Generalized sar target recognition method and system based on end-to-end structure

Abstract

The application discloses a kind of generalization SAR target identification method and system based on end-to-end structure, obtains the dataset containing label and label-free SAR image sample and expands it, forms training sample;Based on multimodal remote sensing image self-supervised pre-training weight fine-tuning target SAR image data on feature representation;Combining representation learning loss, entropy normalization loss and improved divergence loss, the new class discovery model is trained end-to-end until convergence.Processing SAR image data to be identified in generalization SAR target new class discovery framework, target recognition result can be obtained.The application not only can support the identification of known category, but also can efficiently mine unknown new class target and determine its specific category, effectively avoid the confusion of recognition model to new class target and the catastrophic forgetting of old class target, enhance the generalization ability of SAR target recognition model in uncertain open scene.

Description

Technical Field

[0001] This invention relates to the field of target recognition, and in particular to a generalized SAR target recognition method and system based on an end-to-end structure. Background Technology

[0002] In the field of remote sensing, Synthetic Aperture Radar (SAR), as a coherent imaging radar system operating in the microwave band, achieves high resolution by utilizing range-direction pulse compression and azimuth-direction aperture accumulation, and enables high-resolution imaging of large observation areas via airborne, spaceborne, or missile-borne platforms. Benefiting from its all-weather, all-day observation capabilities, SAR automatic target identification plays a crucial role in battlefield reconnaissance and situational awareness. With the diversification of SAR imaging systems and their commercialization, the scale and quality of SAR image data available have significantly improved. Traditional manual interpretation methods are resource-intensive and difficult to meet the demands of real-time, efficient, dynamic, and accurate battlefield surveillance. With the development of deep learning technology, data-driven target identification methods have made breakthrough progress, making the automation and intelligence of SAR intelligent interpretation systems an inevitable trend. As a fundamental problem in SAR image analysis and understanding, SAR image target identification is currently a research hotspot.

[0003] In recent years, thanks to the development of deep learning technology and the increasing computing power, deep learning-based SAR target recognition methods have been successfully applied to many real-world scenarios. However, the contradiction between the processing paradigm of existing recognition algorithms and the needs of real-world open scenarios is prominent, severely restricting the widespread and efficient application of these algorithms in military scenarios. Existing algorithms rely on a large amount of labeled data and are trained and tested under closed-set conditions, meaning that targets share the same label space during training and testing, and the test category is included in the training set category. Most models based on closed-set testing can achieve very high recognition accuracy through supervised learning. However, the continuous expansion of target databases in real-world scenarios brings new challenges to recognition. To ensure good target recognition performance for all categories, data collection and mapping work needs to be carried out again, resulting in extremely high labor costs. More importantly, in military fields such as military reconnaissance, area surveillance, and precision strikes, due to the non-cooperative and low-exposure characteristics of enemy targets, obtaining a sufficient number of high-quality labeled samples becomes more difficult or even impossible.

[0004] Most researchers address the poor generalization ability of models in open environments through novel class discovery algorithms. These algorithms discover or cluster new categories in unlabeled data with the aid of known target information, thereby enabling target recognition when new categories appear. However, most novel class discovery algorithms are based on a two-stage approach: representation learning and semi-supervised clustering. The first stage trains labeled class samples, and the second stage generates pseudo-labels on unlabeled images before supervising the learning of new categories. These methods lack end-to-end training capabilities and focus solely on the clustering effect of new categories, failing to retain sufficient knowledge of known categories.

[0005] Dai et al. used deep embedded clustering to address the problems of open-set target identification and new class discovery for SAR ships. In the open-set target identification stage, they simultaneously used known and unknown class targets for unsupervised training, and then fine-tuned the model using known class labeled data to discover new categories. However, this method is limited in practical scenarios where data containing unknown class targets is unavailable during the open-set identification training stage, thus hindering unsupervised training and limiting its application. Huang et al. improved the data augmentation method during model pre-training in the joint classification method, effectively enhancing the discriminability between known and unknown classes, considering separability across all categories, and effectively avoiding classifying unlabeled samples into the same category.

