The discernment of radar jamming signal played an important role for the downstream tasks. Great performance was achieved by the deep learning methods. Yet large amounts of labeled signals were needed. To solve the problem, a new cross-modality contrast learning method was proposed in this paper. The signal-wise hierarchy, and the image-wise hierarchy were developed to extract the features from IQ data (In-phase & quadrature) and TF-image (Time-frequency). The cross-domain features were then combined. The fused features were delivered to the learning phase. It was composed of the pre-training and the fine-tuning. The unlabeled signals were first employed to optimize the similarity loss to make the positive sample mores similar to the signal than the negative samples. The pre-trained model was then fine-tuned to the recognition task. The proposed method was demonstrated to provide impressive performance by multiple comparative studies.

High quality of synthetic aperture radar (SAR) images is an essential requirement for many applications. However. the increasingly intensive jamming deteriorated the imaging quality. It is therefore needed to achieve jamming suppression. The traditional methods were the empirical model parameterized by simple function, such as adaptive filtering and subspace projection. These methods usually relied on the detail jamming parameters to discriminate the jamming and the signal. To address these problems, a new hybrid U-shaped network (JSUnet) is proposed in this article. First, JSUnet was constructed, which is a hybrid U-shaped network of CNN and Transformer. Second, open source SAR data was obtained and used to build the dataset. Third, a supervised training was designed to train JSUnet. Finally, the trained JSUnet was used to complete the jamming suppression of interfered SAR images end-to-end. Experiments demonstrated the powerful performance of the proposed method and compared it with traditional filtering method and other deep learning-based methods. The proposed method achieved the state- of-the-art result.

Detecting ships in inshore SAR imagery is challenging due to strong clutter, high maritime activity, and weak target reflections. To address these issues, we propose CSDet, a novel Clutter Suppression-Aided SAR ship detection network. CSDet integrates a SAR Clutter Suppression Network (SAR-CSNet) with a detection network. SAR-CSNet effectively suppresses clutter and noise while preserving ship targets’ structural and electromagnetic characteristics. Its outputs are directly fed into the detection network, maintaining the unified end-to-end framework of CSDet. Clutter suppression aims to enhance detection performance, achieved by removing background interference. Additionally, Detection-Assisted Consistency Loss (DACL) is proposed to use detection outputs to guide clutter suppression, improving consistency between the two tasks. The modular design of CSDet allows it to be seamlessly adapted to various detection networks, highlighting significant flexibility and portability. Experimental results show a significant 11.06% improvement in Average Precision (AP), demonstrating its effectiveness in complex SAR inshore scenarios.

The well focused radar image was important for the downstream tasks. Though the classical methods could focus the radar echoes well, yet they highly relied on the accuracy modeling of phase error. The generalization ability was much poor in the realistic scenarios. To solve these problems, a new data-driven SAR image autofocus method (FocusNet) was proposed in this paper. Different from the preceding works, a new encoder-decoder architecture configured by residual connection was presented. So the defocus mechanism can be learned directly from the input SAR image. The explicit modeling of phase errors can be then circumvented. On the basis, the refocused image can be formed by the learned features during decoder phase. One major advantage of proposed method is the generalization ability to more diverse defocus model. For example, the low-frequency and higher-order phase errors can be handled even they were not considered during training. Finally, multiple rounds of experiments were performed. The results prove that the proposed method could provide much finer imaging quality compared with the state-of-the-art. In the measured dataset, the improvement of 0.61% and 5.36% was achieved for the PSNR and SSIM of the amplitude.

The high quality of synthetic aperture radar (SAR) images is the core key to ensure the feasibility of downstream tasks. However, due to some human errors, system errors and other reasons, the echo data collected by the radar is partially missing. It will cause defocusing, resolution loss and other problems in SAR images. The quality of SAR images degraded or even became unusable. Especially when the missingness was completely random, most classical methods were difficult to deal with. To address the above problems, an end-to-end imaging method based on deep learning (DL) methods was proposed. First, an image formation and enhancement network (IFENet) was designed, which combined the advantages of convolution and multi-head attention (MHA). Second, a supervised training strategy was designed to train the model. And an image formation (IF) loss was used to guide the model to converge. Finally, the converged model was used to perform end-to-end mapping from missing echoes to high-quality SAR single-look complex (SLC) images. Experiments have proven that the proposed method was able to form sufficiently high-quality SAR images on missing echoes. And it performed best and achieved the state-of-the-art (SOTA) results compared with other methods.

