Currently, a convolutional neural network (CNN) combined with a Transformer has been widely applied in synthetic aperture radar (SAR) image object detection tasks. The CNN-Transformer hybrid models have the ability to model both local and global features simultaneously; however, several challenges remain. On the one hand, for large-size images to be processed, the multihead self-attention (MHSA) mechanism in Transformer aggregates feature information from the entire image, which results in high training costs and makes target features susceptible to interference from clutter features. On the other hand, the characteristics of SAR imagery are not utilized sufficiently. One special characteristic is that the gradient magnitude map of SAR imagery could reflect the target location and contours better than an intensity image in some cases and has been utilized in some existing detection methods. However, the current way of introducing it into the detection model either introduces too many parameters or loses the detailed information in it. To address these limitations, we propose a gradient-guided dual-stream you only look once (YOLO) detector combined with graph convolution. First, we embed the multihead grapher (MHGrapher) into the YOLO detector. MHGrapher is designed to establish long-range dependencies between features with fewer parameters compared with MHSA; meanwhile, it could suppress the long-range interference clutter features. Furthermore, we explore the hybrid embedding position of MHGrapher to ensure effective collaboration between graph convolution and convolutional feature extraction. Moreover, to introduce the gradient magnitude maps of SAR imagery into the model efficiently, a lightweight sub-network producing high-resolution gradient feature maps is added into the model to form a dual-stream network. The dual streams can extract features from both the intensity images and the gradient magnitude maps, and these features are then fused using the proposed intensity and gradient feature fusion module within a unified framework to enhance the ability of the model to locate and extract target features. The proposed method achieves 0.906 mAP@50 on the SARDet-100K dataset, demonstrating its effectiveness.

Imbalanced classification remains a considerable challenge in machine learning, as conventional models often fail to generalize well for underrepresented tail classes. One effective approach is to leverage the well-represented knowledge of head classes to adjust the tail-class distributions in feature space, a process known as distribution calibration. Despite its effectiveness, existing distribution calibration methods have two key limitations: (i) oversimplified Gaussian assumptions neglect the complexity and diversity of real-world data distributions, leading to suboptimal calibration; (ii) heuristic head-class selection strategies (e.g., top-k) lack adaptiveness, which may either overlook valuable information from unselected head classes or amplify irrelevant information from dissimilar ones. To address these issues, we propose Gaussian Mixture Model-oriented Distribution Calibration (GMDC), a plug-and-play framework that models each class-conditional distribution with a Gaussian Mixture Model (GMM) to better capture complex feature distributions. Instead of selecting top-k head classes and assigning heuristic transfer weights, GMDC computes a dense head-to-tail transport plan via a two-level optimal transport (OT) formulation. Specifically, GMDC solves a global OT to compute a class-level transfer matrix from head to tail classes under a cost that captures both mean and geometric statistics, and further performs local OT to align mixture components for component-level calibration. GMDC introduces no additional trainable modules, making it a lightweight post-hoc calibration module. Extensive experiments on four benchmark datasets and two real-world datasets demonstrate consistent improvements.

Synthetic Aperture Ladar (SAL) provides high-resolution, high-data-rate, and detailed imaging for remote sensing. However, its short wavelength makes SAL systems highly sensitive to vibrations, introducing Doppler frequency shifts and range cell migration that degrade image quality, particularly for extended targets. Traditional vibration compensation methods often face limitations in challenging scenes with severe vibration conditions or when strong scattering points are absent. To address these challenges, a second-order vibration error model is firstly developed to characterize the time-varying errors within each tunable period. Then, a physicallyinformed deep neural network is designed to estimate the vibration coefficients through its encoder, which are then used by physical layers in the decoder to correct the errors. By combining the physical model with a datadriven approach, the proposed method can mitigate severe vibration-induced errors and reduce range cell migration without relying on strong scattering points. Additionally, integration of the physical layers makes the decoder being non-parametric, thus simplifies the network training. Numerical results validate the method's effectiveness and its superiority over traditional spectral correlation algorithm, demonstrating its potential for high-resolution SAL imaging.

