Due to the reduction of imaging accuracy caused by the approximations of signal models, traditional synthetic aperture radar(SAR) moving target motion parameter estimation methods based on frequency domain imaging algorithms may suffer from the problem of accuracy reduction. To solve this problem, a novel moving target motion parameter estimation scheme is proposed for dual-channel SAR based on the time domain back projection(BP) algorithm. First, the BP imaging model of a moving target is constructed for the dual-channel SAR, and the focus position and the phase response of the moving target are analyzed. We show that the radial velocity of the moving target is proportional to the center frequency of the azimuth wavenumber spectrum, which can be used to estimate the radial velocity. Afterwards, the displaced phase center antenna based on BP is deduced for the clutter suppression, and the constant false alarm rate detector is used to detect moving targets. Then, the azimuth offset of the moving target between the two sub-aperture images is used to estimate the azimuth velocity, since it is proportional to the azimuth velocity. Meanwhile, a modified refocussing method is applied for a more accurate azimuth velocity estimation. Furthermore, the slant-range velocity is calculated by the geometric relationship among the radial velocity, azimuth velocity, and the slant-range velocity. The simulation and semi-physical simulation experiments verify that the proposed scheme can achieve higher accuracy in motion parameter estimation than the method based on the frequency domain imaging algorithm in both the side-looking and squint-looking dual-channel SAR.

Current 3-D object detection technology for autonomous driving usually cannot efficiently utilize local sensitive points. Meanwhile, contextual feature extracted from a object is not sufficient, which easily leads to deteriorated detection accuracy of the final object estimation. For the problems, a point–voxel-based 3-D dynamic object detection algorithm is proposed. First, local points are grouped with a camera frustum. Then, the global feature extracted by the submanifold 3-D voxel CNNs is aggregated into frustum key points. Second, a module of vector pool with feature aggregation is used to aggregate multiscale features of the point cloud. Moreover, the frustum raw feature and BEV feature are used for feature extension. Subsequently, the fine multiscale feature extracted from the point cloud is used as input to a subsequent fully convolutional network for final classification and continuous estimation of oriented 3-D boxes. The proposed method was compared with other state-of-the-art algorithms on the KITTI, Waymo, and nuScenes data sets. Experimental results showed that the proposed algorithm was better in accuracy, robustness, and generalization capabilities in 3-D dynamic object detection. Experiments on a real scenario and extensive ablation studies also demonstrated that the proposed algorithm not only effectively controls computational cost but also achieved more efficient results in 3-D object detection.

Current monocular 3D object detection algorithms generally suffer from inaccurate depth estimation, which leads to reduction of detection accuracy. The depth error from image-to-image generation for the stereo view is insignificant compared with the gap in single-image generation. Therefore, a novel pseudo-monocular 3D object detection framework is proposed, which is called Pseudo-Mono. Particularly, stereo images are brought into monocular 3D detection. Firstly, stereo images are taken as input, then a lightweight depth predictor is used to generate the depth map of input images. Secondly, the left input images obtained from stereo camera are used as subjects, which generate enhanced visual feature and multi-scale depth feature by depth indexing and feature matching probabilities, respectively. Finally, sparse anchors set by the foreground probability maps and the multi-scale feature maps are used as reference points to find the suitable initialization approach of object query. The encoded visual feature is adopted to enhance object query for enabling deep interaction between visual feature and depth feature. Compared with popular monocular 3D object detection methods, Pseudo-Mono is able to achieve richer fine-grained information without additional data input. Extensive experimental results on the datasets of KITTI, NuScenes, and MS-COCO demonstrate the generalizability and portability of the proposed method. The effectiveness and efficiency of Pseudo-Mono have been demonstrated by extensive ablation experiments. Experiments on a real vehicle platform have shown that the proposed method maintains high performance in complex real-world environments.