Frequency estimation was a fundamental problem in radar imaging. The classical Fourier spectral analysis suffered from the Rayleigh limit. The imaging performance deteriorated rapidly in low SNR conditions. In addition, the prior knowledge on the number of signal sources was required. To solve the problems, a new data-driven spectral estimation method via frequency attention fusion (FreqAF) was proposed in this letter. Different from the preceding works, the signal spectral were estimated by a deep architecture neural network automatically. The echo signal was first dechirped according to the radar parameters. It is then fed into a deep architecture for spectral estimation. The proposed architecture was composed of three phases, the decomposition, the FreqAF, and the projection. In the decomposition phase, the individual single-frequency components were estimated from the input dechirped signal. The components were dynamically fused in a delicate FreqAF module. The frequencies were obtained finally in the projection phase. Numerical experiments are performed to verify the proposed method.

In the classical radar imaging framework, the echo signals can be well compressed and focused by matched filtering. Yet these methods suffered model mismatch in the non-idea scenarios, such as the active jamming, the motion errors. In these situations, the imaging results were defocused or blurred. To solve these problems, a new learnable SAR imaging method was proposed in this paper. First, a hierarchical U-shaped network ImagingNet was constructed to learn the imaging mechanism from the history data. The base model was formed by a training strategy to optimize the errors between the learning imaging result and the reference image. On this basis, a new teacher-student training strategy was developed to refine the base model, and form the advanced model accordingly. Different from the classical imaging framework, the proposed method could focus the echo signals in the ideal and non-idea scenarios. In addition, the proposed method could achieve the real-time imaging when deployed on the parallel computing platform. Multiple rounds of experiments were performed to verify the proposed method. The performance improvement of 0.209 and 0.502 for SSIM, 4.8 dB and 4.4 dB for PSNR were achieved in the active jamming and motion errors scenarios in comparison to the classical method.

Synthetic aperture radar (SAR) imaging is a popularly used tool in many fields, while the contradiction between resolution and coverage is an open problem. The range resolution is relied on the bandwidth of transmitted signals, while the azimuth resolution is dependent on the length of synthesized aperture. They are related to radar configurations, and infeasible to be tuned freely. To solve the problem, a new SAR image super-resolution method is proposed in this paper. It is composed of pre-imaging, feature learning, and reconstruction. In the first phase, the zero-padded technique is used to achieve spatial alignment. The result is then decomposed along the sub-apertures, forming the sub-aperture images and the sub-band images. They were delivered to the second phase for cross-domain learning and hierarchical fusion. The pre-learning, encoder-decoder, and refinement are configured. The learned features are finally tuned for super-resolution reconstruction. On this basis, a new evaluation system was built to measure the performance. Both the supervised evaluation and the unsupervised evaluation are considered. The improvement of 3.78% is achieved for the amplitude components, while the phase accuracy is 4.62% better than the competitor.

SAR played an important role in target monitoring and identifying. Yet the imaging resolution and the spatial coverage was contradictory. A great many works were presented previously, yet little studies were devoted to complex-valued SAR image. In addition, the evaluation strategy customized to SAR image was unavailable. To address the problems, a new SAR image super-resolution method was proposed in this paper. Different from the preceding works, SAR super-resolution was achieved from the perspectives of metric-driven and task-driven, forming the base model (SARSR) and the advanced model (DetSR). The base model was composed of three phases, pre-imaging, feature refinement, and reconstruction. The periodogram spectral estimation technique was first used to achieve spatial alignment. The spectrogram was fed into a hybrid architecture to learn the high-level representations. The learned features were refined and reconstruction. On the basis, the prior knowledge transferred from target detection task was then used to promote the super-resolution performance, forming the advanced model. Finally, the evaluation system customized to SAR image was presented. A set of metrics composed of vision and imaging measurements were defined to assess the image quality initially. The task-driven evaluation was then presented. Multiple rounds of experiments were performed to verify the proposed method. The results proved that the metrics can be improved more than 10%, while the detection accuracy can be improved by 4%.

The improvement of synthetic aperture radar (SAR) image quality is the constant subject of SAR technology. In previous works, certain physical models were built to form SAR images from raw echoes, such as Range-Doppler algorithm (RDA). Although this family of methods is effective in the ideal conditions, the generalization ability is poor under the special scenarios. Moreover, these methods ignore the large amounts of history SAR data. And their processing speed is not enough to support their application in real-time systems unless draconian restrictions are added. To solve these problems, a learnable radar imaging paradigm driven by deep generative model is proposed in this paper. A deep fully convolutional network is first developed to achieve the mapping from the raw echoes to the imaging results. Then a large amount of historical echo data and corresponding SAR images formed by RDA are used to train and evaluate this network. It has been proven that our imaging paradigm can quickly form high-quality imaging results in various scenarios, through simulations and experiments.

