SAR image target recognition method and device based on small sample incremental learning

Abstract

The application relates to a SAR image target recognition method and device based on small sample incremental learning. In the base class training, a multi-class target confusion training set and a single-class target rotation training set are constructed based on SAR sample images in a base class data set to increase the target classes of training, so that the trained feature extractor can agilely extract and represent unknown target features in a real recognition scene. In the incremental class training, a plurality of small sample training sets are constructed by performing multiple random sampling according to N classes and K samples based on an incremental class data set. During the incremental class training, the feature extractor is decoupled into a static feature extraction part and a dynamic updating part, and the adjustable parameters of the feature extractor and the classifier are optimized by using class imprint cross-entropy loss. In particular, in each batch of incremental class training, the mean of the training samples obtained by the current feature extractor is used as the initial value of the class imprint in the class imprint cross-entropy loss. The method can solve the two problems of catastrophic forgetting and overfitting in small sample training, so that the SAR target recognition method can be applied to an open dynamic environment.

Description

Technical Field

[0001] This application relates to the field of SAR target recognition technology, and in particular to a SAR image target recognition method and apparatus based on small sample incremental learning. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave detector capable of observing targets of interest around the clock and in all weather conditions. It is widely used in many fields, such as disaster early warning, geological exploration, and national defense. As a fundamental application of SAR image interpretation, Automatic Target Recognition (SAR ATR) aims to accurately and quickly locate and identify targets based on their characteristics within SAR images. Due to the rich information it provides, it has received widespread and sustained attention from both academia and industry.

[0003] In recent years, with the development of deep learning (DL) technology, SAR ATR algorithms based on DL technology have also made significant progress. Undoubtedly, with their powerful feature extraction and discrimination capabilities, DL-based algorithms have achieved significant advantages over traditional algorithms in many recognition tasks. However, most of these advanced algorithms focus only on target recognition in ideal, closed environments, typically assuming that the sample categories seen in the prediction phase are consistent with those in the initial training set and do not change with differences in the deployment environment. Clearly, this is highly incompatible with the open and dynamic real-world application scenarios of SAR ATR. With the increasing maturity of SAR imaging technology, the number of high-resolution SAR images is growing rapidly, and new target categories are emerging more frequently. However, in most non-cooperative scenarios, the number of observable samples is extremely limited due to constraints such as the observation environment, imaging platform, and the target itself, resulting in more confusing class discrimination information. In particular, due to limitations in storage and computing resources, data security, and privacy, storing and repeatedly accessing historical data is extremely inefficient and even prohibited.

[0004] Therefore, improving the ability of the DL-based SAR ATR algorithm to continuously learn and discriminate new categories from small samples, i.e., the ability of Few-Shot Class-Incremental Learning (FSCIL), is the key to its agile adaptation in open and dynamic environments. Summary of the Invention

[0005] Therefore, it is necessary to provide a SAR image target recognition method and apparatus based on small-sample incremental learning that can continuously learn and judge new target categories in a small sample in an open and dynamic environment, in order to address the above-mentioned technical problems.

[0006] A SAR image target recognition method based on few-sample incremental learning, the method comprising:

[0007] Obtain the base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images.

[0008] Based on the SAR sample images in the base class dataset, a multi-class target confusion training set and a single-class target rotation training set are constructed.

[0009] The training samples from the multi-class target confusion training set and the single-class target rotation training set, along with the training samples from the original training set, are input into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples.

[0010] Based on the incremental dataset, multiple small sample training sets are constructed by randomly sampling multiple times according to N categories, each containing K samples;

[0011] Multiple small sample training sets are used to sequentially perform multiple rounds of incremental class training on the feature extractor and classifier that have completed base class training. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and the parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. At the beginning of each round of incremental class training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial class imprint value in the class imprint cross-entropy loss.

[0012] Acquire an image of the SAR target to be identified, and input the image of the SAR target to be identified into a feature extractor and a classifier that have completed additive training to achieve target recognition.

[0013] In one embodiment, constructing a multi-class target confusion training set based on SAR sample images in the base class dataset includes:

[0014] According to the property scattering center theory, targets in SAR sample images are all composed of multiple strong scattering components;

[0015] Based on SAR sample images of different target categories, different strong scattering components are randomly combined to construct virtual target SAR sample images of new target categories;

[0016] The multi-target confusion training set is constructed based on multiple virtual target SAR sample images.

