Incremental learning methods can overcome the problem of catastrophic forgetting in synthetic aperture radar (SAR) target detection models when continuously learning new class data. However, due to the characteristics of SAR images, such as large background variations and relatively stable scale differences between different target classes, existing advanced incremental detection methods perform poorly in the SAR domain. Specifically, current incremental detection research lacks adaptability to spatial distribution changes and insufficiently focuses on the model’s localization knowledge, making it significantly limited when addressing the aforementioned SAR characteristics. To tackle these issues, we propose a class-incremental learning method for SAR target detection based on smooth distribution and depth-focused knowledge distillation (SDDFD). First, we design a spatial dimension-based smooth distribution distillation (SDDL) method, which evaluates the effectiveness of spatial response points through concentration assessment and dynamically assigns weights to adapt to different background spatial distributions. Then, based on SDDL, we propose an embeddable depth-focused distillation (DFDL) approach. This approach innovatively enhances shallow localization knowledge and deep classification knowledge from a depth perspective, significantly improving the model’s localization ability. Experimental results show that our method outperforms advanced incremental detection methods in various incremental task scenarios, achieving optimal performance. For example, in one-step incremental tasks, the AP0.5 of SDDFD compared to the state-of-the-art method ERD, improved by 8.8% and 4.0% on the MSAR-1.0 and SAR-AIRcraft-1.0 datasets, respectively.

Existing incremental target detection (ITD) methods heavily depend on the diversity of information. When the scene in the new image diverges significantly from the previous training data, the detector’s ability to detect known targets diminishes considerably. The scene information in synthetic aperture radar (SAR) images is closely linked to target types, as targets of the same class often emerge in similar environments. Consequently, various scenes are frequently introduced in conjunction with new classes, posing substantial challenges to the robustness of the SAR incremental target detector. To tackle this issue, this article proposes the context-aware attention (CAA) and the Gaussian-box similarity metric (GBSM) methods to enhance the scene adaptability of SAR incremental target detectors. First, the CAA operates during the feature knowledge transfer stage, which consists of a global relation module and a local attention (LA) module. It integrates the relationship between the target and its contextual information while preserving contextual awareness through knowledge transformation. Second, the GBSM establishes a constraint factor through both 2-D Gaussian modeling and distribution similarity measurement. It further modifies the incremental localization loss to reduce the impact of target-background contrast on the model’s localization capability. We set up multiple data increment scenarios using the MSAR and SAR-Aircraft datasets. Comparative experimental results show that our method achieves better performance. In addition, the time consumption of training was recorded, and comparisons demonstrate that our method also offers advantages in efficiency.

In complex and dynamic synthetic aperture radar (SAR) scenes, few-shot detection of novel classes suffers from sample scarcity and significant distribution differences between base and novel class features, leading to severe bias and poor generalization in existing few-shot object detection (FSOD) models. To address this issue, we propose a meta-transfer learning method based on dynamic semantic guidance (DSG). This approach combines the strengths of meta-learning and transfer learning, comprising three modules: semantic guidance (SG), distribution alignment metric (DAM), and global feature dynamic aggregation (GFDA). The SG module generates guided features with query semantic information to reduce the distribution gap between base and novel classes, dynamically adapting to few-shot novel class SAR targets. The DAM module applies adversarial training to achieve dynamic feature distribution alignment, improving model bias and generalization. The GFDA module dynamically aggregates and retains critical feature information, enhancing model detection performance. Experimental results on the SRSDD-v1.0, MSAR-1.0, and SAR-AIRcraft-1.0 datasets show that the DSG method outperforms state-of-the-art methods in the SAR field (GMFBA) and the optical domain (G-FSOD), with average detection performance improvements of 1.21%, 1.45%, 1.44%, and 9.76%, 2.86%, 1.8%, respectively.

Traditional supervised learning methods have achieved great success in automatic target recognition (ATR). Unfortunately, if the model is trained exclusively on new class samples, the model will forget all knowledge about the old class samples. This phenomenon is called catastrophic forgetting. Incremental learning method can prevent catastrophic forgetting by keeping a small number of old class samples as exemplars and training them together with new class samples. However, the ratio of old to new class samples is still seriously out of balance. In this paper, an data augmentation method of old class samples is proposed by using random interpolation (RI) between two exemplars to generate fake samples in the process of incremental learning. In this method, the distribution of the original classes is partially restored while also creating a numerical balance between the old and new class samples. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset demonstrate the effectiveness of this method in Incremental SAR ATR.

