Showing papers by "Mengdao Xing published in 2021"

FEC: A Feature Fusion Framework for SAR Target Recognition Based on Electromagnetic Scattering Features and Deep CNN Features

Abstract: The active recognition of interesting targets has been a vital issue for synthetic aperture radar (SAR) systems. The SAR recognition methods are mainly grouped as follows: extracting image features from the target amplitude image or matching the testing samples with the template ones according to the scattering centers extracted from the target complex data. For amplitude image-based methods, convolutional neural networks (CNNs) achieve nearly the highest accuracy for images acquired under standard operating conditions (SOCs), while scattering center feature-based methods achieve steady performance for images acquired under extended operating conditions (EOCs). To achieve target recognition with good performance under both SOCs and EOCs, a feature fusion framework (FEC) based on scattering center features and deep CNN features is proposed for the first time. For the scattering center features, we first extract the attributed scattering centers (ASCs) from the input SAR complex data, then we construct a bag of visual words from these scattering centers, and finally, we transform the extracted parameter sets into feature vectors with the $k$ -means. For the CNN, we propose a modified VGGNet, which can not only extract powerful features from amplitude images but also achieve state-of-the-art recognition accuracy. For the feature fusion, discrimination correlation analysis (DCA) is introduced to the FEC framework, which not only maximizes the correlation between the CNN and ASCs but also decorrelates the features belonging to different categories within each feature set. Experiments on Moving and Stationary Target Acquisition and Recognition (MSTAR) database demonstrate that the proposed FEC achieves superior effectiveness and robustness under both SOCs and EOCs.

Robust CFAR Ship Detector Based on Bilateral-Trimmed-Statistics of Complex Ocean Scenes in SAR Imagery: A Closed-Form Solution

Abstract: A robust constant false alarm rate (RCFAR) detector based on bilateral-trimmed-statistics (BTS-RCFAR) with a closed-form solution is proposed. BTS-RCFAR aims at improving the detection performance in complex ocean scenes such as the multiple-target environment, off-shore, oil-spilled ocean area, etc. In these circumstances, the clutter samples are often contaminated by the outliers. Consequently, the estimated parameters are biased, and the probability density function modeling of the clutter is not accurate. Detection performance deteriorates with either decrease of the detection rate or increase of the false alarm rate. Inspired by Sigma filter, BTS-RCFAR proposes a bilateral-thresholds-based strategy to automatically trim the samples in the local reference window, both the high-intensity and the low-intensity outliers are eliminated. Furthermore, the trimming depth is adaptively derived according to the homogeneity of the clutter backgrounds, where the outliers are completely removed and the real clutter samples can be greatly sustained. Maximum-likelihood-estimator with a closed-form solution is used for parameter estimation using the bilateral-trimmed samples, and log-normal model of the sea clutter can be accurately established. Finally, the test cell is detected given the specified probability of false alarm rate. BTS-RCFAR improves the detection performance in complex ocean scenes by elevating the detection rate and reducing the false alarm rate. Both simulated data and real data are used for validation.

Artificial Intelligence In Interferometric Synthetic Aperture Radar Phase Unwrapping: A Review

Abstract: Interferometric synthetic aperture radar (InSAR) is a radar technique widely used in geodesy and remote sensing applications, e.g., topography reconstruction and subsidence estimation. Phase unwrapping (PU) is one of the key procedures of InSAR signal processing. Artificial intelligence (AI) techniques have proven to be potentially powerful in many fields and have been introduced into the PU domain, achieving superior performance. In this article, we provide a comprehensive overview of AI-based PU techniques in InSAR. We survey the AI-based single-baseline (SB) PU methods and then review the AI techniques related to multibaseline (MB) PU. In addition, we show several experimental examples of these methods, from both simulated and real InSAR data sets, which gives readers an overview of AI-based PU processing's potential and limitations. It is our hope that this article will provide researchers with guidelines and inspiration to further enhance the development of AI-based PU.