[0017] In one embodiment, the shape of the target component that has strong scattering in the SAR sample image includes dihedral, trihedral, cylindrical, and spherical.

[0018] In one embodiment, the random combination of different strong scattering components based on SAR sample images of different target categories is performed using the following formula:

[0019]

[0020] In the above formula, λ represents the coefficient controlling the overlap between real and virtual categories, and Beta(a,a) represents the sampling distribution. Represents the target scattering characteristics of the generated class. x′ represents the scattering characteristics of a selected real-world target. j This represents the target scattering characteristics of another real class selected.

[0021] In one embodiment, constructing a single-class target rotation training set based on SAR sample images in the base class dataset includes:

[0022] The target in the SAR sample image is rotated by a preset angle to obtain a rotated target SAR sample image;

[0023] The single-target rotation training set is constructed based on multiple rotating target SAR sample images.

[0024] In one embodiment, the class-imprint cross-entropy loss is expressed as:

[0025]

[0026]

[0027] In the above formula, This represents the feature extractor for the t-th learning stage, where the superscript t indicates the t-th learning stage, and x... i Let y represent the training sample. i D represents the truth label. t Let w represent the target training set for the t-th learning stage. c Represents a type of imprint.

[0028] In one embodiment, during incremental training, the feature extractor parameters can be decoupled and represented as:

[0029]

[0030] In the above formula, This refers to the static feature extraction portion. This indicates the part where parameters are dynamically updated.

[0031] In one embodiment, the feature extractor and classifier are constructed as a neural network, and this neural network is used as a SAR target recognition network.

[0032] In one embodiment, after acquiring the SAR target image to be identified and inputting the SAR target image to be identified into a feature extractor and a classifier that have completed additive training to achieve target recognition, the method further includes:

[0033] If the target in the SAR target image to be identified is of an unknown category, then the training samples in the small sample training set are used to perform incremental online training on the SAR target recognition network.

[0034] This application also provides a SAR image target recognition device based on few-sample incremental learning, the device comprising:

[0035] The dataset acquisition module is used to acquire base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images.

[0036] The base class training set construction module is used to construct a multi-class target confusion training set and a single-class target rotation training set based on SAR sample images in the base class dataset.

[0037] The base class training module is used to input the training samples from the multi-class target confusion training set and the single-class target rotation training set, as well as the training samples from the original training set, into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples.

[0038] The small sample training set construction module is used to construct multiple small sample training sets based on the incremental class dataset, according to N categories, with each category containing K samples, through multiple random samplings.

[0039] The incremental training module is used to perform multiple rounds of incremental training on the feature extractor and classifier that have completed base class training using multiple small sample training sets. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and the parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. At the beginning of each round of incremental training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial class imprint value in the class imprint cross-entropy loss.

[0040] The SAR image target recognition module is used to acquire SAR target images to be recognized, and input the SAR target images to be recognized into a feature extractor and a classifier that have completed additive training to achieve target recognition.

[0041] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0042] Obtain the base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images.

[0043] Based on the SAR sample images in the base class dataset, a multi-class target confusion training set and a single-class target rotation training set are constructed.

[0044] The training samples from the multi-class target confusion training set and the single-class target rotation training set, along with the training samples from the original training set, are input into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples.

[0045] Based on the incremental dataset, multiple small sample training sets are constructed by randomly sampling multiple times according to N categories, each containing K samples;

[0046] Multiple small sample training sets are used to sequentially perform multiple rounds of incremental class training on the feature extractor and classifier that have completed base class training. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and the parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. At the beginning of each round of incremental class training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial class imprint value in the class imprint cross-entropy loss.

[0047] Acquire an image of the SAR target to be identified, and input the image of the SAR target to be identified into a feature extractor and a classifier that have completed additive training to achieve target recognition.

[0048] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0049] Obtain the base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images.

[0050] Based on the SAR sample images in the base class dataset, a multi-class target confusion training set and a single-class target rotation training set are constructed.

[0051] The training samples from the multi-class target confusion training set and the single-class target rotation training set, along with the training samples from the original training set, are input into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples.

[0052] Based on the incremental dataset, multiple small sample training sets are constructed by randomly sampling multiple times according to N categories, each containing K samples;

[0053] Multiple small sample training sets are used to sequentially perform multiple rounds of incremental class training on the feature extractor and classifier that have completed base class training. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and the parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. At the beginning of each round of incremental class training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial class imprint value in the class imprint cross-entropy loss.

