Multi-resolution sar target recognition method and device based on scale perception domain adaptation

Abstract

The application discloses a kind of based on scale perception domain adaptation multi-resolution SAR target identification method and device, obtain source domain SAR image set and target domain SAR image set;Multi-resolution SAR target identification network is trained based on source domain SAR image set and target domain SAR image set;The target domain SAR image to be identified is identified by using the trained source domain classification network;The scale enhancement feature extraction network with adaptive kernel residual module is designed in the application, multiple kernel convolution features can be selectively fused to capture multi-granularity information in different receptive field (RF), combined with multi-level bilinear fusion module, more robust feature alignment is realized, and then the cross-domain classification task in multi-resolution SAR image is completed.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image target recognition technology, and particularly relates to a multi-resolution SAR target recognition method and device based on scale-aware domain adaptation. Background Technology

[0002] Synthetic Aperture Radar (SAR) systems are widely used in modern intelligent marine monitoring due to their all-weather, day-and-night observation capabilities, which are not limited by natural conditions. In particular, their use in ship classification and detection is of paramount importance for safeguarding national maritime rights and security.

[0003] With the application of deep learning technology in remote sensing, SAR Automatic Target Recognition (ATR) technology based on deep learning has become mainstream. Most supervised deep learning target recognition technologies require the assumption that training and test data come from the same distribution. However, this assumption is not always met in practice. For example, in remote sensing, with the increasing diversification of Earth observation methods, remote sensing images can be acquired through different satellite platforms or sensors. Due to differences in orbital altitude, viewing angle, imaging mechanism, resolution, and environment among various satellite platforms, there are significant distributional differences between multi-source remote sensing data; this distributional difference is called domain drift. Furthermore, due to its imaging mechanism, SAR image data faces greater challenges in annotation compared to natural images, making it difficult to obtain sufficient annotation data. Under these conditions, when a model learned directly from training data is used to identify test data from a different distribution, its performance often drops sharply.

[0004] To address the domain drift problem, domain adaptation techniques have been proposed to reduce the distributional differences between source and target domain data, enabling recognition models trained on source domain data to achieve good classification performance on target domain data with varying distributions. Furthermore, in remote sensing, targets in multi-source data often exhibit significant resolution / scale variations, posing a unique challenge to cross-domain recognition tasks. Existing methods do not fully consider the impact of resolution / scale variations between data on recognition tasks, and uniformly applying existing standard domain adaptation techniques to address domain drift caused by numerous factors among multi-source data has limitations. Designing a more robust recognition model that maintains good generalization performance even with data exhibiting significant resolution / scale variations and differences to successfully complete cross-domain recognition tasks has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-resolution SAR target recognition method and apparatus based on scale-aware domain adaptation. By extracting the scale-enhanced features of the target and then performing target recognition on the scale-enhanced features, the robustness of target recognition can be improved.

[0006] This invention adopts the following technical solution: a multi-resolution SAR target recognition method based on scale-aware domain adaptation, comprising the following steps:

[0007] Obtain the source domain SAR image set and the target domain SAR image set; wherein, the source domain SAR image set consists of source domain SAR images and their category labels, and the target domain SAR image set consists of target domain SAR images;

[0008] A multi-resolution SAR target recognition network is trained based on source domain SAR image sets and target domain SAR image sets. The multi-resolution SAR target recognition network consists of a source domain classification network and a domain adversarial network. The source domain classification network includes a scale-enhanced feature extraction network and a classifier, while the domain adversarial network includes a domain discriminator, a multi-level bilinear fusion module, and a scale-enhanced feature extraction network.

[0009] A pre-trained source domain classification network is used to identify targets in the SAR image of the target domain to be identified.

[0010] Furthermore, the loss function of the multi-resolution SAR target recognition network is:

[0011] L total =L cls +αL adv ,

[0012] Among them, L total L is the loss function of the multi-resolution SAR target recognition network. cls Let L be the loss function of the source domain classification network. adv Let α be the loss function of the domain adversarial network, and α be the compromise weight.