The matched filtering played an important role in radar signal processing. It is actually the convolution of orginal signal with its conjugate transpose version, and hence can be viewed as a predefined operator. However, the predefined matched filtering suffers from some anomaly scenarios. For example, aircraft experiences shaking, speed changes or irregular motion. They made the signal compression unsuccessfully. To solve the problem, a method for signal generation is put forwards in this article by using U-net network. The PSLR and ISLR is used to evaluate the compression of the generated signal, and the 3dB bandwidth is used to quantify the range resolution of the generated signal. The method uses pre-training approach and the results become more accurate as the dataset grows so that we call the method learnable matched filtering. Multiple comparative studies are performed to demonstrate the advantage of the proposed strategy.

The deep learning-based target recognition methods have achieved great success in recent years. The major advantage lies in the ability to learn the high-level representation from the input data adaptively. Yet, they relied on large amounts of training data to cover the complete distribution space of samples. It is infeasible to be met in the practical applications. The fitting ability and the learning power were therefore deteriorated. To solve the problem, a new target augmentation method is proposed in this article. The original complex-valued image is first recast into the phase history. The range profiles are then randomly perturbed, such as the shift in cyclic, corruption in range bins, and the drop of range bins. The perturbed range profiles are used for target imaging, forming the new inverse synthetic aperture radar (ISAR) images. The diversity of the training dataset can be enhanced, and the learning effectiveness of the deep model can be improved accordingly. To verify the proposed method, several rounds of experiments are performed. The results demonstrate the advantages of the proposed method in comparison to state-of-the-art.

Due to the complexity of real-world applications, open set recognition is often more practical than closed set recognition. Compared with closed set recognition, open set recognition needs not only to recognize known classes but also to identify unknown classes. Different from most of the current methods, we proposed three novel frameworks with kinetic pattern to address the open set recognition problems, and they are kinetic prototype framework (KPF), adversarial KPF (AKPF), and an upgraded version of the AKPF, AKPF++. First, KPF introduces a novel kinetic margin constraint radius, which can improve the compactness of the known features to increase the robustness for the unknowns. Based on KPF, AKPF can generate adversarial samples and add these samples into the training phase, which can improve the performance with the adversarial motion of the margin constraint radius. Compared with AKPF, AKPF++ further improves the performance by adding more generated data into the training phase. Extensive experimental results on various benchmark datasets indicate that the proposed frameworks with kinetic pattern are superior to other existing approaches and achieve the state-of-the-art performance.

Radar jamming signal identification at low JNR and with limited training data are both open problems. Though multiple studies were performed previously, there still existed two problems needed to be solved. On the one hand, it is difficult to obtain discriminative presentations from the radar jamming signal that are accompanied with the solid Gaussian White Noise (GWN). On the other hand, the available data for training is difficult to obyain from the actual complex noise environment. To solve these problems, we propose a novel siamese network architecture with self attention named SA-Siam for Radar jamming signal classification. Firstly, transforming the jamming signal to the time-frequency (TF) domain where can represent higher dimensional information. Then the intra-class aggregation and inter-class separability of radar jamming signals are enhanced through the siamese network which is beneficial to learn more discriminative features from TF image especially in the case of limited data. In addition, the self attention block (SA) can further capture spatial correlations of the TF image so as to improve the anti-noise performance of the siamese network. We conducted quantitative and qualitative experiments on own dataset with a set of Deep CNNs and classical siamese algorithms. The results verify that our proposed SA-Siam can fully explore the representation of noised jamming data and dramatically improve the radar jamming classification performance under limited training data.

Alert or monitoring, runs through thousands of years of human history. Now, with the growth of the number of satellites in orbit, hundreds of terabytes of data are transmitted from the satellite to the data center every day. How to efficiently understand the information contained in these huge data in the face of practical needs is an increasingly urgent engineering challenge. However, most current computer vision methods are used for Optical images. Due to the presence of domain gaps between optical images and SAR images ,the processing results are not ideal when Optical and SAR images are mixed. Therefore, in view of the above problems, a network model is proposed to realize the correlation between SAR images and Optical images. The model solves the problem that the imaging mechanism of SAR images differs from Optical images. The domain gaps cause SAR images are not directly used in Optical images computer vision method. This paper proposes an initial set of methods and models that have learned robust representations for Optical and SAR images dataset. So image analysts are able to interchangeably use Optical and SAR images for downstream tasks by using our models.