The wide applications of 3-D synthetic aperture radar (SAR) imaging bring higher requirements for resolution and computational efficiency. The stepped frequency linear frequency modulated signal achieves high range resolution imaging by fusing multiple sub-pulses. However, its wide sub-pulse bandwidth results in a large amount of echo data to be processed, which results in a significant increase in the time consumption of the bandwidth synthesis algorithm. To achieve fast high range resolution 3-D SAR imaging, we propose an inter-array frequency-hopping linear frequency modulated signal model and a 3-D variable carrier frequency back projection algorithm. The proposed signal model transmits only one narrow bandwidth sub-pulse with hopping carrier frequency in each array element, which allows the receiver to sample the echo at a lower frequency. The lower sampling frequency and number of sub-pulse significantly reduce the amount of echo data. The proposed algorithm not only focuses the along-track direction and the cross-track direction of SAR image, but also fuses the low range resolution imaging results obtained by each sub-pulse into a high range resolution imaging result. Benefiting from the fusion of bandwidth synthesis algorithm and imaging algorithm, the computational efficiency is greatly improved. The experimental results demonstrate that the proposed algorithm achieves the comparable resolution and imaging quality as the ideal 3-D back projection (BP) algorithm with a large bandwidth chirp signal. Moreover, the time consumption of the proposed algorithm has been reduced to only 2.25% to 3.02% of that of the advanced high range resolution 3-D BP algorithm.

The success of deep learning in synthetic aperture radar (SAR) automatic target recognition (ATR) relies on a large number of labeled samples; however, there are often wrong (noisy) labels in a large-scale dataset. In this article, we propose a loss curve-fitting-based method, which can identify the noisy labels and train the classification network effectively. We propose to model label noise by unsupervised clustering via fitting loss curve to identify whether the sample’s label is clean or noisy. Then, we train the network using augmented samples with clean labels to correct noisy labels further. The experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset prove that our proposed method can deal with the situation when training a network with different ratios of noisy labels and correct noisy labels effectively. When the noise ratio is small (40%) in the training dataset, our method can correct 97.9% of noisy labels and train the classification network with 98.8% classification accuracy. While the noise ratio is large (80%), our method can correct 78.1% of noisy labels and train the classification network with 79.6% classification accuracy.

Compressed sensing (CS) demonstrates significant potential to improve image quality in 3-D millimeter-wave imaging compared with conventional matched filtering (MF). However, existing sparsity-driven 3-D imaging algorithms always suffer from large-scale storage, excessive computational cost, and nontrivial tuning of parameters due to the huge-dimensional matrix–vector multiplication in complicated iterative optimization steps. In this article, we present a novel range migration (RM) kernel-based iterative-shrinkage thresholding network, dubbed as RMIST-Net, by combining the traditional model-based CS method and data-driven deep learning method for near-field 3-D millimeter-wave (mmW) sparse imaging. First, the measurement matrices in ISTA optimization steps are replaced by RM kernels, by which matrix–vector multiplication is converted to the Hadamard product. Then, the modified ISTA optimization is unrolled into a deep hierarchical architecture, in which all parameters are learned automatically instead of manually tuned. Subsequently, 1000 pairs of oracle images with randomly distributed targets and their corresponding echoes are simulated to train the network. A well-trained RMIST-Net produces high-quality 3-D images from range-focused echoes. Finally, we experimentally prove that RMIST-Net is capable process$512 \times 512$large-scale imaging tasks within 1 s. Besides, we compare RMIST-Net with other state-of-the-art methods in near-field 3-D imaging applications. Both simulations and real-measured experiments demonstrate that RMIST-Net produces impressive reconstruction performance while maintaining high computational speed compared with conventional and sparse imaging algorithms.

Ship classification in synthetic aperture radar (SAR) images is a fundamental and significant step in ocean surveillance. Recently, with the rise of deep learning (DL), modern abstract features from convolutional neural networks (CNNs) have hugely improved SAR ship classification accuracy. However, most existing CNN-based SAR ship classifiers overly rely on abstract features, but uncritically abandon traditional mature hand-crafted features, which may incur some challenges for further improving accuracy. Hence, this article proposes a novel DL network with histogram of oriented gradient (HOG) feature fusion (HOG-ShipCLSNet) for preferable SAR ship classification. In HOG-ShipCLSNet, four mechanisms are proposed to ensure superior classification accuracy, that is, 1) a multiscale classification mechanism (MS-CLS-Mechanism); 2) a global self-attention mechanism (GS-ATT-Mechanism); 3) a fully connected balance mechanism (FC-BAL-Mechanism); and 4) an HOG feature fusion mechanism (HOG-FF-Mechanism). We perform sufficient ablation studies to confirm the effectiveness of these four mechanisms. Finally, our experimental results on two open SAR ship datasets (OpenSARShip and FUSAR-Ship) jointly reveal that HOG-ShipCLSNet dramatically outperforms both modern CNN-based methods and traditional hand-crafted feature methods.