[0013] Furthermore, the loss function of the source domain classification network is as follows:

[0014]

[0015] Among them, L cls (θ G ,θ C ) is a training parameter θ G and θ C The loss function of the source domain classification network, where G(·) represents the scale-enhanced feature extraction network and C(·) represents the classifier. This represents the feature extracted from the i-th SAR image in the source domain S. express The corresponding true class label of the SAR image, n s This represents the number of SAR images in the source domain S.

[0016] Furthermore, the loss function of the domain adversarial network is specifically as follows:

[0017]

[0018] Among them, L adv (θ G ,θ D ) is a training parameter θ G and θ D The loss function of the domain adversarial network, n s G(·) represents the number of SAR images in the source domain S, G(·) represents the scale-enhanced feature extraction network, and D(·) represents the domain discriminator. n represents the features extracted from the i-th SAR image in the source domain S. t This represents the number of SAR images in the target domain t.

[0019] Furthermore, a scale enhancement feature extraction network is used to extract scale enhancement features from SAR images. Specific methods include:

[0020] Extracting feature X from SAR images using a convolutional network;

[0021] The feature X is sequentially passed through stacked adaptive convolution kernel residual modules to obtain the scale enhancement features of the SAR image.

[0022] Furthermore, the adaptive convolution kernel residual module is used to sequentially perform separation, fusion, and selection operations on the features X;

[0023] The separation operation involves feeding feature X into two feature branches with different kernel sizes, and obtaining different feature maps respectively. and

[0024] The fusion operation includes combining feature maps and Perform a summation operation to obtain the fused feature U; then pass the fused feature U through a global average pooling layer and a fully connected layer to obtain the feature representation vector z.

[0025] The selection operation includes calculating the feature map using the feature representation vector z. The weight vector a and the feature map The weight vector b, combined with the feature map and The fusion feature V is obtained by calculating the corresponding weight vector, and the output feature F of the adaptive convolution kernel residual module is calculated based on the fusion feature V.

[0026] Furthermore, the output feature F is calculated as follows:

[0027]

[0028] Furthermore, the multi-level bilinear fusion module is used to extract multi-level fusion features from SAR images, and the specific methods include:

[0029] Features from the residual modules of different adaptive convolutional kernels are obtained from the scale-enhanced feature extraction network and fused to obtain multi-level fused features.

[0030] Furthermore, after obtaining the multi-level fusion features, it also includes:

[0031] The scale enhancement features and multi-level fusion features of the same SAR image are fused and then fed into the discriminator.

[0032] Another technical solution of the present invention: a multi-resolution SAR target recognition device based on scale-aware domain adaptation, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-resolution SAR target recognition device based on scale-aware domain adaptation as described above.

[0033] The beneficial effects of this invention are as follows: This invention designs a scale-enhanced feature extraction network with an adaptive kernel residual module, which can selectively fuse multiple kernel convolution features to capture multi-granular information in different receptive fields (RFs). Combined with a multi-level bilinear fusion module, it achieves more robust feature alignment, thereby completing the cross-domain classification task in multi-resolution SAR images. Attached Figure Description

[0034] Figure 1 This is an overall framework diagram of the multi-resolution SAR target recognition network in an embodiment of the present invention;

[0035] Figure 2 This is an architecture diagram of the adaptive convolution kernel residual module in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of a multi-level bilinear fusion adversarial strategy in an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of some SAR image data in a verification embodiment of the present invention. Detailed Implementation

[0038] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0039] Because remote sensing images can be acquired from different sensors, significant distributional differences exist between multi-source data, a phenomenon known as domain shift. In this invention, "source" specifically refers to the origin of the image; multi-source means remote sensing images from multiple sources, i.e., images acquired by different sensors. Domain adaptation techniques are commonly used to mitigate this problem. Furthermore, large resolution / scale differences (an inherent characteristic of multi-source remote sensing images, exhibiting different resolutions or scales) are often more common in remote sensing images than in natural images, and this has not received sufficient attention. Directly using domain adaptation methods significantly limits the performance of transfer models. Additionally, most current feature alignment methods are implemented under single-scale feature representations, which often affects the quality of feature alignment, especially with data exhibiting large resolution / scale variations.