The recent years had witnessed a resurgence on neural network. Many hidden layers were stacked hierarchically to learn the high-level representations. Great performances were achieved by the learned representations. However, this kind of learning models were highly dependent on large amounts of signals with label information. In the realistic scenarios, it is very difficult and costly to collect the modulated signals with label information. Given small training samples, the fitting power of deep models were limited. To solve the problems, a new family of signal augmentation strategies, segment-wise generation and signal-wise generation are proposed. The former builds new signal by tuning a single signal, while the latter combines several different modulated signals together to produce new signal. Four kinds of techniques, segment shift in cyclic, segment correlation in random, pairwise signals combination, and multiple signals concatenation are presented. The aim is to simulate the unforeseen disturbances during signal sampling. The recognition performance under the realistic scenarios can be then improved. Multiple comparative studies were performed. The results demonstrated the effectiveness of proposed strategy in comparisons to the classical methods, as well as the deep learning algorithms.

Target recognition via deep learning has achieved great performances in the preceding works. Yet this family of method are dependent on large amounts of training data with label information. For radar sensors, it is difficult to collect labeled data in practical due to the absent imaging truth. The commonly used solution is data augmentation. However, seldom studies are devoted to complex-valued radar images. In this paper, a new radar image generation method via target re-imaging is proposed. The original image is first cast into the frequency-aspect domain. The one axis represents the transmitted frequency, while the other presents the synthetic azimuth. The inverse operations of zero-padding and windowing are then applied on the transformed data. Two kinds of imaging techniques, the intra-sample re-imaging and the inter-sample re-imaging are presented to generate new radar images. The generated images are finally used to improve the learning effectiveness of deep models. Multiple comparative studies are performed to demonstrate the advantages of the proposed method.

Single sensor radar can no longer satisfy the increasingly complex electromagnetic environment. More attention is paid to radar sensor networks, which can obtain more information from different nodes and points of view. Moreover, better detection and tracking performance can be achieved through resource sharing and complementary advantages (joint learning). How to improve the utilization efficiency of multiple radar sensors with limited resources is an open problem, which can be transformed into joint learning in scenarios such as limited training data or imbalanced samples. This paper presents a distributed learning model to solve these problems. It has three phases. Self-reweighting loss is developed to dynamically rebalance the gradients of positive and negative samples for each category, after which the imbalance of samples can be alleviated. An image generation technique via target reimaging addresses the problem of limited samples. Self-reweighting loss and image generation are then unified in a federated learning framework. The classifier is adjusted using virtual representations to further improve learning efficiency. Comparative studies on the MSTAR dataset demonstrate the advantages of the proposed method.

Radar automatic target recognition (RATR) based on high resolution range profile (HRRP) has attracted more attention in recent years. In fact, the actual application environment of RATR is open set environment rather than closed set environment. However, previous works mainly focus on closed set recognition, which classifies the known classes by dividing hyper-planes in the feature space, and it will cause classification errors in open set environment. Therefore, open set recognition is proposed to solve this problem, which needs to determine a closed classification boundary for the identification of the known and unknown targets simultaneously. To accomplish this purpose, this paper proposes and proves the extreme value boundary theorem, which demonstrates that the maximum distance from the known features to the cluster center follows the generalized extreme value distribution. According to the proposed theorem, the closed classification boundary of the cluster is easily determined to distinguish between the known and unknown classes. Finally, extensive experiments on measured HRRP data verify the validity of the proposed theorem and the effectiveness of the proposed method.

The huge data collected every minute present an urgent need for image interpretation automatically. Many studies were performed previously, yet the problem was still far from being solved. For the task of detection and recognition, the preceding works usually paid more attentions to feature extraction on a single image. The azimuth information were directly ignored. This is because SAR images of target in many azimuths were absent in practical. It is therefore difficult to achieve image interpretation from the perspective of azimuth. To solve the problem, this paper develop a new recurrent architecture. We regard the images of target in the consecutive azimuths as the samples of sequence signal at the specific times. A new recurrent neural network composed of some convolutional cells is then developed. It inputs the samples prior to the current azimuth, and outputs the prediction in the next azimuth. Multiple studies are performed to demonstrate the effectiveness of proposed strategy.