Water Body Detection in High-Resolution SAR Images With Cascaded Fully-Convolutional Network and Variable Focal Loss

Abstract: The water body detection in high-resolution synthetic aperture radar (SAR) images is a challenging task due to the changing interference caused by multiple imaging conditions and complex land backgrounds. Inspired by the excellent adaptability of deep neural networks (DNNs) and the structured modeling capabilities of probabilistic graphical models, the cascaded fully-convolutional network (CFCN) is proposed to improve the performance of water body detection in high-resolution SAR images. First, for the resolution loss caused by convolutions with large stride in traditional convolutional neural network (CNN), the fully-convolutional upsampling pyramid networks (UPNs) are proposed to suppress this loss and realize pixel-wise water body detection. Then considering blurred water boundary, the fully-convolutional conditional random fields (FC-CRFs) are introduced to UPNs, which reduce computational complexity and lead to the automatic learning of Gaussian kernels in CRFs and the higher boundary accuracy. Furthermore, to eliminate the inefficient training caused by imbalanced categorical distribution in the training data set, a novel variable focal loss (VFL) function is proposed, which replaces the constant weighting factor of focal loss with the frequency-dependent factor. The proposed methods can not only improve the pixel accuracy and boundary accuracy but also perform well in detection robustness and speed. Results of GaoFen-3 SAR images are presented to validate the proposed approaches.

Deep Learning-Based Branch-Cut Method for InSAR Two-Dimensional Phase Unwrapping

Abstract: Two-dimensional (2-D) phase unwrapping (PU) is a critical processing step for many synthetic aperture radar (SAR) interferometry (InSAR) applications. As is well known, the traditional 2-D PU is an ill-posed inverse problem, which means that regardless of how skillful the PU algorithm designer is, it is impossible to design an algorithm that can correctly process all the 2-D PU situations, i.e., we can only design the best PU algorithm in the statistical sense. Therefore, accumulating PU processing experience from different study cases is important for PU algorithm design. Currently, the deep learning (DL) technique provides a potential framework to accumulate processing experience, and a flood of valuable data coming from different InSAR sensors provides the ability to enable the learning-based PU technique outside the traditional model-based technique. In this article, we transform the 2-D PU problem into a learnable image semantic segmentation problem and propose a DL-based branch-cut deployment method (abbreviated as BCNet). To start, we propose the optimal branch-cut connection criterion (referred to as OPT-BC) with the reference unwrapped phase given. Next, using the relationship between the residue and branch-cut as the learning objective, BCNet is trained using the samples provided by OPT-BC to produce the branch-cut result. Finally, the traditional branch-cut method is utilized to perform the postprocessing procedure to obtain the final PU result. The experimental results demonstrate that the proposed BCNet-based PU method is a near-real-time 2-D PU algorithm, and its accuracy outperforms the traditional model- and learning-based 2-D PU methods.

CANet: An Unsupervised Deep Convolutional Neural Network for Efficient Cluster-Analysis-Based Multibaseline InSAR Phase Unwrapping

Abstract: Multibaseline (MB) phase unwrapping (PU) is a vital processing procedure for MB synthetic aperture radar interferometry (InSAR) signal processing and can improve the traditional InSAR by changing the ill-posed problem to the well-posed problem. The existing research has shown that the MB PU problem can be successfully converted into an unsupervised cluster analysis problem. Using the high feature descriptiveness of the deep learning technique, an unsupervised deep convolutional neural network, referred to as CANet, is proposed to cluster all the pixels into different groups according to the input's recognizable pattern of the ambiguity number of the MB interferometric phase. Subsequently, we extend our previous two-stage programming-based MB processing approach (TSPA) to processing MB PU on a sparse irregular network, which is established from the clustering result of CANet. Both theoretical analysis and experimental results show that the proposed method is an effective MB PU method, and its execution time is drastically lower than those of many classical MB PU methods.

Integration of Rotation Estimation and High-Order Compensation for Ultrahigh-Resolution Microwave Photonic ISAR Imagery

Abstract: The microwave photonic (MWP) radar technique is capable of providing ultrawide frequency bandwidth waveforms to generate ultrahigh-resolution (UHR) inverse synthetic aperture radar (ISAR) imagery. Nevertheless, conventional ISAR imaging algorithms have limitations in focusing UHR MWP-ISAR imagery, where high-precision high-order range cell migration (RCM) and phase correction are crucially necessary. In this article, a UHR MWP-ISAR imaging algorithm integrating rotation estimation and high-order motion terms compensation is proposed. By establishing the relationship between parametric ISAR rotation model and high-order motion terms, an average range profile sharpness maximization (ARPSM) is developed to obtain rotation velocity by using nonuniform fast Fourier transform (NUFFT). Second-order range-dependent RCM is corrected with parametric compensation model by using the rotation velocity estimation. Furthermore, the spatial-variant high-order phase error is extracted to compensation by the entire image sharpness maximization (EISM). A new imaging framework is established with two one-dimensional (1-D) parameter estimations: ARPSM and EISM. Extensive experiments demonstrate that the proposed algorithm outperforms traditional ISAR imaging strategies in high-order RCM correction and azimuth focusing performance.