This article proposes a novel framework by combining a modified real-time recurrent regression (mRe³) network and a newly designed trajectory smoothing long short-term memory (LSTM) network for refocusing the ground moving target (GMT) in the synthetic aperture radar (SAR) image. The mRe^3 network that consists of a convolutional neural network (CNN) backbone and two LSTM modules is designed to track the GMT's shadow in an SAR video. Furthermore, we find that the complex trajectory obtained by the tracking network cannot directly be used for refocusing the GMT because of the estimation error. To address the abovementioned problem, a β-order total variation loss-based smoothing LSTM (TVβ-LSTM) is proposed to recover the GMT's trajectory to meet the requirement of refocusing. Besides, the effect of TVβ on the performance of smoothing LSTM is analyzed. By the experiments on simulated and real SAR videos, we find that the mRe^3 has stronger robustness and a better trajectory reconstruction precision compared with the existing tracking methods, especially for the strong interference cases. In addition, the smoothing LSTM can recover the trajectory of the GMT with higher precision and better smoothness. When β is set to 3, with the TVβ-LSTM, the center distance error of a recovered complex trajectory can be reduced from 0.82 to 0.782, while its fluctuation can be suppressed from 6 to 1 mm. By using our framework, the focused GMT with bountiful geometrical features can be obtained even for the K_a-band SAR.

Convolutional neural network (CNN) has become the mainstream method in the field of image recognition for its excellent ability to feature extraction. Most of the CNNs increase the classification accuracy for the rotational objects by imposing the network with rotation invariance or equivariance property, which causes the loss of the target's orientation information. In this work, a rotation-mapping network (RM-Net) that can achieve objects recognition and angle or orientation estimation simultaneously without additional network training is constructed. Besides, an octagona convolutional kernel is introduced to improve the network's performance. The experiments on the simulation SAR datasets show that the proposed RM-CNN can achieve state-of-the-art results in target recognition and angle estimation.

Deep neural networks have been widely applied and researched in synthetic aperture radar (SAR) object detection and achieved a great success. However, deep supervised networks heavily rely on a large amount of labeled data, while the annotation is difficult and time-consuming to obtain. But the unlabeled data are comparably easier to get. Considering that, we introduce a semi-supervised learning framework for SAR object detection, which is built via label propagation and consistent augmentation. The experiments on a SAR ship dataset prove that the introduced semi-supervised training framework can achieve higher detection performance with utilizing the unlabeled data compared with the corresponding supervised object detection network.

Shadows of ground moving targets in video synthetic aperture radar (SAR) has been found very useful in ground moving target indication (GMTI) for they can indicate the real positions of moving targets at different times, which is significant for SAR reconnaissance and surveillance. However, nearly all the shadow-based SAR GMTI methods only focused on detecting shadows in every separate frame and failed to make full use of the continuous observation ability of video SAR. In this paper, we propose to apply a deep learning-based multi-target tracking method to solve this problem and find that the FairMOT network which jointly detects and re-identifies objects in sequential frames is suitable for this task. To verify its performance, video SAR datasets that contain shadows of ground moving targets are obtained by simulation. The experiments on the simulation datasets show that the introduced network in this work can achieve a state-of-the-art result, for instance, the multiple object tracking accuracy (MOTA) can reach 83.4%.

In synthetic aperture radar (SAR) automatic target recognition, it is expensive and time-consuming to annotate the targets. Thus, training a network with a few labeled data and plenty of unlabeled data attracts attention of many researchers. In this article, we design a semisupervised learning framework including self-consistent augmentation rule, mixup-based mixture, and weighted loss, which allows a classification network to utilize unlabeled data during training and ultimately alleviates the demand of labeled data. The proposed self-consistent augmentation rule forces the samples before and after augmentation to share the same labels to utilize the unlabeled data, which can ensure the prominent effect of supervised learning part of the framework for training by balancing amounts of labeled and unlabeled samples in a minibatch, and makes the network achieve better performance. Then, a mixture method is introduced to mix the labeled, unlabeled, and augmented samples for the better involvement of label information in the mixed samples. By using cross-entropy loss for the mixed-labeled mixtures and mean-squared error loss for the mixed-unlabeled mixtures, the total loss is defined as the weighted sum of them. The experiments on the MSTAR data set and OpenSARShip data set show that the performance of the method is not only far better than the state of the art among current semisupervised-based classifiers but also near to the state of the art among the supervised learning-based networks.