[0040] Convolutional Neural Networks (CNNs) have achieved a series of milestones in the computer vision community, thanks to their powerful ability to extract layers of discriminatively meaningful features without any human intervention. The receptive field (RF) is a crucial parameter in CNN models, defined as the region of the input image that can be perceived on the corresponding feature map. Therefore, if the RF size is not chosen appropriately, the CNN model may fail to adequately capture target information. This is particularly problematic for fine-grained classification tasks, where the situation can be exacerbated when the input images are presented at different resolutions / scales.

[0041] To enable the model to adaptively select the RF size and fully capture multi-granularity information, a pre-designed scale-enhanced feature extraction network is used in the scale-aware domain adaptation method. By adaptively and selectively fusing features from convolutional kernels of different sizes, the network obtains scale-enhanced features, improving the model's adaptability to targets at different resolutions / scales and fully capturing multi-granularity features.

[0042] Meanwhile, most existing adversarial domain adaptation methods use the network's final output as the feature input for adversarial learning, and using a single-scale feature representation for feature alignment has limitations. Especially in the field of remote sensing, large differences in resolution / scale are more common than in natural image data, and satisfactory performance is often not achieved in tasks involving such data.

[0043] Therefore, in adversarial learning strategies based on multi-level bilinear feature fusion, this invention fully considers this limitation. Unlike other methods, this invention introduces multi-level, multi-scale fused features into the adversarial learning process to achieve feature alignment at multiple scales, rather than a single scale. This improves the feature alignment quality of the model and enhances its generalization performance across data at different resolutions / scales. It also enables the completion of cross-domain ship target identification tasks.

[0044] Specifically, within the scale-aware domain adaptation framework, a scale-enhanced feature extraction network based on adaptive convolutional kernels is first designed. This network can adaptively, selectively, and fully capture target features at different resolutions / scales to obtain scale-enhanced features of the targets. Compared with feature extraction networks with fixed convolutional kernel sizes, this improves the network's adaptability to multi-source remote sensing targets with varying resolutions / scales. Next, an adversarial learning strategy based on multi-level bilinear feature fusion is designed to fully incorporate scale factors into the domain adaptation adversarial learning process, thereby enhancing the model's generalization ability across multi-source data with large resolution / scale variations.

[0045] More specifically, this invention discloses a multi-resolution SAR target recognition method based on scale-aware domain adaptation, comprising the following steps: acquiring a source domain SAR image set and a target domain SAR image set; wherein, the source domain SAR image set consists of source domain SAR images and their category labels, and the target domain SAR image set consists of target domain SAR images; training a multi-resolution SAR target recognition network based on the source domain SAR image set and the target domain SAR image set; wherein, the multi-resolution SAR target recognition network consists of a source domain classification network and a domain adversarial network; the source domain classification network includes a scale-enhanced feature extraction network and a classifier, and the domain adversarial network includes a domain discriminator, a multi-level bilinear fusion module, and a scale-enhanced feature extraction network; and using the trained source domain classification network to perform target recognition on the target domain SAR images to be identified.

[0046] This invention designs a scale-enhanced feature extraction network with an adaptive kernel residual module, which can selectively fuse multiple kernel convolutional features to capture multi-granular information in different receptive fields (RFs). Combined with a multi-level bilinear fusion module, it achieves more robust feature alignment, thereby completing the cross-domain classification task in multi-resolution SAR images.

[0047] This invention proposes a scale-aware domain adaptation framework to complete cross-domain SAR ship classification tasks. First, an adaptive scale enhancement feature network based on an adaptive kernel residual module is designed to selectively fuse multiple kernel convolutional features to fully capture multi-granularity information. Then, during the adversarial learning process of adversarial domain adaptation, a multi-level bilinear fusion adversarial strategy is used to improve generalization performance under different resolution / scale data. This completes the cross-domain ship target identification task. The overall framework of the scale-aware domain adaptation multi-resolution SAR target identification method is as follows: Figure 1 As shown.

[0048] The recognition network of this method mainly consists of two parts: a source domain classification network composed of a scale-enhanced feature extraction network and a classifier, and a domain adversarial network composed of a scale-enhanced feature extraction network, a multi-level bilinear fusion module, and a domain discriminator.