Processing of Bistatic SAR Data With Nonlinear Trajectory Using a Controlled-SVD Algorithm

Abstract: The nonlinear trajectory and bistatic characteristics of general bistatic synthetic aperture radar (SAR) can cause severe two-dimensional space-variance in the echo signal, and therefore it is difficult to focus the echo signal directly using the traditional frequency-domain imaging algorithm based on the assumption of azimuth translational invariance. At present, the state-of-the-art nonlinear trajectory imaging algorithm is based on singular value decomposition (SVD), which has the problem that SVD may be not controlled, and thus may lead to a high imaging complexity or low imaging accuracy. Therefore, this article proposes a nonlinear trajectory SAR imaging algorithm based on controlled SVD (CSVD). First, the chirp scaling algorithm is used to correct the range space-variance, and then SVD is used to decompose the remaining azimuth space-variant phase, and the first two feature components after SVD are integrated to make them be represented by a new feature component. Finally, the new feature component is used for interpolation to correct the azimuth space-variance. The simulation results show that the proposed CSVD can further improve the image quality compared with SVD-Stolt.

A Pixel Cluster CNN and Spectral-Spatial Fusion Algorithm for Hyperspectral Image Classification With Small-Size Training Samples

Abstract: Convolutional neural networks (CNNs) can automatically learn features from the hyperspectral image (HSI) data, avoiding the difficulty of manually extracting features. However, the number of training samples for the classification of HSIs is always limited, making it difficult for CNN to obtain effective features and resulting in low classification accuracy. To solve this problem, a pixel cluster CNN and spectral-spatial fusion (SSF) algorithm for hyperspectral image classification with small-size training samples is proposed in this article. First, spatial information is extracted by the gray level co-occurrence matrix. Then, spatial information and spectral information are fused by means of bands superposition, forming spectral-spatial features. To expand the number of training samples, the pixels after SSF are combined into pixel clusters according to a certain rule. Finally, a CNN framework is utilized to extract effective features from the pixel clusters. Experiments based on three standard HSIs demonstrate that the proposed algorithm can get better performance than the conventional CNN and also outperforms other studied algorithms in the case of small training set.

Enhanced-Random-Feature-Subspace-Based Ensemble CNN for the Imbalanced Hyperspectral Image Classification

Abstract: Hyperspectral image (HSI) classification often faces the problem of multiclass imbalance, which is considered to be one of the major challenges in the field of remote sensing. In recent years, deep learning has been successfully applied to the HSI classification, a convolutional neural network (CNN) is one of the most representative of them. However, it is difficult to effectively improve the accuracy of minority classes under the problem of multiclass imbalance. In addition, ensemble learning has been successfully applied to solve multiclass imbalance, such as random forest (RF) This article proposes a novel enhanced-random-feature-subspace-based ensemble CNN algorithm for the multiclass imbalanced problem. The main idea is to perform random oversampling of training samples and multiple data enhancements based on random feature subspace, and then, construct an ensemble learning model combining random feature selection and CNN to the HSI classification. Experimental results on three public hyperspectral datasets show that the performance of the proposed method is better than the traditional CNN, RF, and deep learning ensemble methods.

Ground Cartesian Back-Projection Algorithm for High Squint Diving TOPS SAR Imaging

Abstract: This article presents a fast back-projection (BP) algorithm based on subaperture (SA) image coherent combination in a downsampled Cartesian coordinate grid for high squint diving terrain observation by progressive scans (HSD-TOPS) synthetic aperture radar (SAR) ground plane imaging. A two-step spectrum compression (SC) method is proposed to coherently combine the aliasing SA images by exploiting the relationship between the wavenumber and the image frequency. The first-step SC is introduced to align the spectrum support region centers. The second-step SC effectively corrects the space-variant spectrum inclination. The proposed algorithm does not need interpolation in the process of image combination, which ensures the accuracy and the efficiency of the algorithm. Furthermore, the SC method is well-modified to suppress the sidelobes of the focused image. Simulation and measured data processing verify the effectiveness of the proposed method.