With the continuous development of deep learning, a large number of deep learning models are gradually emerging in the field of SAR target recognition. However, SAR targets usually have complex textures and noises, and it is difficult to directly extract effective feature information and rich contextual information between the target object and the background, which inhibits the potential of deep learning models to further improve the recognition ability in the SAR domain. To ameliorate this problem, a GIC-Mechanism that improves the ability to capture global and local feature information interactively is proposed in this paper and applied to the Resnet family of models. In the recognition task of two SAR target datasets, the mechanism designed in this paper with the ability of multi-scale losing feature information interaction and cross-feature mapping layer information interaction improves the recognition performance of the Resnet series model by 2.16%-2.81%, which is effective.

Synthetic aperture radar (SAR) imagery is susceptible to multiple factors, including radar parameters, imaging mode, and imaging angle. Consequently, the resolution of the training and test data often differs, leading to the failure of existing deep learning-based SAR target recognition methods. To address this challenge, this paper introduces a transformer-based cross-resolution target recognition method for SAR images based on feature alignment and aggregation. The objective is to improve the accuracy of cross-resolution target recognition. Experimental results obtained from OpenSARShip, an openly available dataset with ship images captured at various resolutions, demonstrate the superior performance of the proposed method compared to the current state-of-the-art (SOTA) technique.

In recent years, the rapid development of synthetic aperture radar (SAR) technology has enabled researchers to obtain SAR targets from multiple sources within the same class. However, most existing SAR target recognition methods are designed for single-source target recognition. When applied to multi-source target recognition in SAR imagery, the different data distributions from various sources can negatively impact recognition accuracy. To address this issue, this paper proposes a scale distillation method for multi-source target recognition in SAR imagery. This approach involves extracting feature representations of the target, incorporating multi-scale semantic information through the multi-scale feature fusion mechanism (MSFFM), and performing knowledge distillation on the output scale features to enhance discriminative knowledge about the target scales. Experimental results on the FUSAR-Ship and OpenSARShip datasets from different sources demonstrate an average improvement in classification accuracy by 2.5% after applying scale distillation, which is a significant advantage over other methods.

Synthetic aperture radar (SAR) targets are often characterised by high manoeuvrability and strong concealment, resulting in scarce and few-shot SAR data. Due to the scarcity and variability of samples causing significant fluctuations in class centers, the sample distribution is challenging to determine, resulting in the model’s inability to accurately represent the potential representative features of few-shot SAR targets. Therefore, we propose a few-shot SAR target detection method based on Gaussian meta-feature directed aggregation (GMDA) using the meta-learning paradigm. Specifically, a Gaussian distribution is constructed on the support branch to estimate the class distribution of the few-shot SAR target and replace the traditional class prototype. Based on this, we propose the feature information maximization module (FIMM) to avoid the bias of the feature information and achieve the directional expression of the features in order to complete the efficient aggregation. Experiments on SRSDD-v1.0 and MSAR-1.0 datasets show that our method almost beats the state-of-the-art methods at this stage and achieves state-of-the-art performance in all settings.

In-air handwriting necessitates consistent motion tracking, in contrast to millimeter-wave (mmWave) radar-based simple gesture recognition techniques. However, during long-duration gesture tracking, challenges, such as body motion interference and environmental clutter, become more pressing. Moreover, due to the lack of a supporting surface in in-air handwriting, slight arm tremors also can result in unsmooth trajectories. To address these challenges, this article proposes a two-stage processing framework called MMHTSR. In the first stage, the state-space equations are reestablished, and a locally correlated 2-D Gaussian process regression (GPR) algorithm is employed for interframe prediction. By incorporating uncertainty estimation, weights are assigned to the next frame data, effectively suppressing interference from nongestural targets. In the second stage, real-time smoothing and tracking of gesture trajectories are accomplished using a Kalman filter, followed by mapping the trajectories onto the Cartesian coordinate system. Finally, an end-to-end signal processing framework is deployed on a low-cost 60-GHz mmWave radar prototype, and gesture trajectory recognition is achieved using deep learning methods. Experimental results demonstrate that MMHTSR can accurately track motion gestures within the range of approximately 5–40 cm and successfully recognize 30 classes of in-air gesture trajectories, including uppercase letters A–Z and four interactive gesture actions. Furthermore, the proposed framework exhibits robust performance across various scenarios which shows its adaptability.

The features employed for synthetic aperture radar (SAR) target recognition have evolved from traditional SAR target geometric features and pattern features to modern deep features, indicating a trend of increasing recognition accuracy but decreasing feature interpretability. Therefore, the fusion of multidimensional features has been investigated by many researchers. Existing feature fusion methods typically involve simple concatenation or addition of geometric features and pattern features with deep features, or directly incorporating them into deep networks. However, such fusion methods mentioned above inadequately consider the potential conflicts between features and hard to fully exploit multidimensional features. To solve the above problem, a multidimensional feature joint learning framework (MFJL-Framework) that serves the SAR target recognition task is proposed in this article, which consists of three models. Specifically, the SGC-GA-Model can select pattern features for SAR targets based on geometric feature constraints, the Global and local Feature Information interaction Capture model (GFIC-Model) can select deep features with high-level abstract semantics, and the MFFS-Model can complement and fuse these two types of features to maximize the utilization of feature information. Experiments and comprehensive ablation studies on four datasets, namely OpenSARShip-1.0, FUSAR-Ship, MSTAR-T72Variants, and SAR-AIRcraft-1.0, collectively demonstrate that the recognition performance of our proposed FJL-Framework outperforms the current state-of-the-art methods.