In modern compact linear arrays, strong electromagnetic mutual coupling driven by space-wave interactions and surface-wave leakage, along with inherent radio-frequency gain-phase mismatches, severely distorts active element patterns and degrades super-resolution direction-of-arrival (DOA) estimation. Traditional hardware-centric decoupling techniques inevitably increase design complexity, introduce insertion losses, and are inapplicable to already-deployed systems. To overcome these barriers, this paper proposes a hardware-free decoupling approach: Knowledge-Aided Sparse Bayesian Multi-Error Calibration (KASB-MEC). By computationally compensating for electromagnetic distortions at the digital baseband, KASB-MEC effectively neutralizes adjacent scattering and channel inconsistencies, which is physically equivalent to terminating elements with virtual impedances. To resolve the phase-DOA ambiguity and the highly non-convex optimization landscape, the framework introduces a robust soft-constraint mechanism using von Mises and Rician priors. This strategy structurally decouples spatial angles from array errors, completely bypassing fragile hard-constraints. Furthermore, an adaptive dictionary learning strategy breaks the computational bottleneck of high-dimensional Bayesian inference. Validated by extensive numerical simulations and practical anechoic chamber measurements, KASB-MEC achieves fast, accurate, and robust multi-error calibration, offering a plug-and-play, software-defined calibration upgrade for real-time super-resolution sensing arrays.

In practical applications, inverse synthetic aperture radar (ISAR) imaging consistently faces two primary challenges: 1) Low signal-to-noise ratio (SNR); 2) High maneuvering motion. For challenge 1), conventional non-parametric cascaded motion compensation methods often suffer from insufficient envelope alignment accuracy. Regarding the second, existing parametric methods frequently exhibit low accuracy and efficiency due to the complex coupling of translational errors and spatial-variant phase errors (SVPE). To address these limitations, this paper proposes an integrated track-before-detect ISAR (ITBD-ISAR) algorithm designed to achieve 2D joint full motion compensation (2D JFMC). The main contributions are outlined below: 1)Unified 2D Joint Full Motion Compensation Framework: A unified framework is established to effectively compensate for translational motion errors, range and azimuth spatial-variant errors, and migration through range cells, yielding well focused ISAR images. 2) Robust Low-SNR Processing: A TBD-driven motion trajectory inversion algorithm (TBD-MTIA) is introduced to enhance performance under low SNR conditions. By utilizing multi-frame joint processing and energy-weighted allocation, this algorithm effectively suppresses noise, ensuring stable and accurate motion trajectory estimation. 3) Decoupling and Dimensionality Reduction for High Maneuvering Motion: A sub-aperture-fusion-based motion order determination and parameter estimation (SAF-MODPE) method is proposed to extract translational motion information. Leveraging a ”global-to-local” strategy, it decouples translational and rotational motions while reducing parameter dimensionality via sub-aperture decomposition. This enables precise translational compensation and accurate tracking. Combined with TBD and ISAR imaging, the proposed method achieves 2D JFMC, resulting in well-focused imagery. Simulated and real data experiments demonstrate the robustness and superiority of the proposed method at low SNR conditions.

Spatiotemporal sensor biases, if unaccounted for, can significantly degrade the positioning accuracy in target tracking. Spatial bias causes sensor measurements to be biased, and temporal bias leads to inaccurate measurement timestamps. Meanwhile, it is difficult to accurately obtain the sensor positions in multisensor tracking systems. This paper addresses the problem of bearings-only tracking using asynchronous sensors, considering both spatiotemporal biases and sensor position errors. A novel method is proposed that simultaneously compensates for these biases and performs target tracking. Simulation results demonstrate the superiority of the proposed approach in terms of target position estimation, and further confirm its ability to attain the posterior Cramér–Rao lower bound (PCRLB) performance under distance-dependent measurement noise.

