator, is the loss function of the generator, L G is the gradient of the generator loss function with respect to the parameter θ G of the generator, m is the batch size, D(G(z k )) is the discrimination probability of the second false sample output by the discriminator, C(G(z k )) is the discrimination probability of the second false sample output by the classifier, and β is the training hyperparameter.

6. The training method of the ISAR open set recognition network based on the generative discriminative model according to claim 1, wherein The classification loss is: Among them, L c is the classification loss function, and p(y = k i |x i ) represents the probability that the sample x i belongs to the category k i , and N represents the number of known categories; The open boundary constraint is: Among them, L o is an open boundary constraint function, f(x i ) is the feature of the i-th data sample in the batch, p i is the corresponding prototype, d(·) represents the distance distribution, R is a learnable boundary parameter, ε = 0.001, G(z o ) is the third fake sample output by the input generator for the noise vector batch ; The constraint condition composed of the classification loss and the open boundary constraint is: L t = L c + λL o where λ is a weight parameter.

7. An ISAR open set recognition method based on generative discriminant, characterized in that, It includes the steps: S1. Input a test set including known classes and unknown classes into the trained classifier to obtain the closed-set classification probability, and the trained classifier is trained by the training method described in any one of claims 1-6; S2. Fit the extreme value cumulative distribution model for each category according to the Weibull cumulative distribution function and obtain the corresponding cumulative distribution function value F under the known category weibull (λ i , k i ), where λ i is the scale parameter of the i-th category, and k i is the shape parameter of the i-th category; S3. Use the cumulative distribution function value under the known class to correct the closed-set classification probability of the test sample to obtain a corrected probability value; S4. Calculate the probability that the sample to be tested is determined as an out-of-distribution sample; S5. Combine the corrected probability value and the probability of the out-of-distribution sample into a corrected score value, and use the maximum score value in the score values as the recognition result in the open-set scenario.

8. The discriminative generative-based ISAR open set recognition method according to claim 7, characterized in that The closed-set classification probability is: Among them, is the sample to be tested, are the features output by the trained classifier, and p i is the preliminary classification score that the sample to be tested belongs to the i-th known class, and N k is the number of known class categories; The corrected probability value is: where p i ' is the probability after correction for the i-th category, which is the value of the cumulative distribution function of the Weibull distribution; The probability of the out-of-distribution sample is: where p u is the probability of out-of-distribution samples; The corrected score value is: The recognition result in the open-set scenario is: y = argmax(p).

9. The discriminative generative-based ISAR open set recognition method according to claim 7, characterized in that Step S2 includes: S21. Input the training samples of the i-th class in the known-class training set into the trained classifier to obtain the activation vectors of the training samples of the i-th class, and retain the activation vectors correctly classified as the i-th class samples by the classifier to obtain an activation vector set, and calculate the mean value of the activation vector set as the centroid of the target class; S22. Calculate the distance from each activation vector in the activation vector set to the centroid to obtain a distance set; S23. Fit the maximum value distribution in the distance set, and input the maximum value distribution into the Weibull distribution model for fitting to obtain the cumulative distribution function value under the known class.

10. The discriminative generative-based ISAR open set recognition method according to claim 9, characterized in that The cumulative distribution function of the Weibull distribution model is: where k is the shape parameter, λ is the scale parameter, and x is the fitted maximum value distribution.

Images