[0054] Acquire an image of the SAR target to be identified, and input the image of the SAR target to be identified into a feature extractor and a classifier that have completed additive training to achieve target recognition.

[0055] The aforementioned SAR image target recognition method and apparatus based on few-sample incremental learning increases the number of target categories during base class training by constructing a multi-class target confusion training set and a single-class target rotation training set based on SAR sample images in the base class dataset. This allows the trained feature extractor to construct feature spaces for various target categories. During incremental class training, multiple few-sample training sets are constructed by randomly sampling K samples from N categories based on the incremental class dataset. During incremental class training, the feature extractor is decoupled into a static feature extraction part and a dynamic update part, and the adjustable parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. Specifically, during each batch of incremental class training, the mean of the training samples obtained by the current feature extractor is used as the initial value of the class imprint in the class imprint cross-entropy loss. This method can solve the two major problems of catastrophic forgetting and overfitting encountered during few-sample training, enabling this SAR target recognition method to be applied in open and dynamic environments. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a SAR image target recognition method based on few-sample incremental learning in one embodiment.

[0057] Figure 2 This is a schematic diagram illustrating the correlation between the attribute scattering center theory and the target classification method in one embodiment;

[0058] Figure 3 This is a schematic diagram of the complete model training process in this method in one embodiment;

[0059] Figure 4 This is a schematic diagram illustrating the volatility and periodicity of the target structure and features in one embodiment;

[0060] Figure 5 This is a structural block diagram of a SAR image target recognition device based on few-sample incremental learning in one embodiment;

[0061] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] Existing SAR image automatic target recognition (SAR ATR) algorithms that address Few-Shot Class-Incremental Learning (FSCIL) all face two challenges: catastrophic forgetting and overfitting.

[0064] Catastrophic forgetting refers to the phenomenon where, due to a lack of historical category supervision information, algorithms tend to forget previously learned information when learning new knowledge, leading to reduced "stability" in the representation of old knowledge. In real-world SAR ATR applications, the number of new target categories continuously increases with the acquisition of high-resolution SAR imagery. However, due to constraints such as the wide swath size of SAR imagery and data security and auditing requirements, storing and repeatedly accessing large amounts of historical data is prohibited. Furthermore, on most low-power mobile platforms (such as satellites, UAVs, and vehicles), iteratively updating depth algorithms with a large number of samples is extremely inefficient. Therefore, directly fine-tuning the DL algorithm with a small number of new category samples leads to catastrophic forgetting of old knowledge during the update process.

[0065] Overfitting, also known as unreliable empirical risk minimization, refers to the tendency of algorithms, due to their "plasticity," to easily fit weakly transferable feature representations of targets, thus reducing the algorithm's generalization ability in complex scene target discrimination. In most non-cooperative scenarios, obtaining samples of targets from multiple azimuths and under multiple imaging conditions is extremely difficult. Furthermore, the unique imaging mechanism of SAR makes high-quality annotation of target samples heavily reliant on expert knowledge, and is time-consuming and labor-intensive. Constrained by this, DL algorithms based on few-sample updates suffer from severe overfitting problems when dealing with new categories.

[0066] To address the above challenges, the stability of the balancing algorithm for old class representations and its plasticity for learning new knowledge are crucial. In particular, due to the unique imaging mechanism of SAR and the diversity of target attitudes, solving the FSCIL problem in SAR ATR faces numerous challenges. The limited nature of target information: Compared to the clear texture and significant structure of targets in optical images, where discriminative features are "what you see is what you get," targets in SAR images only exhibit a sparse distribution of key scattering structures, exhibiting significant component-like characteristics. Specifically, the discriminative information provided by small samples has large biases and low completeness, severely restricting the generalization and robustness of the algorithm's target class representation. Inter-class confusion and intra-class variability of target information: Due to the unique imaging mechanism of SAR and the similarity of target structural configurations, dissimilar targets exhibit similar scattering topologies at specific azimuth angles, resulting in high inter-class confusion. Conversely, due to the attitude sensitivity of targets, targets of the same class exhibit significantly different scattering information when azimuth angles differ greatly, leading to unstable intra-class information.

[0067] To address the above issues, this application, such as Figure 1 As shown, a SAR image target recognition method based on few-sample incremental learning is provided, including the following steps:

[0068] Step S100: Obtain the base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images.