[0049] The training process can be divided into two steps: training the classification network and feature adaptation. In the classification network training, labeled source domain data is fed into the feature extraction network to obtain source domain features. These source domain features are then fed into the classifier, and the source domain classification network is iteratively trained by optimizing the cross-entropy classification loss function. In the feature adaptation step, labeled source domain data and unlabeled target domain data are fed together into a shared feature extraction network to obtain source and target domain features respectively. These features are then fed into the domain discriminator, which is trained by optimizing a binary classification loss function to correctly predict the domain labels of the features. Simultaneously, the weights of the feature extraction network are updated to reduce the distribution difference between the two datasets, ultimately achieving domain confusion and obtaining domain-invariant features. The gradient reversal layer (GRL) in the adversarial domain adaptation framework is embedded between the feature network and the domain discriminator.

[0050] Finally, in the testing phase, since we have made the distribution differences between the source domain and the target domain as close as possible during the training process, we can directly feed the target domain data into the source domain classification network, which is composed of a classifier trained with source domain data and a feature extraction network with adjusted weights, to directly identify the target data and obtain its classification result.

[0051] In the scale-enhanced feature extraction network, in order to fully capture and extract multi-scale and multi-granular information from different source data, the scale-enhanced features of the data are obtained by stacking adaptive kernel residual modules (AKRM).

[0052] The scale enhancement feature extraction network is used to extract scale enhancement features of SAR images. The specific method includes: extracting features X of SAR images through a convolutional network; and passing features X sequentially through stacked adaptive convolutional kernel residual modules to obtain scale enhancement features of SAR images.

[0053] The detailed structure of the adaptive convolutional kernel residual module is as follows: Figure 2As shown, this module mainly consists of three steps: splitting, fusing, and selecting. Its input is the output features of the convolutional network. In the splitting stage, the original features are fed into two feature branches with different kernel sizes. In this embodiment, a 3×3 convolutional kernel and a 5×5 dilated convolutional kernel with a 2-column spacing are used to improve computational efficiency. The feature maps obtained from the two different branches can be represented as follows: and

[0054] In other words, the adaptive convolutional kernel residual module is used to sequentially perform separation, fusion, and selection operations on feature X. The separation operation involves feeding feature X into two feature branches with different kernel sizes, resulting in different feature maps. and

[0055] The fusion operation includes combining feature maps and A summation operation is performed to obtain the fused feature U. This fused feature U is then passed sequentially through a global average pooling layer and a fully connected layer to obtain the feature representation vector z. Specifically, in the fusion stage, features from different feature branches are first fused using a summation operation to obtain the fused feature. Then, a global average pooling (GAP) and a fully connected layer F are used. fc We obtain a compact feature representation vector z = F fc (GAP(U)).

[0056] The selection operation includes calculating the feature map using the feature representation vector z. The weight vector a and the feature map The weight vector b, combined with the feature map and The fused feature V is obtained by calculating the corresponding weight vector, and the output feature F of the adaptive convolutional kernel residual module is calculated based on the fused feature V. Specifically, in the selection stage, in order to achieve automatic selection of multi-scale information of different convolutional kernel sizes, a weight allocation method based on an attention mechanism is used. Larger weights are assigned to features with greater discriminative significance to obtain scale-enhanced features and improve feature quality. The attention weights are implemented using a softmax function, resulting in attention weight vectors a and b with the same dimension as the number of feature channels. Therefore, the output features are obtained by calculating the attention-weighted sum of features from different branches. Finally, this embodiment also uses residual connections to alleviate the gradient vanishing problem, obtaining the output features of each adaptive convolutional kernel residual module.

[0057] The task of the source domain classification network is to ensure that the network can accurately predict the category labels of the source domain data. Source domain features are obtained through a feature extraction network, and then the network's label predictions are obtained through a classifier consisting of fully connected layers and a Sofrmax activation function. Therefore, the most common category cross-entropy is used as the network's optimization function. Specifically, the loss function of the source domain classification network is as follows:

[0058]

[0059] Among them, L cls (θ G ,θ C ) is a training parameter θ G and θ C The loss function of the source domain classification network, where G(·) represents the scale-enhanced feature extraction network and C(·) represents the classifier. This represents the feature extracted from the i-th SAR image in the source domain S. express The corresponding true class label of the SAR image, n s This represents the number of SAR images in the source domain S.