Synthetic aperture radar (SAR) target recognition plays an indispensable role in interpreting SAR images. However, differences in radar parameters (including factors such as imaging modes and imaging angles) often lead to resolution differences between training and test data, posing challenges for existing methods in recognizing SAR targets under cross-resolution conditions. To address this issue, this article proposes a domain adaptation (DA) framework based on feature aggregation and compensation (FAC) for cross-resolution target recognition in SAR imagery. Initially, we employ a unique local vision transformer (LocalViT) to establish global and local adversarial networks that capture invariant features under cross-resolution conditions. Following this, we design a multiscale feature fusion module (MSFFM) to capture multiscale semantic features of targets at different resolutions. Subsequently, we propose a novel class feature aggregation module (CFAM) to map targets of varying resolutions onto the unit sphere, thereby aggregating features of samples from the same class and distinguishing those of samples from different classes. Finally, we narrow down the difference in frequency-domain information of targets at different resolutions by developing a resolution semantic compensation module (RSCM). This module compensates for the semantic feature about resolution during target recognition across varying resolutions by converting high- and low-frequency information. The experimental results on three SAR datasets (OpenSARShip, FUSAR-Ship, and SRSDD-v1.0) confirm that our method outperforms the state-of-the-art (SOTA) DA methods, with an increase of 2.06%, 1.96%, and 1.61% in three sets of cross-resolution scenes, respectively.

Deep neural networks have shown remarkable effectiveness in SAR target recognition. However, the explainability problem for deep neural networks remains insufficiently addressed. One approach to tackle this challenge is the SHAP method. It enhances the explainability of deep neural networks in SAR target recognition by observing how the target, shadow, and clutter regions play their own distinct roles. The masked regions are typically filled with Zero, Mean, or Random values in optical images. But if the same operation performed on SAR images, it will affect the distribution of clutter and thus introducing new out-of-distribution challenge. In this paper, we propose a novel masking method to enhance the reliability and efficiency of the SHAP method in SAR-ATR applications. Experimental results on the MSTAR and OpenSARShip-1.0 datasets demonstrate that our proposed method provides a more faithful representation to show the importance of every single regions in SAR target recognition. Compared to methods using Zero values, Mean values, and Random baselines, our proposed method significantly enhances the reliability of explainability.

The gradual increase of SAR imagery often accompanies the appearance of new targets, but traditional detection frameworks can only detect existing target classes and cannot detect new. Typically, we must update the model using both new and old data, which puts a strain on storage and computation, but if we only update with new data, the detection performance on old classes will suffer dramatically. For this reason, this paper proposes to use the continual learning (CL) method to solve the problem of SAR target incremental detection. Mainstream CL methods generally consider localization to be a class-irrelevant function, however, this strategy is unsuitable for SAR imagery with significant background changes, leading to poor detection performance. Addressing the above issues, this paper proposes a continual learning object detection (CLOD) method, with an overall framework based on knowledge distillation. The focus of the methodology consists of two parts: firstly, we introduce the predicted location probability representation (PLPR) method, by using spatial discretization and segmented probabilistic statistics, to transform the localization results into probability distributions, thus allowing the localization function to participate in the continual learning process; secondly, we design a proposal selection strategy according to the background characteristics of SAR images, which improves the quality of the proposals during the knowledge review to further optimize the learning effect. Experiments on the latest multi-class SAR target detection dataset MSAR-1.0, show that our method is able to learn new knowledge with less performance penalty for old classes than other methods. In multiple data incremental settings, our method provides a 2%-11% performance improvement over numerous common methods.