[0069] Step S110: Based on the SAR sample images in the base class dataset, construct a multi-class target confusion training set and a single-class target rotation training set.

[0070] Step S120: The training samples from the multi-class target confusion training set and the single-class target rotation training set, along with the training samples from the original training set, are input into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples.

[0071] Step S130: Based on the incremental class dataset, construct multiple small sample training sets by randomly sampling multiple times according to N categories, each category containing K samples.

[0072] Step S140: Multiple small sample training sets are used to perform multiple rounds of incremental class training on the feature extractor and classifier that have completed base class training. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part. The class imprint cross-entropy loss is used to optimize the parameters of the feature extractor and classifier. At the beginning of each round of incremental class training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial value of the class imprint in the class imprint cross-entropy loss.

[0073] Step S150: Obtain the SAR target image to be identified, and input the SAR target image to be identified into the feature extractor and classifier that have completed the classification training to achieve target recognition.

[0074] This application proposes a few-sample incremental SAR image target recognition method based on a self-supervised decoupled dynamic classifier (SDDC). By incorporating the characteristics of SAR image targets and introducing a self-supervised task, along with a specific loss function and learning strategy, the method enhances the agile and continuous learning and discrimination capabilities for new target classes in open and dynamic environments. Specifically, considering the significant component-based and azimuth-sensitive characteristics of SAR image targets, a self-supervised learning task based on scattering component confusion and rotation modules is designed to simulate potential scattering patterns of unknown classes, improving the algorithm's agile transfer and representation of unknown targets. Furthermore, the algorithm employs a class-imprint cross-entropy loss and fine-tunes the model using a parameter decoupled learning strategy to balance the robustness and adaptability of representing new and old classes based on prior knowledge.

[0075] Furthermore, in form, few-shot incremental learning (FSCIL) assumes that the target samples are in the form of a data stream {D}. 1 D 2 D 3 ,..,D t It appears in the form of}. For the t-th stage, the target sample x is included. i With tag y i The data. C t Let C be the target category set for this stage, and its size is |C t |

[0076] In particular, the new categories at different stages do not overlap, that is... During the base class learning phase at t=1, D 1It contains a large number of target samples. During the incremental learning phase where t>1, D t It takes the form of small-sample N-way K-shot (i.e., N classes and K samples). During training, the algorithm needs to train all classes in the current stage (t). Accurate identification.

[0077] In this method, unlike classic recognition tasks that only focus on the discriminative power of the algorithm for predefined categories, the FSCIL problem needs to consider both its transferability and discriminative power. Figure 2 As shown in the figure, transferability reflects the algorithm's ability to quickly capture information about unknown targets, while discriminability represents its ability to accurately identify a large number of similar targets after fine-tuning and updating. According to the attribute scattering center theory, target SAR image information can be uniformly modeled as a specific combination of multiple class-agnostic structures, exhibiting significant componentization characteristics. Therefore, strengthening the agile representation of diverse scattering components is key to improving the algorithm's transferability representation and stable discrimination of new categories based on limited samples.

[0078] Understandably, this method includes a model training part and an actual prediction part. The model training part is the content described in steps S100 to S140, specifically including a base class training phase and an incremental class training phase. However, during the actual training of the model, in addition to these two training parts, there is also a testing phase. The entire model training process can be found in [reference needed]. Figure 3 The actual prediction part is the content described in step S150. Here, "model" refers to the feature extractor and classifier, as well as the SAR target recognition network mentioned later.

[0079] Before model training, the dataset acquisition step, namely step S100, is performed. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images. However, the number of target categories and the total number of SAR sample images in the base class dataset are greater than those in the incremental class dataset. SAR sample images belonging to the same target category exhibit differences, including variations in target elevation angle, azimuth angle, or background image.

[0080] During the base class training phase, sufficient base class data pre-acquired in the collaborative scenario is used. For deep feature extractor f 1 (x; Φ) are used for training, thereby representing the generalizable prior knowledge of targets in SAR images.

[0081] In this embodiment, based on the azimuth sensitivity and component characteristics of the scattering information of the target in the SAR image, the algorithm designs a self-supervised learning task based on the scattering component confusion and rotation module, so as to simulate the unknown class features according to the scattering components of the base class target and participate in the model optimization, thereby improving its generalization representation of the diverse scattering components of the target.

