In this letter, a joint tracking and classification (JTC) algorithm for multiple extended target (ET) is proposed based on the random matrix model (RMM) and trajectory probability hypothesis density (TPHD) filter. In the proposed algorithm, the extension state of the ET is represented by a random matrix, with the target size serving as the basis for classification. The joint recursion of the multi-trajectory density and class probabilities is constructed and implemented within the extended target TPHD filtering framework, which enables the simultaneous estimation of the extended targets' states, classes, and trajectories. Additionally, an optimal sub-pattern assignment (OSPA) distance based on intersection over union (IOU) is developed to evaluate the accuracy in estimating the contours of the targets. Simulation results demonstrate that the proposed algorithm achieves higher tracking performance.

A lightweight SAR ship detection algorithm is necessary to further meet the demands of military applications. This paper proposes an end-to-end efficient SAR ship target detection algorithm based on RT-DETR called YO-DETR. In the feature extraction network, a CNN-based backbone network is used to replace the transformer-based encoder structure, which retains the original feature extraction capability while reducing the number of parameters. Additionally, in order to retain the characteristic of long-distance feature dependency in transformers, the IRMB module is incorporated into the CNN network to enhance long-distance feature interaction. Finally, the introduction of the decoder head reduces the additional time overhead of traditional NMS during inference. Ultimately, the YO-DETR method achieves a 98.2% mAP on the SSDD dataset with only 5.37M parameters and 10.5M weight size, while the FPS (when batch size is set to 32) also achieves 220.4.

The goal of hyperspectral image fusion (HIF) is to reconstruct high spatial resolution hyperspectral images (HR-HSI) via fusing low spatial resolution hyperspectral images (LR-HSI) and high spatial resolution multispectral images (HR-MSI) without loss of spatial and spectral information. Most existing HIF methods are designed based on the assumption that the observation models are known, which is unrealistic in many scenarios. To address this blind HIF problem, we propose a deep learning-based method that optimizes the observation model and fusion processes iteratively and alternatively during the reconstruction to enforce bidirectional data consistency, which leads to better spatial and spectral accuracy. However, general deep neural network inherently suffers from information loss, preventing us to achieve this bidirectional data consistency. To settle this problem, we enhance the blind HIF algorithm by making part of the deep neural network invertible via applying a slightly modified spectral normalization to the weights of the network. Furthermore, in order to reduce spatial distortion and feature redundancy, we introduce a Content-Aware ReAssembly of FEatures module and an SE-ResBlock model to our network. The former module helps to boost the fusion performance, while the latter make our model more compact. Experiments demonstrate that our model performs favorably against compared methods in terms of both nonblind HIF fusion and semiblind HIF fusion.

Superpixel segmentation is essential to the rapid information extraction and image interpretation. In this paper, we develop a superpixel segmentation method for polarimetric synthetic aperture radar (PolSAR) images, utilizing minimum spanning tree algorithm (MST) to achieve a hierarchy of superpixels. Thereinto, the revised Wishart distance and region intensity distance is applied to accurately measure the dissimilarity between two neighboring pixels. The proposed method can generate superpixels of different scales in real time, so it has significant application value. The performance of the proposed method is validated on experimental PolSAR dataset from the ESAR system.

In recent years, deep learning has made breakthroughs in the field of computer vision, the single-stage detection algorithm represented by You Only Look Once (YOLO) has achieved satisfying detection results in SAR ship target detection. For the multi-scale problem of SAR ship targets in complex scenes, we proposed an improved YOLOv5 detection method using Convolutional Block Attention Module (CBAM) and Bidirectional Feature Pyramid Network (BiFPN). The CBAM module and BiFPN are added in YOLOv5 so that it can fully learn the feature information of space and channel dimensions, and enhance information fusion transfer between multi-scale targets. Experiments on our dataset show that the proposed YOLOv5 algorithm achieves 92.8% Average Precision (AP), which gains a 1.9% improvement in AP compared to the standard YOLOv5 algorithm in SAR ship target detection. The problem of missed detection of multi-scale targets is well solved.

Current synthetic aperture radar (SAR) ship detectors achieve excellent performance with sufficient samples while encountering degraded results when the sensors and imaging conditions change. The mismatch of view, shape, and illumination inevitably result in the variations of feature distribution between source domain and target domain, which will lead to detection performance degradation. Therefore, devising a detector with well transferability to new domains remains a challenging issue. To this end, this letter proposes a novel domain adaptive YOLOv5 framework for cross-domain SAR ship detection, which is composed of the following keypoints: 1) a cross-domain co- attention feature correlation module, which models spatial and semantic interdependencies by capturing pixel correspondence between source and target domain in a bidirectional way; 2) a multilevel feature alignment module, which constrains the inter-domain difference of features from different scales by inserting three domain classifiers; and 3) teacher–student mutual learning, which makes full use of unlabeled target data and iteratively generates higher-quality pseudo-labels, thus further improving a teacher model with narrowed domain gap. Model performance is evaluated on three SAR ship datasets, and comprehensive results demonstrate the superiority of our method on multiple domain transfer scenarios, i.e., cross resolution, cross-sensor adaptation, and cross-resolution adaptation under the same sensor.