With the increasing requirements for target classification capabilities in an open scene, various open set recognition algorithms emerge in an endless stream. It becomes urgent to evaluate the recognition performance. However, very few studies have examined this crucial and challenging problem previously. To fill the gap, a new evaluation strategy is proposed in this paper. First, two measurements, Openness and openness based on the number of instances per unknown class are presented to quantify the degree of open scenario. The recognition performance is then evaluated by this unified evaluation criteria across several different perspectives, the number of classes, the number of instances, and the synthesis difficulty, on which the ability of algorithms to classify the known classes and identify the unknown classes can be analyzed quantitatively. Multiple experiments are performed on the real dataset. The experimental results demonstrate the effectiveness of the proposed evaluation strategy.

Radar plays an increasingly important role in the civil applications as well as the military ones. Nowadays, various kinds of jamming strategies were presented to blind the radar. It is urgent to identify and recognize those jamming signals in the realistic applications. The early works relied on the predefined handcrafted features. They are highly dependent on the expert knowledge. Likewise, it is difficult to locate the jamming signal accurately. To solve this problem, this paper proposes a new method for joint detection and identification of jamming signals. A backbone composed of some convolutional blocks are presented to learn the high-level features. They are used to locate the jamming signals in the timefrequency plane. The identification of the interfering signal can be then recognized simultaneously. Multiple experiments are performed to demonstrate the effectiveness of proposed method.

Bridge-over-water detection plays vital role in both civilian and military applications. Though widely studied previously, it is still a challenging problem. This is because bridges are with a high diversity of aspect ratios, shapes and orientations in practical. The detection performance is highly dependent on the annotation accuracy of training samples. To address these problems, this paper proposes a new coarse-to-fine strategy to detect oriented bridges over water. In coarse detection stage, a new backbone containing modulated deformable convolution is developed to enrich the diversity of receptive fields. The detection results are further refined by the prior knowledge. The oriented bounding boxes can be then obtained based on frequency domain analysis and edge detection. Different from previous works, oriented bounding box annotaions are not required in the training of the proposed method. Comparative experiments were conducted. The results demonstrate the effectiveness of the proposed method.

Synthetic Aperture Radar Automatic Target Recognition (SAR-ATR) is the core application of SAR technology. The shortage of training data is a constraint for SAR-ATR. One of the valid methods to solve the problem is SAR image simulation. A large quantity of generative models have achieved impressive performances on SAR image simulation. Therefore, it is absolutely necessary to evaluate whether the simulated images fulfill the requirement of application. This challenging problem has not attracted plenty of attention. Very few studies have been done. To fill the blank, we propose a new evaluation strategy. Two quantitative measurements are proposed to evaluate the simulated images from the local and the global perspective respectively. The local measurement, Fréchet Inception Distance score (FID) is used to measure the distance of feature vector between the real images and the simulated images. Contrarily, the global measurement, Hybrid Recognition Rate curve (HRR) is developed to assess the application capability of the simulated images from a global perspective. Multiple comparative experiments are performed on real SAR dataset. The experimental results demonstrate the effectiveness of the proposed method.

The open set recognition model can identify the known and unknown samples simultaneously. In the radar automatic target recognition application, open set recognition meets better practical requirements than closed set recognition. Theoretically, an overlap usually exists between the known and unknown features, which makes it difficult for the model to identify unknown samples. Therefore, we explore the distribution of the known and unknown features, and find that the unknown features are usually smaller and closer to the center region in the feature space than the known features. Based on this phenomenon, we propose a novel loss function that improves the open set recognition performance by controlling the known features distributed to the surrounding area of the feature space. In addition, extensive experiments are carried out on measured HRRP data. Thus, we verify the effectiveness of the proposed method.

The diminishing land resources compel us to pay more attention to the Marine. The underwater image restoration therefore plays an increasingly important role. The early works usually rely on image super-resolution reconstruction. This family of methods yet suffer from the phenomenon of gradient dispersion, resulting in poor restoration performance. To solve this problem, this paper proposes a new deep neural network. The dense block structure and the adaptive mechanism are combined in a unified framework. The high-frequency representations can be then exploited. Moreover, the best weights of each channel can be determined automatically. Multiple comparative studies are performed. The experimental results prove that the proposed method could solve the gradient dispersion effectively and ignore the redundant information simultaneously.