Convolutional Neural Networks (CNNs) have excellent ability in image recognition, however, the requirement of a large amount of labeled dataset limits its application in the field of synthetic aperture radar (SAR) image processing. In this paper, a kernel rotational network (KR-Net) for SAR target recognition is constructed. When the labeled dataset is small, the KR-net can achieve higher classification rate than standard CNNs benefit from its inherent rotational convolution units. Also, weights sharing strategy is introduced to increase network capacity without multiplying the number of weights parameters. Meanwhile, a simple and feasible multi-branch feature converging method for the KR-Net is proposed to fuse features of rotational convolution units. Experimental results show that our network can achieve state-of-art result in the MSTAR dataset, especially when the training set is small.

Deep neural networks have been widely applied and researched in remote sensing image scene classification and achieved a great success. However, deep supervised network heavily relies on a large amount of labeled data. The annotation is difficult and time-consuming to obtain but the unlabeled data are comparably easier to get. Considering that, we introduce a semi-supervised learning framework for remote sensing image scene classification. The network is trained by a novel adaptive perturbation training method. The experiments on NWPU-RESISC45 dataset prove that the introduced semi-supervised classification method can achieve higher classification accuracy with unlabeled data compared with the corresponding supervised classifier, and the designed adaptive perturbation training can further improve the performance of the semi-supervised learning-based classification network.

For traditional ground moving target indication, multiple channels are necessary to cancel ground clutter. For high-speed aircraft, slow moving target detection is a difficult problem because ground clutter cannot be eliminated completely by channel cancellation. In this paper, we propose a method, which is implemented by single-channel SAR images and improved Faster R-CNN, to detect the moving target and the stationary target. Synthetic aperture radar image which contains amplitude and phase information is put into the neural network to detect the moving and stationary target. We make a dataset to verify the availability of the proposed method. In order to increase the credibility of the dataset, we use FEKO to calculate the target electromagnetic scattering characteristics and use measured data scattering characteristics to generate the ground echo. The simulation proves that the proposed method has good performance in moving target detection and the performance of the proposed method is better than Faster R-CNN.

Object detection in very high-resolution (VHR) remote sensing images is a fundamental and challenging problem due to the complex environments. In this paper, a precise mask region convolutional neural network (precise Mask R-CNN) is presented for object detection and instance segmentation in VHR remote sensing images. This method generates bounding boxes and segmentation masks for each instance of an object in the image. Contrary to regions of interest (RoI) Align whose sample points is pre-defined and not adaptive the size of the bin, the proposed precise RoI pooling can directly compute the two-order integral based on the continuous feature map to avoid loss of precision. The experiments on NWPU VHR-10 dataset show that the presented precise Mask R-CNN improves the accuracy of object detection and instance segmentation for VHR remote sensing images. Furthermore, it promotes the application of instance segmentation in VHR remote sensing.

In this paper, we propose to apply a deep multi-scale encoder-decoder neural network (MsEN-Net) for SAR images enhancement method based on scale-recurrent network (SRN), which consists of encoder and decoder modules trained by the coarse to fine strategy. Simulation with fixed speed errors and experiments with real data are implemented and evaluated by peak signal to noise ratio and structural similarity. Experimental results with both visual and quantitative analysis demonstrate the competitive performance of our proposed method.

Autofocus method is a vital technology for high resolution and wide swath airborne Synthetic Aperture Radar (SAR) imaging. The autofocus algorithms via phase errors estimation and traditional Antenna Phase Centers (APC) errors estimation cannot completely compensate for the phase error of each pixel for the large scene ignoring the spatial variance, which results in corrupted SAR imagery for some part of the scene. In this paper, an autofocus algorithm through precise APC errors estimation based on joint multi-region is proposed to compensate motion error for the whole scene greatly. We established an image intensity model for strong point targets in multi-region with the weight coefficient to estimate APC errors. Moreover, the partial derivative of image intensity is simplified which can easily derive higher order criterion like image sharpness and Conjugate Gradient (CG) is utilized to solve the optimization problem. The simulation and experimental examples verify the effectiveness of the proposed method compared with traditional methods.