[0082] Furthermore, in step S110, a multi-class target confusion training set is constructed based on the SAR sample images in the base class dataset, including: according to the attribute scattering center theory, each target in the SAR sample image is composed of multiple strong scattering components; according to the SAR sample images of different target categories, different strong scattering components are randomly combined to construct virtual target SAR sample images of new target categories; and the multi-class target confusion training set is constructed based on multiple virtual target SAR sample images.

[0083] Specifically, the shapes of components that exhibit strong scattering in SAR sample images include dihedral, trihedral, cylindrical, and spherical.

[0084] In one embodiment, based on the Attribute Scattering Center (ASC) theory, the information of a target in a SAR image can be uniformly modeled as class-independent strong scattering components (such as... Figure 2 In this context, specific combinations of dihedral, trihedral, cylindrical, and spherical shapes are used. Enhancing the representation of diverse components in the algorithm can potentially improve its ability to capture information about unknown targets. Based on this, a scattering confusion operation is designed, the core of which is to randomly combine the key components of different categories of training samples in the current learning batch according to formula (1) to generate virtual targets containing potential class scattering structures and corresponding new class labels, and store them in [the relevant database]. This involves constructing a multi-class target confusion training set, which is then used to optimize the current model.

[0085] In one embodiment, different strong scattering components are randomly combined based on SAR sample images of different target categories using the following formula:

[0086]

[0087] In formula (1), λ represents the coefficient controlling the overlap between real and virtual categories, and Beta(a,a) represents the sampling distribution.

[0088] Preferably, the sampling is based on a Beta(a,a) distribution and the value is restricted to between 0.4 and 0.6 to reduce the confusion between real and virtual classes; α is set to 2.0. Specifically, for classes containing... Base class data of a class Can generate common A virtual target class.

[0089] In one embodiment, a specific algorithm for constructing a multi-class target confusion training set is also provided, as shown in Algorithm 1.

[0090]

[0091] Furthermore, due to the unique nature of SAR imaging mechanisms and the diversity of target attitudes, the class discriminative information of a target in a SAR image exhibits significant fluctuations and periodicities as its azimuth changes. Taking a tank from a specific class in the MSTAR dataset as an example... Figure 4 The diagram illustrates the structural and feature-level similarity results of a target with reference azimuth angles of 90°, 180°, and 270° compared to other targets of the same azimuth angle. Clearly, when the azimuth difference between the comparison target and the reference target approaches 90°, the similarity score decreases significantly and fluctuates. However, due to phase ambiguity, when this angle approaches 180°, the score is closer to that at 0°, exhibiting periodic similarity. To address this, in this embodiment, the original scattering components of the target are rotated by 90° and 180° respectively to generate new categories and participate in optimization, highlighting the differences in target scattering at azimuth angles of 0°, 90°, and 180°, potentially improving the algorithm's sensitivity to the detailed azimuth information of the target. It should be noted that both the 90° and 180° rotation angles serve as examples; in this method, the rotation angle can be freely set according to specific needs.

[0092] Specifically, in step S110, constructing a single-target rotation training set based on SAR sample images in the base class dataset includes: rotating the target in the SAR sample image by a preset angle to obtain a rotated target SAR sample image, and constructing a single-target rotation training set based on multiple rotated target SAR sample images.

[0093] In step S120, the feature extractor and classifier are trained using the multi-class target confusion training set, the single-class target rotation training set, and the original base class dataset (original training set) generated in step S110. Classical cross-entropy is used for optimization to obtain a feature extractor f that possesses both good discrimination of base class targets and generalization representation of unknown categories. 1 (x; Φ).

[0094] After training with the base classes in step S120, the feature extractor constructs a feature space based on the features of each target category. Within this feature space, the features of each target category are classified spatially. In this way, the classifier maps the features to the corresponding categories within the feature space to achieve target recognition.

[0095] During the incremental training phase, i.e. (t≥2), it includes training from Y. t-1 +1 to Y t New category data It appears continuously in the form of small samples. An imprint-cross-entropy (CI-CE) loss function was designed during this training phase to provide prior guidance for learning new knowledge. Furthermore, a parameter decoupling learning (PDL) strategy was used to... 1 (x; Φ) decoupling into static feature extraction With dynamic updates g t The algorithm consists of two parts: (θ) and (θ) to obtain a feature space and a parameterized classifier that can represent both old and new classes well, thus balancing the "stability" of the algorithm in representing old knowledge and the "plasticity" in representing new knowledge.