[0006] For the problem of target recognition in open environments with potential novel classes, existing research on novel class discovery relies on the assumption that unlabeled test data belongs entirely to a completely new and unseen class. However, these existing novel class discovery methods are mainly designed for scenarios where the test set contains only novel classes, and are insufficient to meet the need for effective identification of test sets containing both known and novel class samples in real-world open and dynamically changing scenarios.

[0007] In real-world scenarios, the target to be identified is not limited to an unknown category; its category can be any category in an ever-expanding sample library—that is, unlabeled data simultaneously contains both known and new categories. To meet these more practical application needs, proposing a universally applicable method for identifying new SAR target categories is a pressing issue. First, compared to optical datasets with ample and high-quality data, SAR images suffer from insufficient data volume, complex backgrounds, and blurred contours. During training, the model struggles to extract highly generalizable discriminative features, and is prone to overfitting to known target categories when identifying new target categories. Furthermore, generating high-quality pseudo-labels on unlabeled datasets is crucial for improving model performance in conventional new category discovery methods. Extracting effective knowledge of existing categories from SAR images and transferring it to new categories also affects the accuracy of new category discovery. The complex scattering characteristics of SAR images mean that measuring pseudo-label quality in two-stage new category discovery methods may lead to low network prediction reliability. These challenges further increase the difficulty of ordinary SAR target recognition methods in identifying unknown targets and continuously recognizing known targets. In summary, existing methods for discovering new target classes cannot effectively address the problem of simultaneously identifying old and new target classes in SAR images in open scenarios, which can easily lead to a lack of continuous update and iteration capabilities in practical deployments. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a generalized SAR target recognition method and system based on an end-to-end structure to address the shortcomings of the existing technology, avoid confusion of the recognition model with new types of targets and catastrophic forgetting of old types of targets, and enhance the generalization ability of the SAR target recognition model in uncertain open scenarios.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a generalized SAR target identification method based on an end-to-end structure, comprising the following steps:

[0010] S1. Obtain a SAR image dataset consisting of labeled and unlabeled samples to form a training dataset;

[0011] S2. Extract the discriminative representations of the training set;

[0012] S3. Use the discriminative representation as input to the generalized new class target recognition model, train the generalized new class target recognition model, and obtain the final target recognition model.

[0013] This invention not only supports the recognition of known categories, but also efficiently mines unknown new target categories and determines their specific categories. It effectively avoids the confusion of the recognition model with new target categories and the catastrophic forgetting of old target categories, and enhances the generalization ability of the SAR target recognition algorithm in uncertain open scenarios.

[0014] The training set includes several labeled SAR training images containing known classes, and several unlabeled SAR training images containing new classes.

[0015] The specific implementation process of step S2 includes: using the training set as input to the pre-trained model to obtain the discriminative representation.

[0016] The pre-trained model was obtained by training a multimodal remote sensing image dataset.

[0017] The pre-trained model is the Vision Transformer network.

[0018] In step S3, the loss function of the generalized new class target recognition model is expressed as:

[0019] L total =α(L rep +L cls )+βL entropy ;

[0020] Where α and β represent the proportions of each part of the loss function, and the cross-entropy loss is... This indicates that the conditions shown in parentheses are met. This represents unlabeled samples selected based on three constraints: the sample is unlabeled, the energy value of the sample exceeds a certain threshold, and the sample belongs to a known category. i ) represents the image x i The predicted probability, τ0 is the preset temperature parameter, H(p) = -∑ k p (k) logp (k) p represents the entropy normalization function. (k) The predicted probability is obtained by processing the output of the generalized new class target recognition model using the softmax function. The value of k ranges from 1 to n, where n represents the vector dimension of the output of the generalized new class target recognition model, and the loss is represented by k = k * n * n * n * n. λ is a coefficient that controls the proportion of the representation loss on unlabeled and labeled samples. The classification loss on labeled samples is... l(y i ,p i ) is y i and p i The cross-entropy loss function between p i For a certain viewpoint image x i soft tags, y i For image x i The corresponding ground truth label, B represents the number of training batches, and the classification loss on all input training data is... D skew_KL(p,q) is the skew divergence loss, where p and q are the predicted probabilities of two different image samples. The classification loss on unlabeled samples is... q represents the cross-entropy loss between two labels. i ′ is the image x i A pseudo-tag on '.