EFTL: Complex Convolutional Networks With Electromagnetic Feature Transfer Learning for SAR Target Recognition

Abstract: Considering that synthetic aperture radar (SAR) images obtained directly after signal processing are in the form of complex matrices, we propose a complex convolutional network for SAR target recognition. In this article, we give a brief introduction to complex convolutional networks and compare them with the real counterpart. A complex activation function is applied to analyze the influence of phase information in complex neural networks. Inspired by the theory of network visualization, a special kind of transfer learning based on the electromagnetic property from the attributed scattering center model is applied in our networks to modulate the first convolutional layer. The experiment shows a better performance in terms of classification accuracy compared to random weight initialization.

Fast Direction of Arrival Estimation for Uniform Circular Arrays With a Virtual Signal Subspace

Abstract: In this article, an augmented subspace-based algorithm for detecting the direction of arrival (DOA) of signals is presented. In the developed scheme, a uniform circular antenna array with a full range of lateral capacity is first transformed into a uniform linear antenna array in which the steering matrix has a Vandermonde form. This kind of structure can be exploited to formulate computationally efficient search-free estimation methods. In the process of virtual transformation, a novel DOA unit matrix of an assumed sector where the signals to be detected are located is built and optimized to find response matrices of the real and virtual antenna arrays to signals from the sector. By utilizing singular value decomposition (SVD) of the response matrix of the real antenna array to possible signals from the sector, a stable virtual transformation matrix between the real and virtual antenna arrays is obtained. Meanwhile, the aperture of the original array is also expanded during this process. Then, a virtual cyclic optimization algorithm is introduced to thoroughly mine information from the correlation matrices of the virtual antenna array receiving data and to optimize the signal subspace. Subsequently, the DOA can be efficiently determined through the reconstruction of the steering matrix. In the experimental part of the study, root-mean-square errors and probability of success are used to evaluate the performance of the algorithms. In particular, simulation results demonstrate that the proposed methods provide significant accuracy improvements, especially for low signal-to-noise ratio thresholds.

Ground Moving Target Indication for the Geosynchronous-Low Earth Orbit Bistatic Multichannel SAR System

Abstract: In this article, an effective clutter suppression and moving target imaging approach is proposed for the geosynchronous-low earth orbit (GEO-LEO) bistatic multichannel synthetic aperture radar (SAR) system, which is robust for the fast moving target with Doppler centroid ambiguity. For the GEO-LEO bistatic multichannel SAR system, the characteristic of baseline with azimuth invariant in Doppler Fourier transform domain is not tenable. It is difficult to implement clutter suppression by the unified baseline compensation in azimuth Doppler Fourier transform domain. Fortunately, we discover that the baseline can be approximately regarded as a constant in chirp Fourier transform domain for the GEO-LEO bistatic multichannel SAR system. Hence, the corresponding baseline compensation can be achieved in the azimuth chirp Fourier transform domain for the clutter. With an orthogonality vector of clutter, the clutter can be well suppressed and the moving target is extracted. After that, the moving target imaging approach is developed with the baseband Doppler centroid estimation and Doppler ambiguity number estimation. Finally, the theoretical investigations and the proposed approach in this article are validated by some experiments, where the experiments for clutter suppression and experiments for fast moving targets are included. Especially, an experiment with real SAR scattering scene is involved. In addition, some discussions for blind velocity targets are presented.

SVD-Based Ambiguity Function Analysis for Nonlinear Trajectory SAR

Abstract: A nonlinear trajectory of a radar platform in synthetic aperture radar (SAR) may lead to severe coupling between the range and the azimuth, which may make the ambiguity function (AF) analysis complicated. The numerical algorithm-based AF analysis may be computationally expensive, while the existing analytical algorithm-based AF analysis may cause large errors because it does not consider the coupling between the range and the azimuth. By observing that the singular value decomposition (SVD) is good to deal with the coupling problem, in this article, we propose an effective AF analysis based on SVD. The key idea is to first use a small amount of sampling points for SVD of the coupled term in the AF and then the decoupled vectors are fitted to high-order polynomials for the analytical AF calculation. It converts the double integral into the product of two single integrals in the calculation. From the proposed SVD-based AF analysis, three parameters, namely, 3-dB resolution, peak sidelobe ratio (PSLR), and integrated sidelobe ratio (ISLR), are then effectively computed. The simulated results verify the good performance of the proposed SVD-based AF analysis.