[0096] Specifically, the class imprint cross-entropy loss is expressed as:

[0097]

[0098] In formula (2), Let x represent the feature extractor, where the superscript t indicates the t-th training iteration, and x... i Let y represent the training sample. i D represents the truth label. t w represents the small training set during incremental training. c Represents a type of imprint.

[0099] Because incremental training involves training the feature extractor and classifier in batches using multiple small training sets, the class imprints in formula (2) are typically initialized randomly for each batch. However, this carries the risk of overfitting. Therefore, in this embodiment, the class imprints are generated by the current stage feature extractor. The sample mean is used as the class imprint to fully explore class priors, improve the algorithm's ability to adapt to new classes, and reduce the risk of overfitting. The class imprint is calculated as follows:

[0100]

[0101] In formula (3), Let i be the i-th sample in the t-th learning stage. Given its sample class label, and when it belongs to class c, the indicator function... Equal to 1, W t This represents the set of category feature vectors for the t-th learning stage.

[0102] In the task of identifying ideal closed-world objects, deep learning-based recognition algorithms, using the classic learning paradigm (i.e., iterative optimization based on sufficient data), can accurately represent and distinguish predefined target classes. However, in real-world non-cooperative scenarios, due to the small sample size of new categories and the lack of supervision information for old categories, algorithms using the classic learning paradigm are prone to "catastrophic forgetting" of old knowledge and "overfitting" to new knowledge. Based on the hierarchical learning attributes of CNNs—from local to global, from shallow texture to high-dimensional semantics, and from generalized representation to specific discrimination—a parameter decoupling (PDL) learning strategy is constructed.

[0103] To address the aforementioned issues, in this embodiment, a feature extractor is used. Decoupling for static feature extraction With dynamic updates By cascading, the feature extractor can balance generalized representation of target information with agile adaptation to discriminative features. The feature extractor is represented as follows:

[0104]

[0105] In formula (4), This represents the static feature extraction part. This represents the dynamically updated part. In the static feature extraction part, the neural network parameters remain fixed, based on the parameters obtained after training the base class. In contrast, the parameters of the dynamically updated neural network are adjusted with each incremental class training iteration.

[0106] In this embodiment, the feature extractor and classifier can be constructed into a neural network, and this neural network can be used as a SAR target recognition network.

[0107] Specifically, the feature extractor can employ a feature extraction neural network comprising sequentially connected convolutional layers, activation function layers, pooling layers, batch normalization layers, and Dropout layers. The specific structure can be adjusted according to the actual situation.

[0108] Specifically, there are various types of neural networks that can be used for classifiers, including fully connected layers, Softmax, support vector machines, decision trees, convolutional layers, or recurrent neural networks, which can be selected according to the actual application.

[0109] In this embodiment, after the SAR target recognition network that has completed incremental training performs target recognition on the SAR target image to be identified, the method further includes: if the target in the SAR target image to be identified is of an unknown category, then using the training samples in its small training set to perform incremental online training on the SAR target recognition network. This enables the SAR target recognition network to learn in real time from dynamic and constantly changing SAR target images in non-cooperative open scenarios, continuously expanding the known target categories.

[0110] In the aforementioned SAR image target recognition method based on few-sample incremental learning, during base class training, a multi-class target confusion training set and a single-class target rotation training set are constructed based on SAR sample images in the base class dataset to increase the number of target categories in the training. This allows the trained feature extractor to construct feature spaces for various target categories. During incremental class training, multiple few-sample training sets are constructed based on the incremental class dataset by performing multiple random samplings according to N categories and K samples. During incremental class training, the feature extractor is decoupled into a static feature extraction part and a dynamic update part, and the adjustable parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. In particular, during each batch of incremental class training, the mean of the training samples obtained by the current feature extractor is used as the initial value of the class imprint in the class imprint cross-entropy loss. This method can solve the two major problems of catastrophic forgetting and overfitting faced during few-sample training, enabling this SAR target recognition method to be applied to open and dynamic environments.