[0021] After step S3, the following also includes:

[0022] Acquire a test SAR image to be identified, input the SAR image into a target recognition model, and classify and identify the target categories in the SAR image.

[0023] As an inventive concept, the present invention also provides a generalized SAR target recognition system based on an end-to-end architecture, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the above method.

[0024] As an inventive concept, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon; when the computer program / instructions are executed by a processor, they implement the steps of the above-described method.

[0025] As an inventive concept, the present invention also provides a computer program product, including a computer program / instructions; when the computer program / instructions are executed by a processor, they implement the steps of the above-described method.

[0026] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention uses a model trained on optical remote sensing and SAR images based on a self-supervised visual transformer as a pre-trained model, and then trains it on all data containing both labeled and unlabeled samples, enabling the model to learn stable discriminative feature representations of SAR images. Without introducing additional parameters or changing the original network structure, a local entropy normalization module and a dual-view consistency discrimination module are introduced, allowing the network to automatically filter known samples with high confidence and energy values. The improved consistency constraint loss further enhances the accuracy of sample selection, and the combined effect of these two modules further avoids catastrophic forgetting. The training of the proposed generalized novel class discovery framework includes three loss components: representation learning, entropy normalization of partially known samples, and consistency discrimination. Finally, this generalized novel class discovery model is used to identify target categories in measured unlabeled SAR images. This invention eliminates the need for multi-stage training and employs an end-to-end generalized new class discovery framework. It simultaneously achieves clustering and classification of unknown target classes, balancing recognition accuracy across known old classes and unknown new classes. This overcomes the limitation of other new class discovery algorithms that are prone to forgetting known target classes. Its recognition performance on SAR image vehicle target datasets significantly outperforms conventionally trained classification methods, improving the accuracy of identifying both old and new target classes in generalized SAR target new class discovery tasks. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the generalized SAR target new class discovery method provided in this embodiment;

[0028] Figure 2 This is a structural framework diagram of the open-scene generalized SAR target new class discovery and recognition model based on an end-to-end structure, as presented in this invention.

[0029] Figure 3 This is a schematic diagram comparing the multi-stage and single-stage generalized new class discovery methods in this invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1

[0032] Training data-driven SAR target recognition models relies on a large amount of labeled data, and the actual performance of the model depends on whether the sample category belongs to the training label set space. However, in open scenarios, the category of the test sample is unknown; it may belong to an old category that has already been trained or a new category that has not been trained. Re-collecting samples to train a new model is not only costly and dependent on expert interpretation, but samples of non-cooperative enemy targets are almost impossible to obtain. This limits the practical application of current high-performance SAR target recognition models in open scenarios.

[0033] To address the above issues, in one embodiment, such as Figure 1 As shown, this application proposes a generalized SAR target identification method and system based on an end-to-end architecture, including the following steps:

[0034] Step S1: Obtain the SAR image training dataset and expand it to form training samples; the training set includes several labeled SAR training images with known classes and several unlabeled SAR training images with new classes.

[0035] Specifically, in typical new class discovery methods, a two-stage learning approach first pre-trains on a labeled set, then transfers the learned inter-class knowledge, using this inter-class knowledge as prior information to perform unsupervised clustering on unlabeled data. Unlike the two-stage training scheme, this embodiment employs a single-stage, end-to-end training method, avoiding the problem of incomplete sharing of knowledge between new and old classes present in the two-stage approach. This results in stronger robustness and more stable training.