Deep learning methods have been widely used in Synthetic Aperture Radar Automatic Target Recognition (SAR ATR). However, challenges remain due to limited SAR data and computational constraints on mobile devices, which hinder model training and deployment. In this paper, we propose a Multi-scale Perception and Reference Attention Network (MPRANet) for lightweight SAR ATR, which is a hybrid structure combining convolutional networks and transformers, built upon the ShuffleNetV2 network. Specifically, MPRANet introduces two key improvements compared to the CNN-based ShuffleNetV2. Firstly, we replace the depthwise convolutions (DWConv) in the downsampling and basic units of ShuffleNetV2 with the Multi-scale Parameter-Shared Convolution (MPConv) module. MPConv enables the extraction of multi-scale features of SAR targets with almost no additional parameters, thereby enhancing the network’s feature extraction capabilities. Secondly, we propose a lightweight Reference Attention Transformer (RAformer) to capture global information, addressing the issue of insufficient channel feature interaction in ShuffleNetV2. In RAformer, a Local Linear Mapping Unit (LMU) is designed to perform linear mappings, reducing the introduction of redundant features while ensuring its lightweight and efficient nature. RAformer contains two modules: the Reference Vector Attention (RVA) module, which efficiently models attention relationships, and the Lightweight Feedforward Neural Network (LW-FFN) module, which enhances the network’s ability to capture nonlinear representations. We evaluated the performance of MPRANet using publicly available SAR datasets, including the MSTAR dataset, OpenSARShip dataset, and SAR-AIRcraft-1.0 dataset. The experimental results demonstrate that MPRANet consistently achieves superior recognition performance compared to other lightweight networks of similar complexity.

SAR ship semantic segmentation holds significant value for maritime traffic monitoring, military reconnaissance, and other related fields. However, fully-supervised segmentation methods require high annotation costs and their performance degrades sharply when labels are scarce. As a mainstream semi-supervised paradigm, co-training can leverage unlabeled data through multi-view pseudo supervision, yet it often suffers from view homogeneity. To address this issue, we propose a heterogeneous co-training framework that integrates spiking neural networks (SNNs) and artificial neural networks (ANNs) to establish complementary learning views. First, a dual-branch architecture of SNN and ANN is constructed to inherently create heterogeneous views at the feature representation level, leveraging the temporal dynamics of SNNs and the continuous representation capability of ANNs to enrich the feature space. Second, at the network architecture level, CNN and spiking Transformer are adopted as the backbones for the ANN and SNN branches, respectively, enhancing both intra-model and inter-model diversity. Moreover, to address the spiking Transformer’s limitations in fine-grained segmentation and multi-scale adaptation, we embed a lightweight spiking auxiliary flow within the SNN branch. Extensive experiments demonstrate that the proposed framework outperforms the fully-supervised SNN baseline by 6.81 and 7.93 percentage points in mIoU on SSDD and HRSID under the 1/16 setting, while the retained SNN branch consumes only 54.5% of the theoretical inference energy of ResNet50-based ANN methods.

This work introduces a new condition on bifunctions for equilibrium problems, termed the uniform-type condition, which is weaker than the Lipschitz-type conditions. Based on this new framework, we analyze the convergence of the subgradient extragradient method with nonmonotone step sizes for solving equilibrium problems involving pseudomonotone operators in real Hilbert spaces. Notably, the proposed analytical approach is novel. Numerical experiments, including applications of the proposed algorithm to the Nash-Cournot equilibrium model and a traffic network equilibrium problem, further confirm its efficiency and broad applicability.

Multimodal remote sensing combines optical and synthetic aperture radar (SAR) imagery to improve perception, yet real deployments face spatially varying degradations (e.g., clouds, low light, sensor interference) that can corrupt fusion. To make robustness measurable, we introduce a controlled mixed-severity setting in which only the optical stream is synthetically cloud-degraded while SAR remains intact, providing a standardized testbed for evaluating multimodal detection under modality imbalance. We further present CAIR-Net, a reliability-aware information routing network that follows a denoise-then-fuse principle: a Local Reliability Modulation (LRM) module learns soft, spatial reliability maps to suppress degraded regions before cross-modal interaction, and a Global Information Selection Mechanism (GISM) performs confidence-aware expert routing across optical, fused, and SAR experts. On the mixed-severity benchmark, CAIR-Net consistently outperforms strong unimodal and fusion baselines and exhibits a substantially smaller performance drop under severe clouds. These results indicate that explicit reliability modeling and quality-guided routing provide a practical path toward robust multimodal detection when one modality is partially or nearly completely occluded.

The spatial bias of sensor in the process of target tracking and the temporal bias between the time axis of each sensor and the absolute time axis, if not accounted for, can seriously affect the positioning accuracy. Meanwhile, the sensor position reported by Global Positioning System (GPS) is not accurate. In this paper, the problem of angles-only target motion analysis (TMA) by asynchronous sensors is studied in the presence of spatiotemporal bias and sensor position error. A new target tracking method is proposed by taking the target state, spatiotemporal bias and sensor position as the augmented state vector. Using the filter concept and the minimum mean square error (MMSE) criterion for real-time processing, the augmented state vector can be estimated simultaneously. Simulation results show the superiority of the proposed algorithm for target position estimation, and verify the effectiveness of the proposed in achieving the Posterior Cram & eacute;r-Rao lower bound (PCRLB) performance under the distance-dependent measurement noise.