Though widely studied, target detection in synthetic aperture radar (SAR) image is still a challenging problem. The classical convolutional neural network (CNN) samples spatial locations with the fixed geometric structure, and hence is incapable of learning the representations of varied-scale ships. It's difficult to locate multi-scale targets accurately in the complex scenes. On the other hand, to apply the classical models in SAR image, we need to duplicate the single-channel image to 3-channel one. The preprocess could not introduce the additional semantic information yet producing feature redundancy. To solve these problems, we introduced a new method for ship detection. We deployed the deformable convolutional block to learn features of ships with various scales at arbitrary locations. Different from the preceding works, the former shallow feature maps are also employed to enhance the representations of targets, especially small ships. In addition, the group normalization strategy is configured to alleviate internal covariate shift and accelerate the convergence. There is no need to make a tradeoff between the scale of model architecture and the batch size. Multiple comparative experiments on SSDD dataset are performed to demonstrate the advantage of proposed methods.

The past years have witnessed a revival of neural network and learning strategies. These models configure multiple hidden layers hierarchically and require large amounts of labeled samples to estimate the model parameters. It is yet difficult to be met for target recognition under the realistic environments. For either space borne or airborne radars, collecting multiple samples with label information is very expensive and difficult. In addition, the huge computational cost and poor speed of convergence limit the practical applications. To address the problems, this article presents a new thought of receptive, under which a special hierarchy of feedforward neural network has been built. The proposed strategy consists of two sequential modules: 1) feature generation and 2) feature refinement. We first build pairwise baseline signals by means of the Riesz transform along the range and the azimuth, and extend them to a family of receptive signals using the bandpass filter bank. The input SAR image is then generally convoluted with the set of receptive signals to extract the global features. Certain kinds of information can be then exploited. We make the receptive signals predefined, rather than learned automatically, to handle the environment of a small sample size. In addition, the expert knowledge can be transmitted into the neural network. The resulting features are further refined by a special unit, wherein the input neurons and the latent states are bridged by the weights and the bias randomly generated. They are fixed during the training process. On the other hand, we cast the latent state into the Hilbert space, forming the kernel version of refinement. We aim to achieve the comparable or even better performance yet with limited training resources.

To solve target recognition under extended environments, this paper proposes sparse representation in the transformed domain. Since the signal energy in the frequency domain is mainly concentrated on a small portion of low frequencies, this part of spectrum therefore carry the vital information that distinguishes a class of target from the other. We intend to define a frequency descriptor by the bag of low frequencies. The defined descriptor is used to build sparse signal modeling. The frequency descriptors of the training are concatenated to form an over-complete dictionary. It is used to encode the counterpart of query as a linear combination of themselves. Sparsity has been harnessed to generate the optimal representation, from which the inference can be reached.

For synthetic aperture radar (SAR) image land cover classification, traditional feature-based methods are not always effective because of the heavy multiplicative noise. To solve this problem, we herein propose a new classification method for SAR images considering adaptive spatial contextual information. In contrast to preceding studies, the spatial contextual information of the SAR images is exploited via composite kernels (CKs). Additionally, an image superpixel strategy is employed to design an adaptive neighborhood, which enables the extraction of more accurate spatial information than a fixed-size neighborhood. Specifically, a modified superpixel map is first generated to produce the neighborhood. With this neighborhood, a context kernel is then defined by means of the Gaussian radial basis function. The resulting context kernel is combined with the conventional feature kernel via the designed CKs scheme. The relative proportion of these two kernels is controlled by a weight parameter. The label of each pixel is predicted by feeding the final CKs into a support vector machine classifier. Experiments on two real SAR images demonstrate that the proposed method can greatly improve the classification performance, both visually and quantitatively, in comparison to other traditional feature-based methods.

Downward-looking linear array 3-D synthetic aperture radar (SAR) has attracted increasing attention in the field of radar imaging. As widely reported, the volume of data can be significantly reduced by a random sparse linear array. However, the 2-D under-sampled azimuth-cross-track data brought by the sparse linear array will produce high-level side-lobes, as well as the aliasing and the false-alarm targets. To deal with those problems, this paper introduces a recently developed theory, matrix completion (MC). The new theory could recover a matrix with a small subset of known elements of the matrix. It is founded on the assumption that the matrix is essentially low rank. For downward-looking 3-D SAR with a random sparse linear array, the received 3-D data can be treated as a series of uncorrelated 2-D matrices by the separated channel process. First, range compression can be realized by means of pulse compression. Then, the sets of the 2-D under-sampled azimuth-cross-track matrix can be completed into a full-sampled one via MC trick. The resulting 3-D images can be focused by synthetic aperture technique along the azimuth direction and beamforming operation along the cross-track direction, with the recovered full-sampled matrix. The proposed algorithm achieves high resolution and low-level side-lobes with the acceptable computational cost and memory consumption. It is verified by several numerical simulations and multiple comparative studies on real data. The experimental results clearly demonstrate the imaging performance across different under-sampling rates and signal-noise rates.