[0036] Step S2: The SAR image training data obtained in Step 1 is used as network input for robust feature extraction. Labeled and unlabeled samples are input into the network for representation learning. Borrowing from the DINO approach, pre-trained weights obtained from training on a multimodal remote sensing dataset are used to initialize the recognition network, which is based on the Vision Transformer model. On the SAR image training dataset, a simple cross-entropy loss is used for representation learning optimization of labeled samples, while a self-distillation strategy is used for optimization of unlabeled samples, to maximize the similarity of representations corresponding to augmented views of the same input image.

[0037] Specifically, since the DINO model is a high-performance nearest neighbor classifier, it demonstrates its ability to cluster features effectively in the feature space. Secondly, existing literature demonstrates that self-supervised visual transformers possess the ability to learn salient features of target objects without manual annotation. To obtain a good feature representation consistent with image characteristics on existing SAR image data, the pre-trained weights obtained by training a self-supervised visual transformer in the literature (Y. Wang, C.M. Albrecht and XX. Zhu, "Self-Supervised Vision Transformers for Joint SAR-Optical Representation Learning," IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium, Kuala Lumpur, Malaysia, 2022, pp. 139-142, doi:10.1109 / IGARSS46834.2022.9883983.) are used as the initialization model for the proposed method. This model is trained on the large-scale multi-label scene classification dataset BigEarthNet-MM. Because it contains multimodal data from optical remote sensing and SAR images, the trained model can fully extract features from single-modal or multimodal data. Experiments have shown that the model pre-trained with multimodal data can be well transferred to single-modal tasks and outperforms supervised learning methods under small sample conditions. Therefore, based on this pre-trained model, the representation on the SAR image training data shown in step one is further fine-tuned. Supervised contrastive learning is performed on labeled samples, and unsupervised contrastive learning is performed simultaneously on labeled and unlabeled samples.

[0038] When selecting positive sample pairs in unsupervised contrastive learning, it's important to consider that randomly cropped image patches may contain background or shadows. Using image pairs containing these elements as positive samples can negatively impact model training. If the augmented image does not contain any target, or if the region containing the target is smaller than a certain threshold, the data augmentation is deemed invalid, and the image patch needs to be re-cropped.

[0039] Images of the same image from different perspectives are defined as x. i and x i The unsupervised contrast loss can be expressed as:

[0040]

[0041] Where B represents the batch size for model training, z i =g(f(x) i)) represents the features extracted by the backbone network f and the mapping head g, z n Represents any sample x other than the i-th sample. n The features mentioned above are processed using L2 regularized normalization, τ u This represents the temperature parameter. The image x at a certain viewpoint... i The soft tag is denoted as p i Another perspective image x i The pseudo-label on ' is denoted as q. i If ′, then the classification loss on unlabeled data can be expressed as:

[0042]

[0043] in This represents the cross-entropy loss between two labels. Similarly, the supervised contrastive loss can be derived.

[0044] Unlike unsupervised contrastive loss, since the class labels are known, positive sample pairs can be constructed based on the label information. This loss can be expressed as:

[0045]

[0046] Among them B l N represents the batch size of labeled samples in the same training batch. i Indicates the label and x in the same batch i With the same number of samples, τ c This represents the temperature parameter, which is set to 1 by default. (Image x...) i The corresponding truth label is denoted as y. i The classification loss on labeled data can then be expressed as:

[0047]

[0048] In summary, the overall representation learning function can be expressed as:

[0049]

[0050] Step S3 involves iteratively training the generalized novel class discovery framework based on the SAR image representations obtained in Step S2. Considering that in the two-stage novel class discovery method, the model tends to favor known classes when predicting new and old classes, resulting in an imbalanced distribution of generated pseudo-labels, an entropy normalization-based constraint is designed to reduce prediction bias and force the model to generate a more uniform label distribution across all classes, thereby improving the quality of generated pseudo-labels on unlabeled samples. The generalized novel class discovery target recognition model is trained until all loss functions converge, resulting in a generalized novel class discovery model with good generalization ability for both new and old classes. The following sections describe the local entropy normalization module and the dual-view consistency discrimination module.