[0060] Because of the significant distributional differences between different data domains, classifiers trained directly on source domain data will experience substantial performance degradation when used to predict labels for target domain data. Therefore, feature adaptation is necessary to align feature representations across different domains, enabling the classifier to accurately predict labels for both source and target domain data while maintaining good classification performance.

[0061] In the feature adaptation phase, data from both the source and target domains are simultaneously fed into a scale-enhanced feature extraction network to obtain their feature representations. To further improve the model's generalization performance on large-resolution / scale-difference data, this method employs a multi-level bilinear fusion adversarial strategy. Unlike other existing methods that only use adversarial strategies based on features at a single scale, this method uses a multi-level, multi-scale feature adversarial approach—specifically, the designed multi-level bilinear fusion adversarial strategy (MLBFM)—to achieve more robust feature alignment.

[0062] Its main process is as follows Figure 1 , Figure 3As shown, the multi-level bilinear fusion module is used to extract multi-level fusion features of SAR images. The specific method includes: obtaining the features output by different adaptive convolution kernel residual modules from the scale enhancement feature extraction network and fusing them to obtain multi-level fusion features; and fusing the scale enhancement features and multi-level fusion features of the same SAR image and then sending them to the discriminator.

[0063] Specifically in this embodiment, such as Figure 1 As shown, the features output by the first adaptive convolutional kernel residual module and the features output by the second adaptive convolutional kernel residual module are fused and then fed into the discriminator. Feature maps from different levels and scales of the network are fused using bilinear pooling (BP) to obtain fused features. Where A and B are mapping matrices, The features output by the first adaptive convolution kernel residual module, Let F be the feature output of the second adaptive convolutional kernel residual module, and P be the classification matrix. Finally, we obtain F. domain (·) = concat(V′,Z) and feed it into the domain discriminator, where V′ is the feature output by the third adaptive convolution kernel residual module.

[0064] In domain adversarial networks (DANs), the domain classifier's task is to accurately determine whether a sample comes from the source or target domain. To align features between the source and target domains, a backward gradient layer (RBL) is used within the adversarial domain adaptation framework. This RBL is inserted between the feature extractor and the domain discriminator in the DAN. Its function is to invert the gradient passed through the discriminator and then backpropagate it to the feature extraction network for parameter optimization. Thus, in this network, the feature extraction network is trained to deceive the domain classifier, while the domain classifier aims to correctly identify the sample's domain label. This achieves domain confusion through a minimax game, ultimately aligning the features of the two different domains. Therefore, the loss function of this DAN is specifically:

[0065]

[0066] Among them, L adv (θ G ,θ D ) is a training parameter θ G and θ D The loss function of the domain adversarial network, n s G(·) represents the number of SAR images in the source domain S, G(·) represents the scale-enhanced feature extraction network, and D(·) represents the domain discriminator. n represents the features extracted from the i-th SAR image in the source domain S. t This represents the number of SAR images in the target domain t.

[0067] Therefore, the loss function of the multi-resolution SAR target recognition network is:

[0068] L total =L cls +αL adv ,

[0069] Among them, L total L is the loss function of the multi-resolution SAR target recognition network. cls Let L be the loss function of the source domain classification network. adv Let α be the loss function for the domain adversarial network, and α be the trade-off weight. During network training, the network is optimized using a stochastic gradient descent (SGD) optimizer with a learning rate of 0.0001, a momentum of 0.9, and other parameters set to default.

[0070] The target recognition method of this invention first designs an adaptive scale-enhanced feature extractor with an adaptive kernel residual module, which selectively fuses multiple kernel convolutional features to capture multi-granular information in different receptive fields (RFs). Then, by incorporating multi-level fused features of different scales into the adversarial learning process, a multi-level bilinear fusion adversarial strategy in domain adaptation is realized to complete the cross-domain ship classification task in multi-resolution SAR images.

[0071] To verify the effectiveness of the present invention, the following verification embodiments were also carried out.