This paper addresses the problem of joint tracking and classification (JTC) of a single extended target with a complex shape. To describe this complex shape, the spatial extent state is first modeled by star-convex shape via a random hypersurface model (RHM), and then used as feature information for target classification. The target state is modeled by two vectors to alleviate the influence of the high-dimensional state space and the severely nonlinear observation model on target state estimation, while the Euclidean distance metric of the normalized Fourier descriptors is applied to obtain the analytical solution of the updated class probability. Consequently, the resulting method is called the “JTC-RHM method.” Besides, the proposed JTC-RHM is integrated into a Bernoulli filter framework to solve the JTC of a single extended target in the presence of detection uncertainty and clutter, resulting in a JTC-RHM-Ber filter. Specifically, the recursive expressions of this filter are derived. Simulations indicate that: (1) the proposed JTC-RHM method can classify the targets with complex shapes and similar sizes more correctly, compared with the JTC method based on the random matrix model; (2) the proposed method performs better in target state estimation than the star-convex RHM based extended target tracking method; (3) the proposed JTC-RHM-Ber filter has a promising performance in state detection and estimation, and can achieve target classification correctly.

Deep neural networks have achieved great progress in 3D scene understanding. However, recent methods mainly focused on objects with canonical orientations in contrast with random postures in reality. In this letter, we propose a hierarchical neural network, named Local Frame Network (LFNet), based on the local rotation invariant coordinate frame for robust point cloud analysis. The local point patches in different orientated objects are transformed into an identical distribution based on this coordinate frame, and the transformed coordinates are taken as input features to eliminate the influence of rotations at the input level. Meanwhile, a discrete convolution operator is defined in the constructed coordinate frame to extract rotation invariant features from local patches, which can further remove the influence of rotations at the convolution level. Moreover, a Spatial Feature Encoder (SFE) module is utilized to perceive the spatial structure of the local region. Mathematical analysis and experimental results on two public datasets demonstrate that the proposed method can eliminate the influence of rotations without data augmentation and outperforms other state-of-the-art methods.

In this article, a new interpretation of the polar format algorithm (PFA) for general bistatic spotlight synthetic aperture radar (SAR) imaging is presented. From the viewpoint of 2-D decoupling, we examine the commonly adopted implementation of the PFA, i.e., the separable 1-D range and azimuth resampling procedures for their roles in range cell migration (RCM) correction, respectively. Utilizing this new formulation, we analyze the effect of range and azimuth resampling on the residual 2-D phase error and reveal the inherent structure characteristics of the residual 2-D phase error in the wavenumber domain. By exploiting the available a priori knowledge on the phase error structure, a structure-aided 2-D autofocus approach to refocus the defocused PFA imagery is proposed. The proposed approach fully exploits the potentiality of the available data and the a priori knowledge about the phase error that need to estimate, so the accuracy of the residual 2-D phase error estimation and correction can be greatly improved. Finally, experimental results are presented to show the effectiveness of the proposed approach.

Object detection for optical remote sensing images have undergone rapid development in recent years, due to the advanced techniques of deep learning. However, the diverse objects with different scales and aspect ratios increase the difficulty of detection. In this paper, we improve the detection performance by applying guided anchor generation and feature fusion. In specific, the guided anchoring block directly predicts the positions and shapes of the anchor boxes, taking the place of sliding windows and preset shapes in the original region proposal network (RPN). Then, it can obtain diverse and adaptable anchor boxes, avoiding the generation of redundant and monotonous anchor boxes. In addition, the RoI features obtained from different pyramid levels are fused together to predict the categories and locations of the objects. The experiments conducted on DOTA dataset demonstrate the effectiveness of the proposed method.

The filtered backprojection (FBP) algorithm is a popular choice for complicated trajectory synthetic aperture radar (SAR) image formation processing due to its inherent nonlinear motion compensation capability. However, how to efficiently refocus the defocused FBP imagery when the motion measurement is not accurate enough is still a challenging problem. In this paper, a new interpretation of the FBP derivation is presented from the Fourier transform point of view. Based on this new viewpoint, the property of the residual 2-D phase error in FBP imagery is analyzed in detail. Then, by incorporating the derived a priori knowledge on the 2-D phase error, an accurate and efficient 2-D autofocus approach is proposed. This new approach performs the parameter estimation in a dimension-reduced parameter subspace by exploiting the a priori analytical structure of the 2-D phase error, therefore it possesses much higher accuracy and efficiency than the conventional blind methods. Finally, experimental results clearly demonstrate the effectiveness and robustness of the proposed method.

In stepped-frequency (SF) radar systems, the high-order phase error due to target motion in each burst can blur the range profile seriously. The methods based on motion parameter estimation perform well in most cases. However, the motion parameter estimation is time consuming, which challenges the real-time imaging of a radar target. To solve this problem, a new easy-to-implement variant of SF signal format is proposed in this paper. By employing the new signal, the high-order phase error can be simply eliminated, and the motion parameter estimation is unnecessary, which makes the imaging efficient. The effectiveness of the proposed method is illustrated and analyzed with both simulations and real data.

The extended Kalman filter (EKF) is well known as a state estimation method for a nonlinear system and has been used to train a multilayered neural network (MNN) by augmenting the state with unknown connecting weights. However, EKF has the inherent drawbacks such as instability due to linearization and costly calculation of Jacobian matrices, and its performance degrades greatly, especially when the nonlinearity is severe. In this letter, first a more robust learning algorithm for an MNN-based on unscented Kalman filter (UKF) is derived. Since it gives a more accurate estimate of the linkweights, the convergence performance is improved. The algorithm is then extended further to develop a NN-aided UKF for nonlinear state estimation. The NN in this algorithm is used to approximate the uncertainty of the system model due to mismodeling, extreme nonlinearities, etc. The UKF is used for both NN online training and state estimation simultaneously. Simulation results show that the new algorithm is very effective and is closer to optimal fashion in nonlinear filtering compared with traditional methods.