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen compositions by recombining learned state and object primitives. Many recent methods can be viewed as following a two-stage paradigm that first disentangles visual primitives and then aligns them with their textual counterparts. However, this paradigm suffers from two limitations: coarse global embeddings limit fine-grained cross-modal alignment, while attention-based disentanglement tends to spread over irrelevant regions and weakens primitive separability. To address these issues, we propose a new CZSL framework built on adaptive-weighted optimal transport. Specifically, we design a fine-grained alignment module that represents visual and textual concepts as weighted point sets and performs transport-based matching for more discriminative cross-modal alignment. We further develop a doubly constrained transport-based attention module to reduce attention spread and improve primitive disentanglement. Central to the framework is an adaptive-marginal principle that relaxes the uniform-weight assumption in standard optimal transport, enabling the model to emphasize informative tokens and local regions while reducing the influence of irrelevant responses. Extensive experiments on MIT-States, UT-Zappos, and C-GQA demonstrate that the proposed method achieves improved compositional generalization and competitive performance in both closed-and open-world settings.

Synthetic aperture radar (SAR) is an active detection system integral to automatic target recognition (ATR). The unique scattering and geometric features of SAR images are highly sensitive to target aspect angles. Traditional SAR ATR methods generally assume that the dataset is balanced, meaning that the various types of targets have full target aspect coverage. However, when the aspect angles are incomplete, the recognition performance sig nificantly deteriorates. Given this issue, in this paper, we propose a SAR image generation method based on label-aspect-conditioned diffusion model (SAR-LADM). Its principle and structure are tailored specifically to SAR target characteristics for SAR image generation and imbalanced recognition tasks. Specifically, we first introduce the diffusion model to generate SAR images. The robustness of its training process enhances the reliability of the generated samples. Subsequently, the conditionally controlled Transformer encoder (CCTE) is designed to facil itate controllable SAR image generation. This is achieved by intelligently integrating and utilizing the category label and aspect of targets in the Transformer encoder. Finally, the SAR-LADM is applied to the imbalanced SAR ATR, and addresses the issue of imbalance by generating target samples with missing aspects. Experiments on the moving and stationary target acquisition and recognition (MSTAR) and aircraft datasets, involving generation and imbalanced recognition tasks, demonstrate that the proposed method maintains stable generation quality and significantly enhances SAR ATR performance in imbalanced scenarios.

High-resolution range profile (HRRP) target recognition has advanced with deep learning, yet most existing methods rely primarily on data-driven feature extraction and still face challenges in physical interpretability and noise robustness. To address these issues, this paper proposes a physics-guided dual-branch fusion (PGDBF) model. It consists of two parallel branches: a data-driven branch extracts discriminative features from raw HRRPs, while a physics-guided branch estimates sparse peak parameters (position and intensity) and reconstructs the signal envelope under sparsity constraints. Cross-attention adaptively fuses the two branches. Experiments on measured ten-class aircraft HRRP data show that PGDBF achieves higher accuracy and improved robustness under additive Gaussian noise in the evaluated fixed-route scenario. Visualizations confirm that the estimated peaks align with dominant HRRP energy, linking model variables to physically meaningful peak locations and intensities. These results suggest that integrating explicit peak parameter estimation with data-driven learning is a promising direction for improving HRRP recognition robustness and interpretability under low-SNR conditions.

This research introduces an innovative and high-performance multi-objective parallel feasible direction approach for hypergraph partitioning, utilizing rank-two semidefinite programming relaxation frameworks. The algorithm generates high-quality initial partitions through strategies such as recursive bisection, and multi-domain adjustment. These partitions are further optimized using multi-partition synchronized parallel clustering optimization algorithms and embedded into a multilevel framework as initial partitions. The numerical results demonstrate superior performance in both solution quality and computational efficiency across diverse test cases. Our method achieves approximately 5% average improvement in bipartitioning quality, with up to 44.4% enhancement over K-SpecPart on instances like ibm06. The algorithm also shows better scalability than KaHyPar, with slower runtime growth as partition count increases. Parameter sensitivity analysis confirms consistent performance across different configurations, highlighting remarkable robustness. Additionally, ablation studies validate the effectiveness of each component of the proposed method. In summary, the presented algorithm effectively balances partitioning quality with computational efficiency, offering a competitive solution for large-scale hypergraph partitioning.