DSDet: A Lightweight Densely Connected Sparsely Activated Detector for Ship Target Detection in High-Resolution SAR Images

Abstract: Traditional constant false alarm rate (CFAR) based ship target detection methods do not work well in complex conditions, such as multi-scale situations or inshore ship detection. With the development of deep learning techniques, methods based on convolutional neural networks (CNN) have been applied to solve such issues and have demonstrated good performance. However, compared with optical datasets, the number of samples in SAR datasets is much smaller, thus limiting the detection performance. Moreover, most state-of-the-art CNN-based ship target detectors that focus on the detection performance ignore the computation complexity. To solve these issues, this paper proposes a lightweight densely connected sparsely activated detector (DSDet) for ship target detection. First, a style embedded ship sample data augmentation network (SEA) is constructed to augment the dataset. Then, a lightweight backbone utilizing a densely connected sparsely activated network (DSNet) is constructed, which achieves a balance between the performance and the computation complexity. Furthermore, based on the proposed backbone, a low-cost one-stage anchor-free detector is presented. Extensive experiments demonstrate that the proposed data augmentation approach can create hard SAR samples artificially. Moreover, utilizing the proposed data augmentation approach is shown to effectively improves the detection accuracy. Furthermore, the conducted experiments show that the proposed detector outperforms the state-of-the-art methods with the least parameters (0.7 M) and lowest computation complexity (3.7 GFLOPs).

Microwave Correlation Forward-Looking Super-Resolution Imaging Based on Compressed Sensing

Abstract: Forward-looking correlated imaging plays an increasingly important role in modern radar imaging systems. It overcomes disadvantages of traditional side or squint synthetic aperture radar (SAR) which is dependent on specific relative motion between the radar and target scene. A new microwave forward-looking correlated 3-D imaging method based on random radiation field combined with sparse reconstruction is proposed in this article. Firstly, phased array radar (PAR) is adopted to form different and random antenna patterns. Then, combined with the compressed sensing (CS) theory, the target image can be recovered with very few samples which can break through Rayleigh resolution limitation. Furthermore, the proposed method can achieve resolution at least 5.5 times higher than real aperture imaging. To raise computation efficiency of sparse reconstruction, an improved quasi-Newton iteration method based on graphics processing unit (GPU) platform is developed. Meanwhile, a GPU-based (NVIDIA Tesla K40c) accelerated computing method can significantly reduce the processing time compared with the time given by a personal computer (PC). Both simulation and field experiment verify the validity of the proposed method.

CNN-LRP: Understanding Convolutional Neural Networks Performance for Target Recognition in SAR Images

Abstract: Target recognition is one of the most challenging tasks in synthetic aperture radar (SAR) image processing since it is highly affected by a series of pre-processing techniques which usually require sophisticated manipulation for different data and consume huge calculation resources. To alleviate this limitation, numerous deep-learning based target recognition methods are proposed, particularly combined with convolutional neural network (CNN) due to its strong capability of data abstraction and end-to-end structure. In this case, although complex pre-processing can be avoided, the inner mechanism of CNN is still unclear. Such a “black box” only tells a result but not what CNN learned from the input data, thus it is difficult for researchers to further analyze the causes of errors. Layer-wise relevance propagation (LRP) is a prevalent pixel-level rearrangement algorithm to visualize neural networks’ inner mechanism. LRP is usually applied in sparse auto-encoder with only fully-connected layers rather than CNN, but such network structure usually obtains much lower recognition accuracy than CNN. In this paper, we propose a novel LRP algorithm particularly designed for understanding CNN’s performance on SAR image target recognition. We provide a concise form of the correlation between output of a layer and weights of the next layer in CNNs. The proposed method can provide positive and negative contributions in input SAR images for CNN’s classification, viewed as a clear visual understanding of CNN’s recognition mechanism. Numerous experimental results demonstrate the proposed method outperforms common LRP.

