Performance complexity study of several approaches to automatictarget recognition from SAR images

A framework which allows for the direct comparison of alternate approaches to automatic target recognition (ATR) from synthetic aperture radar (SAR) images is described and applied to variants of several ATR algorithms. This framework allows comparisons to be made on an even footing while minimizing the impact of implementation details and accounts for variation in image sizes, in angular resolution, and in the sizes of orientation windows used for training. Alternate approaches to ATR are characterized in terms of the best achievable performance as a function of the complexity of the model parameter database. Several approaches to ATR from SAR images are described and the performance achievable by each for a range of database complexities is studied and compared. These approaches are based on a likelihood test under a conditionally Gaussian model, log-magnitude least squared error, and quarter power least squared error. All approaches are evaluated for a wide range of parameterizations and the dependence on these parameters of both the resulting performance and the resulting database complexity is explored. Databases for all of the approaches are trained using identical sets of images and their performance is assessed under identical testing scenarios in terms of probability of correct classification, confusion matrices, and orientation estimation error. The results indicate that the conditionally Gaussian approach outperforms the other two approaches on average for both target recognition and orientation estimation, that accounting for radar power fluctuation improves performance for all three methods, and that the conditionally Gaussian approach normalized for power delivers average performance that is equal or superior to all other considered approaches     

Automatic spectral target recognition in hyperspectral imagery

Automatic target recognition (ATR) in hyperspectral imagery is a challenging problem due to recent advances of remote sensing instruments which have significantly improved sensor's spectral resolution. As a result, small and subtle targets can be uncovered and extracted from image scenes, which may not be identified by prior knowledge. In particular, when target size is smaller than pixel resolution, target recognition must be carried out at subpixel level. Under such circumstance, traditional spatial-based image processing techniques are generally not applicable and may not perform well if they are applied. The work presented here investigates this issue and develops spectral-based algorithms for automatic spectral target recognition (ASTR) in hyperspectral imagery with no required a priori knowledge, specifically, in reconnaissance and surveillance applications. The proposed ASTR consists of two stage processes, automatic target generation process (ATGP) followed by target classification process (TCP). The ATGP generates a set of targets from image data in an unsupervised manner which will subsequently be classified by the TCP. Depending upon how an initial target is selected in ATGP, two versions of the ASTR can be implemented, referred to as desired target detection and classification algorithm (DTDCA) and automatic target detection and classification algorithm (ATDCA). The former can be used to search for a specific target in unknown scenes while the latter can be used to detect anomalies in blind environments. In order to evaluate their performance, a comparative and quantitative study using real hyperspectral images is conducted for analysis.     

Automatic target recognition for hyperspectral imagery using high-order statistics

Due to recent advances in hyperspectral imaging sensors many subtle unknown signal sources that cannot be resolved by multispectral sensors can be now uncovered for target detection, discrimination, and identification. Because the information about such sources is generally not available, automatic target recognition (ATR) presents a great challenge to hyperspectral image analysts. Many approaches developed for ATR are based on second-order statistics in the past years. This paper investigates ATR techniques using high order statistics. For ATR in hyperspectral imagery, most interesting targets usually occur with low probabilities and small population and they generally cannot be described by second-order statistics. Under such circumstances, using high-order statistics to perform target detection have been shown by experiments in this paper to be more effective than using second order statistics. In order to further address a challenging issue in determining the number of signal sources needed to be detected, a recently developed concept of virtual dimensionality (VD) is used to estimate this number. The experiments demonstrate that using high-order statistics-based techniques in conjunction with the VD to perform ATR are indeed very effective     

Superresolution HRR ATR with high definition vector imaging

A new 1-D template-based automatic target recognition (ATR) algorithm is developed and tested on high range resolution (HRR) profiles formed from synthetic aperture radar (SAR) images of targets taken from the Moving and Stationary Target Acquisition and Recognition (MSTAR) data set. In this work, a superresolution technique known as High Definition Vector Imaging (HDVI) is applied to the HRR profiles before the profiles are passed through ATR classification. The new I-D ATR system using HDVI demonstrates significantly improved target recognition compared with previous I-D ATR systems that use conventional image processing techniques. This improvement in target recognition is quantified by improvement in probability of correct classification (PCC). More importantly, the application of HDVI to HRR profiles helps to maintain the same ATR performance with reduced radar resource requirements     