Automatic target recognition has been widely studied over the years, yet it is still an open problem. The main obstacle consists in extended operating conditions, e.g.., depression angle change, configuration variation, articulation, and occlusion. To deal with them, this paper proposes a new classification strategy. We develop a new representation model via the steerable wavelet frames. The proposed representation model is entirely viewed as an element on Grassmann manifolds. To achieve target classification, we embed Grassmann manifolds into an implicit reproducing Kernel Hilbert space (RKHS), where the kernel sparse learning can be applied. Specifically, the mappings of training sample in RKHS are concatenated to form an overcomplete dictionary. It is then used to encode the counterpart of query as a linear combination of its atoms. By designed Grassmann kernel function, it is capable to obtain the sparse representation, from which the inference can be reached. The novelty of this paper comes from: 1) the development of representation model by the set of directional components of Riesz transform; 2) the quantitative measure of similarity for proposed representation model by Grassmann metric; and 3) the generation of global kernel function by Grassmann kernel. Extensive comparative studies are performed to demonstrate the advantage of proposed strategy.

Recently, the problem of recovering sparse linear representation of a query in terms of a redundant dictionary has received great interest. The query sample is represented as a linear combination of the atoms of the dictionary. The unique representation is obtained via sparsity constraint, and the decision is made in terms of the characteristics of representation on reconstruction. The decision rule for sparse signal modeling can be viewed as a typical application of Bayesian estimation, where the likelihood function is inversely proportional to the reconstruction error. Different from the conventional rule, where the decision is directly made according to the overall reconstruction error associated with each class, this letter proposes a soft decision via Dempster-Shafer theory of evidence. To model the imprecision on uncertainty measurement, we introduce the samplewise ambiguity and the classwise ambiguity during the quantification of probability mass. Each sample that participates in the recovery of the query is considered as an item of evidence that supports certain hypothesis in regard to the class membership of query. The amount of evidence is quantified by a function of the distance between the query and the weighted training sample, where the weights result from the sparse representation coefficient. Then, various pieces of evidence derived from the candidate samples are pooled by means of Dempster's rule of combination, from which a soft decision can be reached.

In this letter, synthetic aperture radar (SAR) target recognition via classification on Riemannian geometry is presented. To characterize SAR images, which have broad spectral information yet spatial localization, a 2-D analytic signal, i.e., the monogenic signal, is used. Then, the monogenic components are combined by computing a covariance matrix whose entries are the correlation of the components. Since the covariance matrix, a symmetric positive definite one, lies on the Riemannian manifold, it is unreasonable to be dealt with by the standard learning techniques. To address the problem, two classification schemes are proposed. The first maps the covariance matrix into the vector space and feeds the resulting descriptor into a recently developed framework, i.e., sparse representation-based classification. The other embeds the Riemannian manifold into an implicit reproducing kernel Hilbert space, followed by least square fitting technique to recover the test. The inference is reached by evaluating which class of samples could reconstruct the test as accurately as possible.

In this letter, a novel classification via information aggregation through Dempster-Shafer's (DS) evidence theory has been presented to target recognition in a SAR image. Although the DS theory of evidence has been widely studied over the decades, less attention has been paid to its application for target recognition. To capture the characteristics of a SAR image, this letter exploits a new multidimensional analytic signal named monogenic signal. Since the components of the monogenic signal are of a high dimension, it is unrealistic to be directly used. To solve the problem, an intuitive idea is to derive a single feature by these components. However, this strategy usually results in some information loss. To boost the performance, this letter presents a classification framework via information aggregation. The monogenic components are individually fed into a recently developed algorithm, i.e., sparse representation-based classification, from which the residual with respect to each target class can be produced. Since the residual from a query sample reflects the distance to the manifold formed by the training samples of a certain class, it is reasonable to be used to define the probability mass. Then, the information provided by the monogenic signal can be aggregated via Dempster's rule; hence, the inference can be reached.