[0111] To address the urgent need for agile and continuous learning and discriminative capabilities in real-world non-cooperative SAR ATR applications, this paper investigates few-sample incremental learning for SAR ATR and proposes a self-supervised decoupled dynamic classifier (SDDC). Considering the component-based and azimuth-sensitive characteristics of target SAR images, this method employs a self-supervised learning task based on scattering component confusion and rotation (SCMR), class imprint cross-entropy (CI-CE) loss, and parameter decoupling learning (PDL) strategy to address the challenges specific to the problem and the domain. Experiments on the MSTAR-FSCIL and SAR-AIRcraft-1.0-FSCIL datasets, covering various targets, imaging platforms, and observation environments, validate the performance advantages of the proposed method compared to existing state-of-the-art algorithms, further demonstrating its adaptability and effectiveness in open and dynamic SAR ATR environments.

[0112] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0113] In one embodiment, such as Figure 5 As shown, a SAR image target recognition device based on few-shot incremental learning is provided, comprising: a dataset acquisition module 200, a base class training set construction module 210, a base class training module 220, a few-shot training set construction module 230, an incremental class training module 240, and a SAR image target recognition module 250, wherein:

[0114] The dataset acquisition module 200 is used to acquire base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images.

[0115] The base class training set construction module 210 is used to construct a multi-class target confusion training set and a single-class target rotation training set based on SAR sample images in the base class dataset.

[0116] The base class training module 220 is used to input the training samples from the multi-class target confusion training set and the single-class target rotation training set, as well as the training samples from the original training set, into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples.

[0117] The small sample training set construction module 230 is used to construct multiple small sample training sets based on the incremental class dataset, according to N categories, with each category containing K samples, through multiple random samplings.

[0118] The incremental training module 240 is used to perform multiple rounds of incremental training on the feature extractor and classifier that have completed base class training using multiple small sample training sets. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and the parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. At the beginning of each round of incremental training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial value of class imprint in the class imprint cross-entropy loss.

[0119] The SAR image target recognition module 250 is used to acquire SAR target images to be recognized and input the SAR target images to be recognized into a feature extractor and a classifier that have completed additive training in order to achieve target recognition.

[0120] Specific limitations regarding the SAR image target recognition device based on few-sample incremental learning can be found in the limitations of the SAR image target recognition method based on few-sample incremental learning mentioned above, and will not be repeated here. Each module in the aforementioned SAR image target recognition device based on few-sample incremental learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0121] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a SAR image target recognition method based on small-sample incremental learning. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0122] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0123] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0124] Obtain the base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images.

[0125] Based on the SAR sample images in the base class dataset, a multi-class target confusion training set and a single-class target rotation training set are constructed.

[0126] The training samples from the multi-class target confusion training set and the single-class target rotation training set, along with the training samples from the original training set, are input into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples.

[0127] Based on the incremental dataset, multiple small sample training sets are constructed by randomly sampling multiple times according to N categories, each containing K samples;

[0128] Multiple small sample training sets are used to sequentially perform multiple rounds of incremental class training on the feature extractor and classifier that have completed base class training. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and the parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. At the beginning of each round of incremental class training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial class imprint value in the class imprint cross-entropy loss.

[0129] Acquire an image of the SAR target to be identified, and input the image of the SAR target to be identified into a feature extractor and a classifier that have completed additive training to achieve target recognition.

[0130] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0131] Obtain the base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images.

[0132] Based on the SAR sample images in the base class dataset, a multi-class target confusion training set and a single-class target rotation training set are constructed.

[0133] The training samples from the multi-class target confusion training set and the single-class target rotation training set, along with the training samples from the original training set, are input into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples.

[0134] Based on the incremental dataset, multiple small sample training sets are constructed by randomly sampling multiple times according to N categories, each containing K samples;

[0135] Multiple small sample training sets are used to sequentially perform multiple rounds of incremental class training on the feature extractor and classifier that have completed base class training. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and the parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. At the beginning of each round of incremental class training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial class imprint value in the class imprint cross-entropy loss.

[0136] Acquire an image of the SAR target to be identified, and input the image of the SAR target to be identified into a feature extractor and a classifier that have completed additive training to achieve target recognition.