一种提高SAR目标识别率的有效方法

在合成孔径雷达自动目标识别SAR ATR中，SAR像的预处理是提高识别率的关键技术之一。给出了一种简单有效的SAR图像预处理方法，该方法首先对SAR目标像进行对数变换后，再做傅立叶变换。经预处理后的SAR像用支持矢量机SVM分类器进行目标识别。实验结果表明：本方法不但有效地提高了目标识别率，而且保证了目标的平移不变性并具有良好的推广能力。     

Critique of some neural network architectures and claims forcontrol and estimation

While there is great potential for successful use of neural network (NN) algorithms in automatic target recognition (ATR) and other pattern identification/classification applications, significant barriers have been encountered that, to date, defy rigorous use of NNs within feedback control designs. The status of several problems and contradictions involving NNs relating to control and estimation theory applications (and to practical failure detection within INS/GPS navigation systems) are summarized here. To give a positive spin and for a balanced perspective, we also mention many novel laudable NN results obtained by invoking the techniques and results of control and estimation theory     

Support vector machines for SAR automatic target recognition

Algorithms that produce classifiers with large margins, such as support vector machines (SVMs), AdaBoost, etc, are receiving more and more attention in the literature. A real application of SVMs for synthetic aperture radar automatic target recognition (SAR/ATR) is presented and the result is compared with conventional classifiers. The SVMs are tested for classification both in closed and open sets (recognition). Experimental results showed that SVMs outperform conventional classifiers in target classification. Moreover, SVMs with the Gaussian kernels are able to form a local “bounded” decision region around each class that presents better rejection to confusers     

Bayesian gamma mixture model approach to radar target recognition

This paper develops a Bayesian gamma mixture model approach to automatic target recognition (ATR). The specific problem considered is the classification of radar range profiles (RRPs) of military ships. However, the approach developed is relevant to the generic discrimination problem. We model the radar returns (data measurements) from each target as a gamma mixture distribution. Several different motivations for the use of mixture models are put forward, with gamma components being chosen through a physical consideration of radar returns. Bayesian formalism is adopted and we obtain posterior distributions for the parameters of our mixture models. The distributions obtained are too complicated for direct analytical use in a classifier, so Markov chain Monte Carlo (MCMC) techniques are used to provide samples from the distributions. The classification results on the ship data compare favorably with those obtained from two previously published techniques, namely a self-organizing map and a maximum likelihood gamma mixture model classifier.     

Maneuvering Target Tracking Using Adaptive State Estimation

Two approaches to a nonlinear state estimation problem are presented. The particular problem addressed is that of tracking a maneuvering target in three-dimensional space using spherical observations (radar data). Both approaches rely on semi-Markov modeling of target maneuvers and result in effective algorithms that prevent the loss of track that often occurs when a target makes a sudden, radical change in its trajectory. Both techniques are compared using real and simulated radar measurements with emphasis on performance and computational burden.     

Comparison of Two Target Classification Techniques

It has been shown that radar returns in the resonance region carry information regarding the overall dimensions and shape of targets. Two radar target classification techniques developed to utilize such returns are discussed. Both of these techniques utilize resonance region backscatter measurements of the radar cross section (RCS) and the intrinsic target backscattered phase. A target catalog used for testing the techniques was generated from measurements of the RCS of scale models of modern aircraft and naval ships using a radar range at The Ohio State University. To test the classification technique, targets had their RCS and phase taken from the data base and corrupted by errors to simulate full-scale propagation path and processing distortion. Several classification methods were then used to determine how well the corrupted measurements fit the measurement target signatures in the catalog. The first technique uses nearest neighbor (NN) algorithms on the RCS magnitude and (range corrected) phase at a number (e.g., 2, 4, or 8) of operating frequencies. The second technique uses an inverse Fourier transformation of the complex multifrequency radar returns to the time domain followed by cross correlation. Comparisons are made of the performance of the two techniques as a function of signal-to-error noise power ratio for various processing options.     

Automatic target recognition using enhanced resolution SAR data

Using advanced technology, a new automatic target recognition (ATR) system has been developed that provides significantly improved target recognition performance compared with ATR systems that use conventional synthetic aperture radar (SAR) image-processing techniques. This significant improvement in target recognition performance is achieved by using a new superresolution image-processing technique that enhances SAR image resolution (and image quality) prior to performing target recognition. A computationally efficient two-level implementation of a template-based classifier is used to perform target recognition. The improvement in target recognition performance achieved using superresolution image processing in this new ATR system is quantified     