[0051] Local entropy normalization module:

[0052] Specifically, some new class discovery methods employ class average entropy regularization constraints to make the model focus on improving the accuracy of new class recognition. However, this approach sacrifices the accuracy of known classes to some extent, leading to catastrophic forgetting. To preserve knowledge of old classes as much as possible, such as... Figure 2 As shown in the dashed box, this scheme first determines potential known samples based on the network's classification head predictions. Then, it uses an information entropy function to constrain these samples, making their predicted values ​​approximately uniformly distributed. This avoids misclassifying known class samples into new categories and retains knowledge related to known categories during the learning process of new categories, facilitating the transfer of knowledge about known categories. The specific implementation of this module includes three steps: 1) selecting unlabeled samples; 2) using a threshold to filter samples with high energy values ​​among the unlabeled samples; 3) further filtering samples based on whether the predicted values ​​belong to known categories; and 4) using entropy normalization loss to ensure the stability of the selected known samples.

[0053] The training data is defined as D = {(x i ,y i )}, which simultaneously includes x i Labeled and unlabeled samples are represented using a binary mask vector. This indicates whether the sample is a labeled sample. If M=0, it indicates that the sample is an unlabeled sample.

[0054] Because the exponential form of the Softmax function used in deep networks for classification is not proportional to the data distribution, it is difficult to represent the input data well. It is prone to generating extremely high confidence levels for out-of-distribution samples, thus misclassifying samples into known categories and interfering with model predictions. Energy-based models essentially map a point in the input space to a real value through an energy function, thereby describing the system state; the lower the energy value, the more stable the system state. Considering that the energy function often has a linear relationship with the data distribution, energy-based models can effectively estimate and reflect the energy value of sample features, thus characterizing the degree of anomaly of new samples relative to the training set. Analogous to classification models and energy models, this approach can define the energy function as:

[0055]

[0056] Where f represents the feature extractor, i.e. Figure 2 The encoder shown in the blue box, g i Let represent the logit output of the i-th category. The above formula can convert the category output value in the recognition task into an energy function. Given that step S2 has combined pre-trained weights and supervised and unsupervised contrastive learning to obtain a SAR image representation with strong discriminative power, a significant difference can also be shown in the energy values ​​of samples of known old classes and unknown new classes.

[0057] If the energy distribution of a sample is highly similar to the energy distribution of known classes, it is assigned to the nearest known class. This process simulates the energy values ​​of known and unknown classes using a Weibull distribution, and uses this as a basis to determine whether the target originates from a known or unknown class. The method of determining potential known class samples based on energy distribution can effectively eliminate new class samples from unlabeled data, reduce the impact of overfitting in the sample space by deep convolutional networks, and improve the accuracy of generalized new class discovery to some extent.

[0058] After analyzing the unlabeled samples using the available energy distribution, unlabeled samples with higher confidence levels were selected again, and then... Label these samples. For unlabeled sample x... i The above process can be represented as:

[0059]

[0060] in This indicates that the condition shown in parentheses is met, where δ is the threshold. This applies to samples x in batch size b. i The corresponding latent class label is defined as y = [y1, y2, ..., y]. i ]∈{1,2,...,K},y i From the predicted probability p i The index corresponding to the maximum value is determined by the following formula:

[0061]

[0062] Where B represents the number of training batches. Based on the three constraints of whether the sample is unlabeled, whether the energy value exceeds a certain threshold, and whether it belongs to a known category, the knowledge of the old class samples that need to be retained can be determined. Similarly, using... The samples that meet the above conditions are labeled.

[0063]

[0064] in This represents the set of labels for samples of known categories. To ensure the stability of learning known category information during training, information entropy normalization loss is used to constrain the samples:

[0065]

[0066] Where τ0 is the preset temperature parameter, H(p) = -∑ k p (k) logp (k) This represents the entropy normalization function.

[0067] Dual-view consistency discrimination module:

[0068] In the aforementioned local entropy normalization module, when determining potential known samples among unlabeled samples based on constraints, this scheme employs a dual-view consistency discrimination module to mitigate the phenomenon of the network excessively favoring known class samples and neglecting the mining of new class knowledge, thus leading to misclassification of unlabeled samples. If the confidence scores of the sample corresponding to the enhanced images under different views differ significantly, the uncertainty of whether the sample belongs to the known class is high. Therefore, the original sample can only be determined to belong to the known class when the confidence scores of both enhanced images are greater than a certain threshold.