[0072] To validate the data, SAR ship target images from different sources with varying resolutions are first collected as input to the network. In this method, Fudan University's FUSAR-Ship dataset is used as the source domain data, and Shanghai Jiao Tong University's OpenSARShip dataset is used as the target domain data. The FUSAR-Ship data, acquired from the Gaofen-3 (GF-3) satellite, contains over 5000 target tile images across 15 categories. The OpenSARShip dataset is acquired from the Sentinel-1 satellite. Some data examples are shown below. Figure 4 As shown. To avoid class imbalance in our dataset, which could negatively impact the results, we selected three main shared categories in this dataset: bulk carriers, cargo ships, and tankers. These are also the most common ship categories, accounting for over 95% of ships in the real world. All three categories have relatively sufficient sample sizes.

[0073] The basic information of the dataset is shown in Table 1. It is easy to observe that the constructed cross-domain dataset is characterized by significant differences in image resolution / scale and information granularity between different domains, which is often absent in natural images. More importantly, the constructed dataset exhibits high intra-class similarity, both of which present unique challenges for cross-domain classification of SAR ships.

[0074] Table 1. Basic information about the dataset used in the validation process.

[0075]

[0076]

[0077] To verify the efficiency of the method, this invention was compared with several mainstream domain adaptation algorithms. The algorithms compared included Domain Adversarial Neural Network (DANN), Adversarial Discriminative Domain Adaptation (ADDA), Batch Nuclear-norm Maximization (BNM), Deep Subdomain Adaptation Network (DSAN), DeepCoral, Deep Adaptation Network (DAN), and a source-only model. Compared to the aforementioned algorithms, this invention achieved the highest average recognition accuracy in all cross-domain recognition tasks. Specific recognition results are shown in Table 2, which fully demonstrates the efficiency of this invention in cross-domain target recognition under conditions of large resolution differences.

[0078] Table 2. Average recognition accuracy of the present invention and the comparative method on the constructed cross-domain dataset.

[0079]

[0080] In summary, to overcome the problem that feature networks with a single receptive field cannot fully capture targets of different scales, this invention employs a scale-enhanced feature extraction network. By selectively assigning different attention weights to features of different scales, it adaptively fuses multi-scale features from convolutional kernels of different sizes. This improves the network's adaptability to target data from different sources but with significant differences in resolution scale, enabling it to fully capture multi-scale and multi-granular information to obtain scale-enhanced features and improve feature quality.

[0081] This invention fully considers the impact of resolution / scale on the feature alignment process in domain adaptation, which has not received sufficient attention in existing domain adaptation methods. It employs an adversarial strategy based on multi-level bilinear fusion to overcome the limitations of existing domain adaptation methods when handling data with large resolution differences. Unlike most existing domain adaptation methods based on adversarial learning under single-scale features, this method can effectively improve the feature alignment between multi-source remote sensing datasets with large resolution / scale differences, enhance the model's generalization performance on data with large resolution / scale differences, and thus improve the accuracy of cross-domain recognition.

[0082] This invention, based on an adversarial domain adaptation framework, fully considers the impact of resolution / scale differences on cross-domain recognition tasks in the remote sensing field. It employs a scale-enhanced feature extraction network that adaptively and selectively fuses multi-scale and multi-granular information to acquire scale-enhanced features, improving the network's adaptability to targets at different scales and enabling it to capture more discriminative information to improve feature quality. Furthermore, unlike existing domain adaptation methods that rely solely on adversarial learning under a single-scale feature representation, this invention uses a fusion method that integrates multi-level, multi-scale features from the network. Adversarial learning is then performed under this fused feature representation to achieve multi-scale feature alignment, making the network more robust to data with large resolution and scale differences. This improves the model's generalization performance and enables cross-domain recognition of SAR ship targets under large resolution and scale differences. It effectively overcomes the limitations of domain adaptation algorithms caused by large resolution differences in SAR images, achieving a more robust domain adaptation algorithm with better generalization capabilities, thus greatly expanding the applicability of this invention.

[0083] The present invention also discloses a multi-resolution SAR target recognition device based on scale-aware domain adaptation, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the multi-resolution SAR target recognition device based on scale-aware domain adaptation as described above.

[0084] The aforementioned device can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. This device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the device may include more or fewer components, or a combination of certain components, or different components; for example, it may also include input / output devices, network access devices, etc.

[0085] The processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0086] In some embodiments, the memory may be an internal storage unit of the device, such as a hard disk or RAM. In other embodiments, the memory may be an external storage device, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units. The memory is used to store operating systems, applications, bootloaders, data, and other programs, such as the program code of computer programs. The memory can also be used to temporarily store data that has been output or will be output.

