Mass production of high-quality synthetic SAR training imagery is essential for boosting the performance of deep-learning (DL)-based SAR automatic target recognition (ATR) algorithms in an open-world environment. To address this problem, we exploit both the widely used Moving and Stationary Target Acquisition and Recognition (MSTAR) SAR dataset and the Synthetic and Measured Paired Labeled Experiment (SAMPLE) dataset, which consists of selected samples from the MSTAR dataset and their computer-generated synthetic counterparts. A series of data augmentation experiments are carried out. First, the sparsity of the scattering centers of the targets is exploited for new target pose synthesis. Additionally, training data with various clutter backgrounds are synthesized via clutter transfer, so that the neural networks are better prepared to cope with background changes in the test samples. To effectively augment the synthetic SAR imagery in the SAMPLE dataset, a novel contrast-based data augmentation technique is proposed. To improve the robustness of neural networks against out-of-distribution (OOD) samples, the SAR images of ground military vehicles collected by the self-developed MiniSAR system are used as the training data for the adversarial outlier exposure procedure. Simulation results show that the proposed data augmentation methods are effective in improving both the target classification accuracy and the OOD detection performance. The purpose of this work is to establish the foundation for large-scale, open-field implementation of DL-based SAR-ATR systems, which is not only of great value in the sense of theoretical research, but is also potentially meaningful in the aspect of military application.

This paper presents pioneer research on joint scenetarget analysis and propose novel cross-modality real-synthetic target feature fusion method. To begin, multi-sensor remote sensing images are jointly leveraged for geographical region classification. After that, a novel Context-Aware Region Masking and Situation AWareness (CARMSAW) strategy is employed for target classification based on the inherent target properties and capabilities reflected by SAR and infrared (IR) imagery, and the Cross-modality Real-synthetic Image Merging (CRIM) strategy is employed for feature enhancement. Specifically, to tackle with the random deviations of the real SAR imagery from the ideal ones, the synthetic SAR signature generated based on the target CAD model is treated as a “skeleton” with known structure for real-sync target feature alignment. To facilitate the recognition of aircrafts, we leverage on the IR images to construct an “exoskeleton” for the target SAR signature, so that the dimension/shape/contour of the target and its electromagnetic features are united. Furthermore, we propose a novel color-guided component-level attention mechanism, in which the SAR image is partitioned into several subregions highlighted or blacked-out adaptively based on their significance level. To demonstrate the effectiveness of the proposed CARMSAW strategy, a series of experiments are carried out based on the SAR-optical image pairs from the SEN1-2 dataset, the SpaceNet6 dataset, and a self-constructed ship detection dataset featuring the Port of Rotterdam. To verify the performance the proposed CRIM method, experiment results based on both the self-constructed SAR-IR dataset and the MSTAR-SAMPLE dataset in the public domain are provided.

Bistatic synthetic aperture radar (SAR) with spatially separated transmitter and receiver is advantageous over monostatic SAR systems in trajectory flexibility and anti-stealth/anti-jamming capability. On the other hand, since bistatic SAR imaging involves more technical complexities and incurs higher cost, the research in the field of bistatic automatic target recognition has been mainly relying on the simulated SAR imagery. Reckoning with the lack of supporting database in the public domain, the researchers at Nanjing University of Aeronautics and Astronautics constructed a proprietary bistatic SAR database featuring multiple types of representative military vehicles with the self-developed miniSAR system. Moreover, a multichannel multiview feature fusion network (MMFFN) is devised by incorporating the Vision-Transformer. Simulation results show that the proposed MMFFN offers a classification accuracy improvement of 4.86%-16.63% over the baseline network (i.e. the plain Vision Transformer) in a series of experiments featuring small-to-large observation angle deviations between the training and the test data.

In recent years, convolutional neural networks (CNNs) have demonstrated significant potential for synthetic aperture radar (SAR) target recognition. SAR images possess a strong sense of granularity and contain texture features of varying scales, including speckle noise, dominant scatterers and target contours, which are not typically considered in traditional CNN models. This paper proposes two residual blocks, termed EMC ² A blocks, with multiscale receptive fields (RFs) based on a multibranch structure and designs an efficient isotopic architecture deep CNN (DCNN) called EMC ² A-Net, whose structure is interpretable from a probability and mathematical statistics perspective. EMC ² A blocks employ parallel dilated convolution with different dilation rates to effectively capture multiscale contextual features without significantly increasing the computational load. To further enhance the efficiency of multiscale feature fusion, this paper presented a multiscale feature cross-channel attention module, known as the EMC ² A module, which adopts a local multiscale feature interaction strategy without dimensionality reduction. This strategy adaptively adjusts the weights of each channel using efficient one-dimensional (1D)-circular convolution and sigmoid function to guide attention at the global channel-wise level. Comparative results on the MSTAR dataset demonstrate that EMC ² A-Net outperforms the other available models of the same type and possesses a relatively lightweight network structure. The ablation experiment results further demonstrate that the EMC ² A module significantly enhances the model’s performance by utilizing only a few parameters and appropriate cross-channel interactions.

Synthetic aperture radar (SAR) automatic target recognition (ATR) is remarkably challenging since the SAR image defies the foundation for human and computer vision, i.e., the Gestalt perceptual principles. We propose to address this problem by fusing the target features reflected in SAR and infrared (IR) images via a novel dual‐channel context‐guided feature‐alignment network (CGFAN) that is capable of fusing the cross‐modality projective‐invariant features extracted from unpaired SAR and IR images. First, region of interest (ROI) matching between SAR and IR images is realized based on special landmarks exhibiting consistent cross‐modality features. After that, generative models trained with historical SAR and IR images are used to synthesize SAR images based on the IR images collected in real time for the current mission. Since SAR imaging takes more time than IR imaging, by using these synthesized SAR images as auxiliary data, the spatial‐coverage rate in a typical collaborative SAR/IR ATR mission carried out by drone swarms is effectively improved. The proposed CGFAN is tested against the proprietary monostatic‐bistatic circular SAR and IR dataset constructed by the researchers at our institution, which consists of nine types of military vehicles. Experimental results show that the proposed CGFAN offers better ATR performance than the baseline networks.A novel dual‐channel CGFAN that is capable of fusing the cross‐modality projective‐invariant features extracted from unpaired SAR and IR images is proposed. First, ROI matching between SAR and IR images are realized based on special landmarks exhibiting consistent cross‐modality features. After that, generative models trained with historical SAR and IR images are used to synthesize SAR images based on the IR images collected in real time for the current mission.