Multisystem Interferometric Data Fusion Framework: A Three-Step Sensing Approach

Abstract: The recent, sharp increase in the availability of interferometric data captured by different synthetic aperture radar (SAR) interferometry (InSAR) sensors poses a new scientific question that whether there is a processing framework that can combine these observations to obtain a more credible InSAR product (i.e., digital elevation model (DEM) and surface deformation estimation). In this article, we extend our previous two-stage programming-based multibaseline processing framework for combining the interferograms generated from disparate InSAR systems with different system parameters to enhance the InSAR performance at the signal processing stage. The proposed multisystem interferometric data fusion framework, abbreviated as TSDFF, includes three processing steps: multisystem interferogram registration, multisystem phase unwrapping, and absolute phase fusing. The advantage of TSDFF is that it can allow the data sets from different InSAR sensors to help each other to get rid of the limitation of the Itoh condition so that the application scope of each InSAR sensor will be effectively enlarged (e.g., measuring violent surface change or mountainous DEM). In addition, to quantitatively analyze the measurement bias robustness bound of TSDFF, the TSDFF-Fusion theorem is proposed, which offers significant application guidance for TSDFF at different noise levels. The real and simulated experimental results reveal the effectiveness of TSDFF for fusing the data sets from disparate InSAR systems.

An Improved SRGAN Based Ambiguity Suppression Algorithm for SAR Ship Target Contrast Enhancement

Abstract: Due to the specific characteristics of synthetic aperture radar (SAR), there will be ambiguity interference in SAR images, resulting in low contrast of the ship target to the clutter. This letter proposes an improved super-resolution generative adversarial network (ISRGAN) based ambiguity suppression algorithm for SAR ship target contrast enhancement. The proposed ISRGAN is the first attempt of using GAN for SAR ambiguity suppression. As a post-processing procedure, it does not need prior information of SAR systems, so it can be applied to various observation scenes and different acquisition modes. The generator of ISRGAN embeds the residual dense network (RDN) to optimally fuse the global and local features of the image, and it effectively improves the completeness of the feature information used for SAR ship target contrast enhancement. The superiority of ISRGAN on ambiguity suppression is validated on the Chinese Gaofen-3 imagery.

Data-Driven Motion Compensation for Airborne Bistatic SAR Imagery Under Fast Factorized Back Projection Framework

Abstract: Due to the independence of azimuth invariance and high implementing efficiency, a fast time-domain algorithm has significant advantages for airborne bistatic synthetic aperture radar (BiSAR) data process with general geometric configuration. In this article, the practical problem of unexpected motion errors of the airborne platform is carefully analyzed under a fast factorized back-projection (FFBP) framework for a general BiSAR process and a coherent data-driven motion compensation (MOCO) algorithm integrated with FFBP is proposed. By utilizing wavenumber decomposition, the analytical spectrum of a polar grid image is obtained where the motion error can be conveniently investigated in image spectrum domain and the coherence between azimuthal phase error (APE) and motion-induced nonsystematic range cell migration (NsRCM) can be perfectly revealed. Then, a new data-driven MOCO method for both APE and NsRCM correction is developed with the FFBP process. Different from the data-driven MOCO in most frequency-domain algorithms, the residual NsRCM introduced by the FFBP process is particularly analyzed and addressed in the MOCO, which significantly improves the image quality in focusing. Promising results from both simulation and raw data experiments are presented and analyzed to validate the advantages of the proposed algorithm for the airborne BiSAR process.

SMOTE-Based Weighted Deep Rotation Forest for the Imbalanced Hyperspectral Data Classification

Abstract: Conventional classification algorithms have shown great success in balanced hyperspectral data classification. However, the imbalanced class distribution is a fundamental problem of hyperspectral data, and it is regarded as one of the great challenges in classification tasks. To solve this problem, a non-ANN based deep learning, namely SMOTE-Based Weighted Deep Rotation Forest (SMOTE-WDRoF) is proposed in this paper. First, the neighboring pixels of instances are introduced as the spatial information and balanced datasets are created by using the SMOTE algorithm. Second, these datasets are fed into the WDRoF model that consists of the rotation forest and the multi-level cascaded random forests. Specifically, the rotation forest is used to generate rotation feature vectors, which are input into the subsequent cascade forest. Furthermore, the output probability of each level and the original data are stacked as the dataset of the next level. And the sample weights are automatically adjusted according to the dynamic weight function constructed by the classification results of each level. Compared with the traditional deep learning approaches, the proposed method consumes much less training time. The experimental results on four public hyperspectral data demonstrate that the proposed method can get better performance than support vector machine, random forest, rotation forest, SMOTE combined rotation forest, convolutional neural network, and rotation-based deep forest in multiclass imbalance learning.