[0069] KL divergence loss is typically used to constrain the predicted output values ​​of images from different viewpoints. Here, to enhance the stability of gradient changes and reduce estimation errors, an improved KL loss, namely the tilted divergence loss, is chosen to measure the consistency of the distributions corresponding to the two views of the same image. Specifically, the parameter α controls the mixing ratio of the two distributions p and q. The improved loss function between the two probability distributions can be expressed as:

[0070] D skew_KL (p,q)=D KL (p,αp+(1-α)q)

[0071] Enhance sample x from different perspectives within the unlabeled sample set. i and x j Substituting into the above equation, we define its corresponding probability prediction distribution as p. i and pj The dual-view consistency loss can be expressed by the following formula:

[0072]

[0073] Among them B u This indicates the batch size of the unlabeled samples.

[0074] Adding this constraint to the classification loss for unlabeled data enhances the reliability of mining potentially known samples. Therefore, the classification loss on all input training data can be expressed as:

[0075]

[0076] In summary, using the improved consistency discriminant loss can ensure that the model can effectively handle unknown new classes. While maintaining the classification performance of known classes, it can use the knowledge of known classes to assist in the discovery of new classes, thereby achieving the classification of known classes and the clustering of new classes.

[0077] By incorporating the local entropy normalization constraint and the dual-view consistency discrimination constraint corresponding to the two modules in step 3 into the baseline classification loss, the training loss function of the proposed end-to-end generalized new class discovery method can be obtained. The overall loss function can be expressed as:

[0078] L total =α(L rep +L cls )+βL entropy

[0079] After training until all loss functions converge, a generalized SAR target new class discovery model with strong generalization ability and robustness is obtained.

[0080] Step S4: Obtain the SAR image to be identified, and input the SAR image into the trained generalized new class discovery model to classify and identify the target category in the SAR image.

[0081] Specifically, after training the model, the model's recognition accuracy can be evaluated on test samples. For the selected dataset, a generalized new class discovery model is trained on dataset D, which contains both labeled and unlabeled samples. During testing, the model's prediction results are used to evaluate the accuracy. and the actual true value of the target y i Measure the accuracy of clustering.

[0082]

[0083] Here Represents the set of all permutations of class labels in unlabeled samples. This represents the number of unlabeled samples. The above formula includes the process of matching clusters with the true labels and taking the highest accuracy, specifically implemented using the Hungarian algorithm. This invention evaluates the recognition performance for all categories, including performance on the old class set (category belongs to...). Performance on instances of the new class set (category belongs to) and performance on new class sets (category belongs to) (Examples).

[0084] The aforementioned generalized SAR target recognition method, apparatus, and computer equipment based on an end-to-end architecture include: using a model trained on optical remote sensing and SAR images based on a self-supervised visual transformer as a pre-trained model, and then training it on all data containing both labeled and unlabeled samples, enabling the model to learn stable discriminative feature representations of SAR images. Without introducing additional parameters or altering the original network structure, a local entropy normalization module and a dual-view consistency discrimination module are introduced, allowing the network to automatically filter known samples with high confidence and energy values. An improved consistency constraint loss further enhances the accuracy of sample selection, and the combined effect of these two modules further avoids catastrophic forgetting. Training the proposed generalized novel class discovery framework includes three loss components: representation learning, entropy normalization of partially known samples, and consistency discrimination. Finally, this generalized novel class discovery model is used to identify target categories in measured unlabeled SAR images. This method eliminates the need for multi-stage training and employs an end-to-end generalized new class target discovery and recognition framework. It simultaneously achieves clustering and classification of unknown target classes, balancing recognition accuracy across known old classes and unknown new classes, and eliminating the limitation of other new class discovery algorithms that are prone to forgetting known target classes. Its recognition performance on SAR image vehicle target datasets significantly outperforms conventionally trained classification and recognition methods, improving the accuracy of identifying both old and new target classes in generalized SAR target new class discovery tasks.