For downward-looking linear array 3-D synthetic aperture radar, the resolution in cross-track direction is much lower than the ones in range and azimuth. Hence, superresolution reconstruction algorithms are desired. Since the cross-track signal to be reconstructed is sparse in the object domain, compressive sensing algorithm has been used. However, the imaging processing on the 3-D scene brings large computational loads, which renders challenges in both data acquisition and processing. To cover this shortage, truncated singular value decomposition is utilized to reconstruct a reduced-redundancy spatial measurement matrix. The proposed algorithm provides advantages in terms of computational time while maintaining the quality of the scene reconstructions. Moreover, our results on uniform linear array are generally applicable to sparse nonuniform linear array. Superresolution properties and reconstruction accuracies are demonstrated using simulations under the noise and clutter scenarios.

This paper introduces a novel classification strategy based on the monogenic scale space for target recognition in Synthetic Aperture Radar (SAR) image. The proposed method exploits monogenic signal theory, a multidimensional generalization of the analytic signal, to capture the characteristics of SAR image, e.g., broad spectral information and simultaneous spatial localization. The components derived from the monogenic signal at different scales are then applied into a recently developed framework, sparse representation-based classification (SRC). Moreover, to deal with the data set, whose target classes are not linearly separable, the classification via kernel combination is proposed, where the multiple components of the monogenic signal are jointly considered into a unifying framework for target recognition. The novelty of this paper comes from: the development of monogenic feature via uniformly downsampling, normalization, and concatenation of the components at various scales; the development of score-level fusion for SRCs; and the development of composite kernel learning for classification. In particular, the comparative experimental studies under nonliteral operating conditions, e.g., structural modifications, random noise corruption, and variations in depression angle, are performed. The comparative experimental studies of various algorithms, including the linear support vector machine and the kernel version, the SRC and the variants, kernel SRC, kernel linear representation, and sparse representation of monogenic signal, are performed too. The feasibility of the proposed method has been successfully verified using Moving and Stationary Target Acquiration and Recognition database. The experimental results demonstrate that significant improvement for recognition accuracy can be achieved by the proposed method in comparison with the baseline algorithms.

In this paper, classification via joint sparse representation of the monogenic signal is presented for target recognition in SAR imagery. First, the monogenic signal is performed to capture the characteristics of SAR image. Since it is infeasible to directly apply the raw component to classification due to the high data dimension and redundancy, three augmented feature vectors are defined via uniform downampling of the real part, the imagery part, and the instantaneous phase. The monogenic features are then fed into a recently developed framework, sparse representation-based classification (SRC). Rather than produce individual sparse pattern, this paper generates the similar sparsity pattern for three feature vectors by imposing a mixed norm on the representation matrix. Extensive experiments on MSTAR database demonstrate that the proposed method could significantly improve the recognition accuracy.

In this paper, the classification via nonnegative and local linear regression model is proposed for SAR image-based target recognition. Recently, a simple yet effective method, linear regression for pattern recognition has been presented. By assuming that images from a single-object class lie on a linear subspace, it represents the test image as a linear combination of class-specific galleries. The representation is obtained by solving a typical inverse problem with least-square strategy. Since the negative weights play a counteractive role in reconstruction, it may be unreasonable to generate the negative weights. In addition, those elements close to the test sample should contribute much more than the ones far from the test. Thus this paper limits the feasible set of the representation by nonnegative and locality constraint. The decision is ruled in favor of the class with the minimum reconstruction error. Extensive experiments on MSTAR database demonstrate that the proposed methods significantly improve the accuracy than the standard one.

In this letter, the classification via sparse representation of the monogenic signal is presented for target recognition in SAR images. To characterize SAR images, which have broad spectral information yet spatial localization, the monogenic signal is performed. Then an augmented monogenic feature vector is generated via uniform down-sampling, normalization and concatenation of the monogenic components. The resulting feature vector is fed into a recently developed framework, i.e., sparse representation based classification (SRC). Specifically, the feature vectors of the training samples are utilized as the basis vectors to code the feature vector of the test sample as a sparse linear combination of them. The representation is obtained via l1-norm minimization, and the inference is reached according to the characteristics of the representation on reconstruction. Extensive experiments on MSTAR database demonstrate that the proposed method is robust towards noise corruption, as well as configuration and depression variations.

This paper addresses the statistical segmentation of SAR (Synthetic Aperture Radar) image combining PM (Perona Malik) nonlinear diffusion model and MRF (Markov Random Field) model. First, the original SAR image is filtered using the modified PM nonlinear diffusion model, in which the diffusion coefficients along the tangent direction and the normal direction are approximated and simplified. Afterwards, the filtered image is segmented using MRF model, in which the clique potential is computed using both the label configuration and the intensity information. The proposed method is marked by PM-MRF for short. Experimental results show that PM-MRF competes favorably with the traditional one to segment SAR image homogeneously.