Vehicle recognition is a fundamental problem in SAR image interpretation. However, robustly recognizing vehicle targets is a challenging task in SAR due to the large intraclass variations and small interclass variations. Additionally, the lack of large datasets further complicates the task. Inspired by the analysis of target signature variations and deep learning explainability, this paper proposes a novel domain alignment framework named the Hierarchical Disentanglement-Alignment Network (HDANet) to achieve robustness under various operating conditions. Concisely, HDANet integrates feature disentanglement and alignment into a unified framework with three modules: domain data generation, multitask-assisted mask disentanglement, and domain alignment of target features. The first module generates diverse data for alignment, and three simple but effective data augmentation methods are designed to simulate target signature variations. The second module disentangles the target features from background clutter using the multitask-assisted mask to prevent clutter from interfering with subsequent alignment. The third module employs a contrastive loss for domain alignment to extract robust target features from generated diverse data and disentangled features. Lastly, the proposed method demonstrates impressive robustness across nine operating conditions in the MSTAR dataset, and extensive qualitative and quantitative analyses validate the effectiveness of our framework.

In this paper, the convolution neural networks (CNN) are developed for the classification and localization of mixed near-field and far-field sources by using the geometry of symmetric nested array. We first transform the received data into frequency domain. Then, we preprocess the phase difference matrix to decouple mixed sources. Considering that the counter-diagonal elements in the phase difference matrix only contain the direction of arrival (DOA) parameter of each mixed source, we utilize the upper right elements as the input of CNN to estimate the DOA of each mixed source. In order to avoid the influence of noise, we construct a particular range vector without the estimated DOA and employ the output of autoencoder to classify the mixed sources. Finally, we further exploit the output of autoencoder and apply the CNN without bias vectors to estimate the range of near-field sources. In contrast to the traditional learning-based approaches regarding the parameter estimation as a classification problem, the proposed approach considers the parameter estimation as a regression problem and can significantly reduce the complexity of networks. Simulation results demonstrate that the proposed learning-based method can improve the precision for the mixed source localization. Moreover, the proposed method is robust to the off-grid parameters and the different numbers of mixed sources.

Recently, deep unfolded networks have been widely utilized in direction of arrival (DOA) estimation due to the reduced computational complexity and improved estimation accuracy. However, few consider the nested array for off-grid DOA estimation, where the estimated DOAs are not on the prespecified grids. In this paper, we propose a deep unfolded FOCal Underdetermined System Solver (FOCUSS) network and a deep unfolded Alternating Direction Method of Multiplies (ADMM) to address the problem, which respectively aim to improve estimation accuracy and further reduce computational complexity. We first apply first-order Taylor expansion and vectorize the covariance matrix into a real-valued single snapshot for network input. We then train the proposed networks to obtain on-grid DOA spatial spectrum and off-grid values, where the off-grid DOA estimation is calculated by the peaks of on-grid DOA spatial spectrum and corresponding off-grid values. We demonstrate that the proposed networks with interpretable parameters can accelerate the convergence rate and achieve better generalization. Simulations verify the performance of proposed networks in comparison with the existing methods.

The generation of 3D models from a single image has recently received much attention, based on which point cloud generation methods have been developed. However, most current 3D reconstruction methods only work for relatively pure backgrounds, which limits their applications on real images. Meanwhile, more fine-grained details are required to provide finer models. This paper proposes an end-toend efficient generation network, which is composed of an encoder, a 2D-3D fusion module and a decoder. First, a single-object image and a nearest-shape retrieval from ShapeNet are fed into the network; then, the two encoders are integrated adaptively according to their information integrity, followed by the decoder to obtain fine-grained point clouds. The point cloud from the nearest shape effectively instructs the generation of finer point clouds. To have a consistent spatial distribution from multi-view observations, our algorithm adopts projection loss as an additional supervisor. Experiments on complex and pure background images show that our method attains state-of-the-art accuracy compared to volumetric and point set generation methods, particularly toward fine-grained details, and it works well for both complex backgrounds and multiple view angles.

The absence of publicly available, large-scale, high-quality datasets for Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) has significantly hindered the application of rapidly advancing deep learning techniques, which hold huge potential to unlock new capabilities in this field. This is primarily because collecting large volumes of diverse target samples from SAR images is prohibitively expensive, largely due to privacy concerns, the characteristics of microwave radar imagery perception, and the need for specialized expertise in data annotation. Throughout the history of SAR ATR research, there have been only a number of small datasets, mainly including targets like ships, airplanes, buildings, etc. There is only one vehicle dataset MSTAR collected in the 1990s, which has been a valuable source for SAR ATR. To fill this gap, this paper introduces a large-scale, new dataset named ATRNet-STAR with 40 different vehicle categories collected under various realistic imaging conditions and scenes. It marks a substantial advancement in dataset scale and diversity, comprising over 190,000 well-annotated samples—$10\times$ larger than its predecessor, the famous MSTAR. Building such a large dataset is a challenging task, and the data collection scheme will be detailed. Second, we illustrate the value of ATRNet-STAR via extensively evaluating the performance of 15 representative methods with 7 different experimental settings on challenging classification and detection benchmarks derived from the dataset. Finally, based on our extensive experiments, we identify valuable insights for SAR ATR and discuss potential future research directions in this field. We hope that the scale, diversity, and benchmark of ATRNet-STAR can significantly facilitate the advancement of SAR ATR.

Current methods synthetic aperture radar–automatic target recognition (SAR-ATR) research methods still struggle with overfitting due to small amounts of training data, as well as black-box opacity and high computational requirements. Unmanned aerial vehicles (UAVs), as the mainstream means of acquiring SAR data, place higher requirements on ATR algorithms due to their flexible maneuvering characteristics. Our work starts by studying the electromagnetic (EM) backscattering mechanism and the physical properties of SAR. We construct a heterogeneous graph for the first time to fully exploit both the EM scattering (EMS) information of the target components and their interactions. Moreover, the multilevel multihead attention mechanism is introduced to the graph net to learn features from various topological structure levels. Additionally, we include a convolutional neural network (CNN)-based feature extraction net to replenish intuitive visual features. The above two nets form the lightweight dual-stream framework (LDSF). LDSF uses a feature fusion subnetwork (FFS) to adaptively fuse the dual-stream features to maximize the final classification performance. The experiments use two more rigorous evaluation protocols on MSTAR and OpenSARShip, namely, once-for-all (OFA) and less-for-more (LFM), which can rigorously assess the efficacy and generalization capability of the algorithms. The superiority of LDSF is verified.