Integrating the Reconstructed Scattering Center Feature Maps With Deep CNN Feature Maps for Automatic SAR Target Recognition

Abstract: Automatic target recognition has been one of the hottest research in synthetic aperture radar (SAR) data processing. Noticing that popular recognition methods cannot utilize multiple features of SAR complex data, a method fused scattering center feature and deep convolutional neural network (CNN) feature is proposed in this letter. This method contains three key parts, namely, scattering center extraction and reconstruction block, CNN feature extraction block, and final feature fusion and classification block. In this process, the scattering center feature and CNN feature are fused at the level of feature maps, which retain the space information of 2-D feature maps. What is more, the proposed half end-to-end strategy realizes the automatic update of weighting parameters in feature extraction network and subnetwork, which promotes a better recognition efficiency. Experimental results on measured SAR data show that the proposed method can achieve better accuracy than other single feature-based methods and feature fusion methods.

Realizing Target Detection in SAR Images Based on Multiscale Superpixel Fusion.

Abstract: For target detection in complex scenes of synthetic aperture radar (SAR) images, the false alarms in the land areas are hard to eliminate, especially for the ones near the coastline. Focusing on the problem, an algorithm based on the fusion of multiscale superpixel segmentations is proposed in this paper. Firstly, the SAR images are partitioned by using different scales of superpixel segmentation. For the superpixels in each scale, the land-sea segmentation is achieved by judging their statistical properties. Then, the land-sea segmentation results obtained in each scale are combined with the result of the constant false alarm rate (CFAR) detector to eliminate the false alarms located on the land areas of the SAR image. In the end, to enhance the robustness of the proposed algorithm, the detection results obtained in different scales are fused together to realize the final target detection. Experimental results on real SAR images have verified the effectiveness of the proposed algorithm.

Learning an SAR Image Despeckling Model Via Weighted Sparse Representation

Abstract: Synthetic aperture radar (SAR) images are inherently degraded by the speckle noise due to the coherent imaging, which may affect the performance of subsequent image analysis task. To address this problem, a weighted sparse representation-based method is proposed in this article for SAR image despeckling. The homomorphic transformation is first adopted to convert multiplicative noise into additive one. Second, similar patches are grouped together to learn the adaptive dictionaries and sparse coefficients based on nonlocal self-similarity constraint. Moreover, weighted regularizations are adopted for coefficients to boost the performance. Finally, despeckling images are obtained via exponential transformation. Experimental results on synthetic and real-world SAR images demonstrate that our proposed method outperforms several state-of-the-art methods in terms of both quantitative measurements and visual quality.

Toward More Robust and Real-Time Unmanned Aerial Vehicle Detection and Tracking via Cross-Scale Feature Aggregation Based on the Center Keypoint

Abstract: Unmanned aerial vehicles (UAVs) play an essential role in various applications, such as transportation and intelligent environmental sensing. However, due to camera motion and complex environments, it can be difficult to recognize the UAV from its surroundings thus, traditional methods often miss detection of UAVs and generate false alarms. To address these issues, we propose a novel method for detecting and tracking UAVs. First, a cross-scale feature aggregation CenterNet (CFACN) is constructed to recognize the UAVs. CFACN is a free anchor-based center point estimation method that can effectively decrease the false alarm rate, the misdetection of small targets, and computational complexity. Secondly, the region of interest-scale-crop-resize (RSCR) method is utilized to merge CFACN and region-of-interest (ROI) CFACN (ROI-CFACN) further, in order to improve the accuracy at a lower computational cost. Finally, the Kalman filter is adopted to track the UAV. The effectiveness of our method is validated using a collected UAV dataset. The experimental results demonstrate that our methods can achieve higher accuracy with lower computational cost, being superior to BiFPN, CenterNet, YoLo, and their variants on the same dataset.

Ensemble CNN Based on Pixel-Pair and Random Feature Selection for Hyperspectral Image Classification with Small-Size Training Set

Abstract: Recently, convolutional neural network (CNN) is widely used in hyperspectral image classification (HSIC) because of its strong self-learning and efficient feature expression ability. However, the CNN model faces the “overfitting” problem when the number of training samples is small. To improve the classification accuracy of CNN under the condition of limited training set, an ensemble CNN method based on pixel-pair and random feature selection (RFS) for HSIC is proposed in this paper. With the purpose of expanding training samples, the pixel-pair feature (PPF) is used in the presented study. Besides, ensemble CNN based on RFS is applied to further improve the classification performance. Experimental results based on two standard hyperspectral images demonstrate that the proposed method achieves better classification performance than the PPF based on CNN (PPF-CNN) and RFS based on SVM (RFS-SVM) methods.