[0087] It should be noted that the specific contents of the above-mentioned device are based on the same concept as the method embodiment of the present invention. For details on its specific functions and the technical effects it brings, please refer to the method embodiment section, which will not be repeated here.

Claims

1. A multi-resolution SAR target recognition method based on scale-aware domain adaptation, characterized in that, Includes the following steps: Obtain a source domain SAR image set and a target domain SAR image set; wherein, the source domain SAR image set consists of source domain SAR images and their category labels, and the target domain SAR image set consists of target domain SAR images; A multi-resolution SAR target recognition network is trained based on the source domain SAR image set and the target domain SAR image set; wherein, the multi-resolution SAR target recognition network consists of a source domain classification network and a domain adversarial network; the source domain classification network includes a scale-enhanced feature extraction network and a classifier, and the domain adversarial network includes a domain discriminator, a multi-level bilinear fusion module and the scale-enhanced feature extraction network; The trained source domain classification network is used to perform target recognition on the SAR image of the target domain to be identified. The scale enhancement feature extraction network is used to extract scale enhancement features from SAR images, and the specific methods include: Extracting feature X from SAR images using a convolutional network; The feature X is sequentially passed through stacked adaptive convolution kernel residual modules to obtain the scale enhancement features of the SAR image; The multi-level bilinear fusion module is used to extract multi-level fusion features from SAR images. Specific methods include: Features output by different adaptive convolutional kernel residual modules are obtained from the scale enhancement feature extraction network and fused to obtain multi-level fused features; the scale enhancement features and multi-level fused features of the same SAR image are fused and then fed into the discriminator. The features output by the first adaptive convolution kernel residual module and the features output by the second adaptive convolution kernel residual module Multi-level fused features are obtained by fusing through bilinear pooling. ,in, , For the mapping matrix, This is a classification matrix; The scale enhancement features and multi-level fusion features of the same SAR image are fused to obtain... ,in, The features output by the third adaptive convolution kernel residual module; The adaptive convolution kernel residual module is used to sequentially perform separation, fusion, and selection operations on the feature X; The separation operation includes feeding the feature X into two feature branches with different kernel sizes, and obtaining different feature maps respectively. and ; The fusion operation includes combining feature maps and Perform a summation operation to obtain the fusion features. Then the fusion features The feature representation vector is obtained by sequentially passing through a global average pooling layer and a fully connected layer. ; The selection operation includes using feature representation vectors. Calculate the feature map separately The weight vector a and the feature map The weight vector b, combined with the feature map and The fused features are obtained by calculating the corresponding weight vectors. Then based on the fusion characteristics Calculate the output features of the adaptive convolution kernel residual module. ; The output features The calculation method is as follows: 。 2. The multi-resolution SAR target recognition method based on scale-aware domain adaptation as described in claim 1, characterized in that, The loss function of the multi-resolution SAR target recognition network is: ， in, Let be the loss function of the multi-resolution SAR target recognition network. Let be the loss function of the source domain classification network. Let be the loss function of the domain adversarial network. This is a compromise weight.

3. The multi-resolution SAR target recognition method based on scale-aware domain adaptation as described in claim 2, characterized in that, The loss function of the source domain classification network is specifically as follows: ， in, For having trainable parameters and The loss function of the source domain classification network, This refers to the scale-enhanced feature extraction network. Represents a classifier. This represents the feature extracted from the i-th SAR image in the source domain S. express The corresponding true category label of the SAR image, This represents the number of SAR images in the source domain S.

4. The multi-resolution SAR target recognition method based on scale-aware domain adaptation as described in claim 2, characterized in that, The loss function of the domain adversarial network is specifically as follows: ， in, For having trainable parameters and The loss function of the domain adversarial network, This represents the number of SAR images in the source domain S. This refers to the scale-enhanced feature extraction network. Representation domain discriminator, This represents the feature extracted from the i-th SAR image in the source domain S. This represents the number of SAR images in the target domain t.

5. A multi-resolution SAR target recognition device based on scale-aware domain adaptation, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a multi-resolution SAR target recognition method based on scale-aware domain adaptation as described in any one of claims 1-4.

Images