[0085] Example 2

[0086] Embodiment 2 of the present invention provides a system corresponding to Embodiment 1 above, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program in the memory to implement the steps of the method of Embodiment 1 above.

[0087] In some implementations, the memory may be high-speed random access memory (RAM), and may also include non-volatile memory, such as at least one disk storage device.

[0088] In other implementations, the processor can be any type of general-purpose processor, such as a central processing unit (CPU) or a digital signal processor (DSP), and there is no limitation here.

[0089] Example 3

[0090] Embodiment 3 of the present invention provides a computer-readable storage medium corresponding to Embodiment 1 above, on which a computer program / instructions are stored. When the computer program / instructions are executed by a processor, they implement the steps of the method of Embodiment 1 above.

[0091] A computer-readable storage medium can be a tangible device that holds and stores instructions for use by an instruction execution device. A computer-readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof.

[0092] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0096] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A universal SAR target recognition method based on an end-to-end structure, characterized in that, The method comprises the following steps: S1, acquiring a SAR image dataset composed of labeled and unlabeled samples to form a training dataset; S2, extracting a discriminative representation of the training dataset; S3, taking the discriminative representation as an input of a universal new class discovery target recognition model, training the universal new class discovery target recognition model, and obtaining a final target recognition model; The loss function of the universal new class discovery target recognition model is represented as: ； wherein, denotes the proportion of each part of the loss function, the cross-entropy loss is , denotes that if the condition shown in the parentheses is met, the value is 1, otherwise the value is 0, denotes that the sample meets three restriction conditions, the three restriction conditions refer to belonging to an unlabeled sample, the energy value corresponding to the sample exceeds a certain threshold, and the sample belongs to a known class, denotes the predicted probability of the image, and T is a preset temperature parameter, denotes an entropy normalization function, denotes the predicted probability obtained by processing the output of the general new class discovery target recognition model by a softmax function, the value range of k is 1-n, n denotes the vector dimension of the output of the general new class discovery target recognition model, the loss is represented by L, and is a proportionality coefficient for controlling the loss on the unlabeled sample and the labeled sample, the classification loss on the labeled sample is , is the cross-entropy loss function between and , is the soft label of a certain view image , is the true value label corresponding to the image , B denotes the number of training batches, and the classification loss on all input training data is , is a tilted divergence loss, p and q are predicted probability values corresponding to image samples of the same sample in different views, and the classification loss on the unlabeled sample is , denotes the cross-entropy loss between two labels, is the pseudo label on the image , the image , and the image are image samples of the same sample in different views.

2. The end-to-end structure based generalized SAR target recognition method according to claim 1, characterized in that, The training dataset comprises a plurality of labeled SAR training images containing known classes and a plurality of unlabeled SAR training images containing new classes.

3. The end-to-end structure based generalized SAR target recognition method according to claim 1, wherein, The specific implementation process of step S2 comprises: Taking the training dataset as an input of a pre-training model to obtain the discriminative representation.

4. The end-to-end structure based generalized SAR target recognition method according to claim 3, characterized in that, The pre-training model is trained by a multi-modal remote sensing image dataset.

5. The end-to-end structure based generalized SAR target recognition method according to claim 3, characterized in that, The pre-training model is a Vision Transformer network.

6. The end-to-end structure based generalized SAR target recognition method according to claim 1, characterized in that, After step S3, the method further comprises: Acquiring a test SAR image to be recognized, inputting the SAR image into the target recognition model, and classifying and recognizing the target class in the SAR image.

7. A generalization SAR target recognition system based on end-to-end structure, comprising a memory, a processor and a computer program stored on the memory; characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1-6.

8. A computer readable storage medium having stored thereon computer programs / instructions; characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of any one of claims 1-6.

9. A computer program product comprising computer programs / instructions; characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of any one of claims 1-6.

Images