Synthetic Aperture Radar (SAR) targets often exhibit characteristics such as high mobility and strong concealment, resulting in scarce SAR data and the manifestation of few-shot data properties. These few-shot SAR targets are susceptible to interference from complex background information and mutual interference of target features, making it challenging to distinguish SAR targets from the background. Additionally, there is confusion in features among different targets, leading to models being highly insensitive to few-shot SAR targets under complex distribution conditions in new tasks. Similarly, these few-shot SAR targets exhibit significant sample scarcity and sample variations, resulting in pronounced fluctuations in class centers and difficulty in determining sample distributions. This leads to challenges in accurately representing the potential representative features of few-shot SAR targets by the model. To address these issues, further enhancement of SAR target features is necessary to provide a robust foundation for the ultimate aggregation module. Therefore, based on the meta-learning paradigm, we propose a method for few-shot target detection in SAR imagery via intensive meta-feature aggregation (IMFA), aiming to reinforce SAR target features for improved representation. Specifically, firstly, we propose a novel hierarchical multi-head cross attention (HMCA) to capture global multiscale contextual information in different subspaces and analyze representative features between different targets to distinguish SAR targets from the background. Then, based on HMCA, we introduce a novel feature coupling module (FCM) to couple support features with cognitive information from the query image on the support branch. This is done to reduce the confusion and mutual interference of features between targets while enhancing the model’s generalization ability on new tasks. Finally, on the support branch with query-aware information, we construct a Gaussian distribution to estimate the class distribution of few-shot SAR targets and replace traditional class prototypes. On this basis, we propose the feature information maximization module (FIMM) to avoid feature information shift, greatly strengthening the expression of potential features. Through these steps, reinforced meta-features can be obtained, enabling efficient aggregation. Experiments on the SRSDD-v1.0 and MSAR-1.0 datasets demonstrate that our method has consistently outperformed state-of-the-art approaches in all configurations, achieving state-of-the-art performance.

In recent years, some studies used computer to generate synthetic aperture radar (SAR) images. Although the synthetic SAR image looks realistic, there is a domain gap and distribution differences between synthetic SAR images and measured SAR images which makes it difficult to use synthetic SAR images directly. The image translation network can reduce the gap between different domain images, but existing optical translation networks are hard to correctly learn and translate the background information of SAR images. In this paper, a synthetic-to-measured SAR target image translation network, called SAR-UT network, based on U-NET network and Transformer is proposed to bridge the domain gap between synthetic and measured SAR target images. The proposed method can extract and fuse the texture features of each layer more effectively, so as to translate SAR images with more realistic details and higher image quality. Experiments on Synthetic and Measured Paired Labeled Experiment (SAMPLE) dataset show that the proposed translation network can well learn the potential features of SAR image and translate the synthetic SAR image into the image closer to the measured SAR image. Experimental result shows that the performance of SAR-UT is better than the state-of-the-art methods.

Synthetic aperture radar (SAR) images are difficult to acquire, and the number of images of different targets often varies greatly, resulting in a large number of imbalanced class distributions in practical applications. Existing classification models usually focus too much on the classes with a large number of samples and less on the minority classes with higher-value, which leads to the degradation of classification performance. A novel method for imbalanced classification in SAR images based on reinforcement learning adaptive cost adjustment is proposed in this paper. The method can adaptively search a cost factor and adjust it continuously according to the predicted classification effect so that the performance of all classes of samples is the best possible. This method can obtain a more accurate cost factor and modify the decision boundary of the classifier to achieve a more accurate classification of the minority classes. Experiments on the MSTAR dataset show that the proposed method achieves binary and multi-classification, significantly alleviates the effect of imbalanced data, shows better classification results on all kinds of samples, and outperforms existing methods in overall performance.

In synthetic aperture radar (SAR) images, there are a large number of dense multi-scale objects, especially dense multi-scale ships docked along the coast. Existing object detection methods are difficult to simultaneously detect dense multi-scale objects in complex background. A novel method for dense multi-scale object detection in SAR images based on Feature-Transferable Pyramid Network (FTPN) is proposed in this paper. In the stage of feature extraction, the feature maps of each layer are connected effectively and the feature maps of various scales are extracted. This method can extract the features of dense multi-scale objects more effectively, so as to realize simultaneous detection of dense multi-scale objects in SAR images. Experiments on SSDD dataset, AIR-SARShip-2.0 dataset and Gaofen-3 dataset show that the proposed method can achieve dense multi-scale object detection, and the overall performance is better than the state-of-the-art methods.

In SAR images, there are objects with large scale difference, which is called cross-scale objects. For example, there are both large-scale airport objects and small-scale ship objects in SAR images. However, the current multi-scale object detection methods are difficult to detect objects with large scale difference. To address this issue, we propose a cross-scale object detection method for SAR images based on Scale Expansion Pyramid Network(SEPN) in this paper. The proposed SEPN can extract the salient features of the objects with a large scale difference, and by closely connecting the scale expansion layer with the convolutional layer during the downsampling process of the Feature Pyramid Network (FPN), the receptive field of the feature extracted by the convolutional layer can be adaptively extended, and finally it achieves the effect of cross-scale object detection in SAR image. Experiments on SSDD dataset and Gaofen-3 dataset show the effectiveness of our proposed methods in detecting objects of different scales in different scenes of SAR images.