[0137] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0138] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0139] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A SAR image target recognition method based on small-sample incremental learning, characterized in that, The method includes: Obtain the base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images. Based on the SAR sample images in the base class dataset, a multi-class target confusion training set and a single-class target rotation training set are constructed. The training samples from the multi-class target confusion training set and the single-class target rotation training set, along with the training samples from the original training set, are input into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples. Based on the incremental dataset, multiple small sample training sets are constructed by randomly sampling multiple times according to N categories, each containing K samples; Multiple small sample training sets are used to sequentially perform multiple rounds of incremental class training on the feature extractor and classifier that have completed base class training. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and class imprint cross-entropy loss is used to optimize the parameters of the feature extractor and classifier. At the beginning of each round of incremental class training, the mean of the same-class training samples obtained from the previous stage of feature extractor is used as the initial class imprint value in the class imprint cross-entropy loss. The class imprint cross-entropy loss is expressed as: In the above formula, Indicates the first Each learning stage feature extractor, superscript Indicates the first Each learning stage Indicates training samples, Indicates the truth label, Indicates the first Training set for each learning stage objective Indicates class imprint; Acquire an image of the SAR target to be identified, and input the image of the SAR target to be identified into a feature extractor and a classifier that have completed additive training to achieve target recognition.

2. The SAR image target recognition method according to claim 1, characterized in that, The construction of a multi-class target confusion training set based on SAR sample images in the base class dataset includes: According to the property scattering center theory, targets in SAR sample images are all composed of multiple strong scattering components; Based on SAR sample images of different target categories, different strong scattering components are randomly combined to construct virtual target SAR sample images of new target categories; The multi-target confusion training set is constructed based on multiple virtual target SAR sample images.

3. The SAR image target recognition method according to claim 2, characterized in that, The shapes of components that exhibit strong scattering in SAR sample images include dihedral, trihedral, cylindrical, and spherical.

4. The SAR image target recognition method according to claim 3, characterized in that, The following formula is used to randomly combine different strong scattering components based on SAR sample images of different target categories: In the above formula, This indicates the coefficient controlling the overlap between real and virtual categories. Indicates the sampling distribution. Represents the target scattering characteristics of the generated class. This represents the scattering characteristics of a selected real-world target. This represents the target scattering characteristics of another real class selected.

5. The SAR image target recognition method according to claim 2, characterized in that, The construction of a single-class target rotation training set based on SAR sample images in the base class dataset includes: The target in the SAR sample image is rotated by a preset angle to obtain a rotated target SAR sample image; The single-target rotation training set is constructed based on multiple rotating target SAR sample images.

6. The SAR image target recognition method according to claim 5, characterized in that, In incremental training, the feature extractor parameters can be decoupled and represented as follows: In the above formula, This refers to the static feature extraction portion. This indicates the part where parameters are dynamically updated.

7. The SAR image target recognition method according to any one of claims 1-6, characterized in that, The feature extractor and classifier are constructed into a neural network, which is then used as a SAR target recognition network.

8. The SAR image target recognition method according to claim 7, characterized in that, After acquiring the SAR target image to be identified and inputting the SAR target image to be identified into the feature extractor and classifier that have completed additive training to achieve target recognition, the process further includes: If the target in the SAR target image to be identified is of an unknown category, then the training samples in the small sample training set are used to perform incremental online training on the SAR target recognition network.

9. A SAR image target recognition device based on small-sample incremental learning, characterized in that, The apparatus implements the SAR image target recognition method based on small-sample incremental learning as described in any one of claims 1-8, including: The dataset acquisition module is used to acquire base class and incremental class datasets. Both the base class dataset and the incremental class dataset include SAR sample images of multiple target categories, and each target category includes multiple SAR sample images. The base class training set construction module is used to construct a multi-class target confusion training set and a single-class target rotation training set based on SAR sample images in the base class dataset. The base class training module is used to input the training samples from the multi-class target confusion training set and the single-class target rotation training set, as well as the training samples from the original training set, into the feature extractor and the classifier for base class training. The feature extractor constructs a feature space based on the features of the training samples. The small sample training set construction module is used to construct multiple small sample training sets based on the incremental class dataset, according to N categories, with each category containing K samples, through multiple random samplings. The incremental training module is used to perform multiple rounds of incremental training on the feature extractor and classifier that have completed base class training using multiple small sample training sets. The feature extractor is decoupled into a static feature extraction part and a dynamic parameter update part, and the parameters of the feature extractor and classifier are optimized using class imprint cross-entropy loss. At the beginning of each round of incremental training, the mean of the same class training samples obtained by the feature extractor in the previous stage is used as the initial class imprint value in the class imprint cross-entropy loss. The SAR image target recognition module is used to acquire SAR target images to be recognized, and input the SAR target images to be recognized into a feature extractor and a classifier that have completed additive training to achieve target recognition.

Images