Weak target detection commonly encounters intense clutter interference, which overshadows weak signals and complicates the task. Taking advantage of the powerful data mining capability of neural networks, more and more deep learning-based methods are applied to radar target detection. Among the approaches, those founded upon unsupervised learning methodologies exhibit remarkable merit because they dispense with the requirement for target samples within the training step, making them highly applicable in practical target detecting scenarios. However, existing methods suffer from limitations in leveraging the range-Doppler (R-D) two-dimensional correlation and finely modeling in multiple clutter scenarios. In this paper, an unsupervised Transformer-based detector (TrDet) is proposed to break through the boundary of modeling capability. First, with the designed two-dimensional position embedding (2-DPE) and global query embedding (GQE) techniques, an unsupervised training strategy for R-D spectrum based on Transformer framework is utilized to achieve refined clutter modeling. Then, radar target detection is formulated as an out-of-distribution (OOD) detection task to mitigate clutter interference. Moreover, the masked variational Transformer-based detector (MVTrDet) is further proposed to prevent target information leakage when the target is in close proximity to the clutter in Doppler domain. Compared with several relative algorithms, our proposed methods are better suited for radar target detection in complex clutter environments. The experimental results derived from both measured data and simulated data verify the effectiveness of our proposed methods.

Detecting high-speed range-spread targets poses significant challenges due to interpulse range migration (RM) and target energy dispersion across multiple range bins. This article addresses these challenges by proposing a joint velocity-robust keystone transform and range-spread target detection network (VRKT-RSTDNet) algorithm. The VRKT algorithm is proposed to mitigate interpulse energy loss caused by RM. It maps ambiguous and half-ambiguous velocities into the unambiguous region through corresponding velocity compensation, performs slow-time correction and coherent integration to accumulate target energy, and reorganizes the Doppler frequency spectrum to obtain the final result. The VRKT algorithm achieves accurate RM correction and robust energy accumulation across arbitrary target velocities. For range-spread target detection, a deep neural network named RSTDNet is proposed, which integrates multiscale convolutions and residual connections for multiscale hierarchical feature extraction. The training dataset is generated based on the electromagnetic scattering model of a typical target, enabling RSTDNet to learn priors for effective detection across varying velocities, attitudes, and signal-to-noise ratios. Simulation results demonstrate that the VRKT algorithm outperforms conventional KT methods in half-ambiguous velocity regions, achieving an integration gain improvement. Meanwhile, compared to classical range-spread target detection algorithms, such as energy integration, order statistics, M-out-of-N, and scattering density dependent generalized likelihood ratio test, RSTDNet demonstrates superior detection performance across various target attitudes. Furthermore, the overall VRKT-RSTDNet method is validated for high-speed range-spread target detection.

The initial partitioning phase is crucial in the multi-level framework, generating the primary solution that underpins all subsequent refinement stages. State-of-the-art hypergraph partitioners follow the heuristic strategy and the ILP for initial partitioning. However, they suffer from two limitations: (a) the inherent stochasticity of heuristic methods and (b) the computational overhead of exact ILP solvers when the hypergraph grows. In this work, we introduce a transformative approach by reformulating hypergraph partitioning as a continuous, angle-based unconstrained optimization problem that systematically incorporates global structural information through the hypergraph Laplacian matrix. Our formulation simultaneously overcomes the stochastic limitations of heuristic methods while avoiding the computational bottlenecks of exact ILP solvers. Furthermore, we present an advanced gradient method approximate optimal stepsize (GM AOS)-based partitioning strategy. It integrates a degree-order-aware scheme, the GM AOS algorithm, angle-based partitioning, and block-paired GM AOS optimization, collectively improving the quality of the partitioning process. Experimental results with extensive evaluations show that for some imbalance factors, our algorithm can substantially improve the cutsize by an average of 15% with a competitive runtime when compared to the SOTA partitioners (hMetis, KaHyPar, and TritonPart).