一种基于红外图像的目标自动识别算法

红外成像制导具有在各种复杂战术环境下自主搜索、捕获、识别和跟踪目标的能力,代表了当代红外制导技术的发展趋势。提出了一种红外图像预处理、跟踪、分类的自动目标识别算法,利用小波变换、形态学方法对红外图像进行预处理,提取不同频带的惯性不变矩作为特征量,利用神经网络进行分类识别,结果表明该算法具有很高的识别率,对于精确制导武器的目标识别研究具有一定的参考价值。     

Efficient ATR using compression

We examine various model-based automatic target recognition (MBATR) classifiers to investigate the utility of model-catalog compression realized via signal-vector quantization (VQ) and feature extraction. We specifically investigate the impact of various compression rates and common automatic target recognition (ATR) scenario variations such as noise and occlusion through simulations on high-range resolution (HRR) radar and synthetic aperture radar (SAR) data. For this data, we show that significant computational savings are possible for modest decreases in classification performance.     

GMM-based target classification for ground surveillance Doppler radar

An automatic target recognition (ATR) algorithm, based on greedy learning of Gaussian mixture model (GMM) is developed. The GMMs were obtained for a wide range of ground surveillance radar targets such as walking person(s), tracked or wheeled vehicles, animals, and clutter. Maximum-likelihood (ML) and majority-voting decision schemes were applied to these models for target classification. The corresponding classifiers were trained and tested using distinct databases of target echoes, recorded by ground surveillance radar. ML and majority-voting classifiers obtained classification rates of 88% and 96%, correspondingly. Both classifiers outperform trained human operators.     

Wavelet preprocessing for high range resolution radarclassification

We develop a wavelet denoising scheme to aid an automatic target recognition (ATR) system in recognizing aircraft from high range resolution radar (HRR) signatures. A template matching classification technique is used with templates formed from synthetically generated signatures. The goal of the classification system is to achieve classification accuracy equivalent to that obtained with measured HRR signatures. Results suggest that a large portion of HRR signature content is nondiscriminatory. The wavelet denoising process removes the nondiscriminatory information, thereby leading to remarkable increases in classification accuracy. Results are shown for HRR signatures from six aircraft     

Classification of radar targets using synthetic neural networks

Radar target classification performance of neural networks is evaluated. Time-domain and frequency-domain target features are considered. The sensitivity of the neural network algorithm to changes in network topology and training noise level is examined. The problem of classifying radar targets at unknown aspect angles is considered. The performance of the neural network algorithms is compared with that of decision-theoretic classifiers. Neural networks can be effectively used as radar target classification algorithms with an expected performance within 10 dB (worst case) of the optimum classifier     

Radar target classification using doppler signatures of human locomotion models

The problem of target classification for ground surveillance Doppler radars is addressed. Two sources of knowledge are presented and incorporated within the classification algorithms: 1) statistical knowledge on radar target echo features, and 2) physical knowledge, represented via the locomotion models for different targets. The statistical knowledge is represented by distribution models whose parameters are estimated using a collected database. The physical knowledge is represented by target locomotion and radar measurements models. Various concepts to incorporate these sources of knowledge are presented. These concepts are tested using real data of radar echo records, which include three target classes: one person, two persons and vehicle. A combined approach, which implements both statistical and physical prior knowledge provides the best classification performance, and it achieves a classification rate of 99% in the three-class problem in high signal-to-noise conditions.     

Adaptive boosting for SAR automatic target recognition

The paper proposed a novel automatic target recognition (ATR) system for classification of three types of ground vehicles in the moving and stationary target acquisition and recognition (MSTAR) public release database. First MSTAR image chips are represented as fine and raw feature vectors, where raw features compensate for the target pose estimation error that corrupts fine image features. Then, the chips are classified by using the adaptive boosting (AdaBoost) algorithm with the radial basis function (RBF) network as the base learner. Since the RBF network is a binary classifier, the multiclass problem was decomposed into a set of binary ones through the error-correcting output codes (ECOC) method, specifying a dictionary of code words for the set of three possible classes. AdaBoost combines the classification results of the RBF network for each binary problem into a code word, which is then "decoded" as one of the code words (i.e., ground-vehicle classes) in the specified dictionary. Along with classification, within the AdaBoost framework, we also conduct efficient fusion of the fine and raw image-feature vectors. The results of large-scale experiments demonstrate that our ATR scheme outperforms the state-of-the-art systems reported in the literature     

基于遗传算法的目标检测定位方法

将目标的检测定位归结为一类组合优化问题,利用遗传算法的进化策略,采用改进的遗传算法,提出了2种基于遗传算法的目标检测定位方法:基于遗传算法的相关匹配定位法和图象序列目标检测定位法。并给出了在可见光图象序列上的实验结果。