Junshui Ma

Accurate on-line support vector regression

Batch implementations of support vector regression (SVR) are inefficient when used in an on-line setting because they must be retrained from scratch every time the training set is modified. Following an incremental support vector classification algorithm introduced by Cauwenberghs and Poggio (2001), we have developed an accurate on-line support vector regression (AOSVR) that efficiently updates a trained SVR function whenever a sample is added to or removed from the training set. The updated SVR function is identical to that produced by a batch algorithm. Applications of AOSVR in both on-line and cross-validation scenarios are presented.Inbothscenarios, numerical experiments indicate that AOSVR is faster than batch SVR algorithms with both cold and warm start.

Online novelty detection on temporal sequences

In this paper, we present a new framework for online novelty detection on temporal sequences. This framework include a mechanism for associating each detection result with a confidence value. Based on this framework, we develop a concrete online detection algorithm, by modeling the temporal sequence using an online support vector regression algorithm. Experiments on both synthetic and real world data are performed to demonstrate the promising performance of our proposed detection algorithm.

Nonlinear multiclass discriminant analysis

An alternative nonlinear multiclass discriminant algorithm is presented. This algorithm is based on the use of kernel functions and is designed to optimize a general linear discriminant analysis criterion based on scatter matrices. By reformulating these matrices in a specific form, a straightforward derivation allows the kernel function to be introduced in a simple and direct way. Moreover, we propose a method to determine the value of the regularization parameter /spl tau/, based on this derivation.

Using support vector machines as HRR signature classifiers

Support Vector Machines (SVMs) are an emerging machine learning technique that has found widespread application in various areas during the past four years. The success of SVMs is mainly due to a number of attractive features, including a) applicability to the processing of high dimensional data, b) ability to achieve a global optimum, and c) the ability to deal with nonlinear data. One potential application for SVMs is High Range Resolution (HRR) radar signatures, typically used for HRR-based Automatic Target Recognition (ATR). HRR signatures are problematic for many traditional ATR algorithms because of the unique characteristics of HRR signatures. For example, HRR signatures are generally high dimensional, linearly inseparable, and extremely sensitive to aspect changes. In this paper we demonstrate that SVMs are a promising alternative in dealing with the challenges of HRR signatures. The studies presented in this paper represent an initial attempt at applying SVMs to HRR data. The most straightforward application of SVMs to HRR-based ATR is to use SVMs as classifiers. We experimentally compare the performance of SVM-based classifiers with several conventional classifiers, such as k-Nearest-Neighbor (kNN) classifiers and Artificial Neural Network (ANN) Classifiers. Experimental results suggest that SVM classifiers possess a number of advantages. For example, a) applying SVM classifiers to HRR data requires little prior knowledge of the target data, b) SVM classifiers require much less computation than kNN classifiers during testing, and c) the structure of a trained SVM classifier can reveal a number of important properties of the target data.

2D HRR Radar Data Modeling and Processing

High Range Resolution (HRR) -based Automatic Target Recognition (ATR) has attracted increasing attention due to a number of potential advantages over alternative radar techniques in moving target identification. Most current HRR-based ATR studies have been conducted using 1D HRR signatures. However, these 1D HRR signatures are generally plagued by scintillation effects, and thus demonstrate highly irregular behavior that dramatically degrades the performance and robustness of algorithms based on these signatures. In order to circumvent this difficulty, an alternative HRR radar data representation and processing technique is presented in this paper. This technique models and extracts the target characteristics directly, based on the 2D HRR raw data. In this paper, we first derive a general, but complex HRR radar model, and then simplify this model by instantiating a set of real-world radar and target parameters for the model. This simplification process produces two HRR radar models with different degrees of simplicity. After establishing this set of models, the typical HRR data processes, such as feature extraction and clutter suppression, are reduced to one problem, which is model-parameter estimation. Based upon the most simplified HRR model we proposed, we devise two model- parameter estimation algorithms. One is a scatterer extraction algorithm based on available 1D Parameter Estimation (1DPE), while the other is based on 2D discrete Fourier Transform (2DFT). In order to examine the performance of these two algorithms a set of simulations are conducted. The experimental results are presented, and the performance comparison between 1DPE and 2DFT is presented.

Parameter estimation algorithms based on a physics-based HRR moving target model

In contrast to Synthetic Aperture Radar (SAR), High Range Resolution (HRR) radar may economically provide satisfactory target resolution when applied to moving targets scenarios. We have devised a series of new physics-based HRR moving target models with different degrees of simplification. These models represent the scatterers from both targets and clutter equally. By employing these models, we can unify the studies of both clutter suppression and target feature extraction into a single topic of model parameter estimation. Therefore, finding reliable parameter estimation algorithms based on these models becomes an important topic for target identification using HRR signatures. This paper derives and presents two feasible parameter estimation algorithms. The first algorithm (1DPE) reduces the 2D-estimation problem to two 1D-estimation problems, and solves the problems by employing some mature 1D-estimation algorithms. The second algorithm (2DFT) utilizes the 2D Discrete Fourier Transform (DFT) to estimate the model parameters by simply applying the 2D DFT to the HRR data, and obtaining the estimation of model parameters from the peaks of the 2D DFT. In order to verify the performance of these algorithms, we performed a series of simulation experiments and the experimental results are presented in this paper. Finally, a brief comparison of these two algorithms is also presented.

Kernel-based feature extraction and its application on HRR signatures

Kernel-based Feature Extraction (KFE) is an emerging nonlinear discriminant feature extraction technique. In many classification scenarios using KFE allows the dimensionality of raw data to be reduced while class separability is preserved or even improved. KFE offers better performance than alternative linear algorithms because it employs nonlinear discriminating information among the classes. In this paper, we explore the potential application of KFE to radar signatures, as might be used for Automatic Target Recognition (ATR). Radar signatures can be problematic for many traditional ATR algorithms because of their unique characteristics. For example, some unprocessed radar signatures are high dimensional, linearly inseparable, and extremely sensitive to aspect changes. Applying KFE on High Range Resolution (HRR) radar signatures, we observe that KFE is quite effective on HRR data in terms of preserving/improving separability and reducing the dimensionality of the original data. Furthermore, our experiments indicate the number of extracted features that are needed for HRR radar signatures.

Complex HRR range signatures

The need to automatically identify moving targets is becoming increasingly important in modern battlefields. However, Synthetic Aperture Radar (SAR) is problematic when applied to moving targets scenarios because moving targets tend to smear SAR images. High-Range Resolution (HRR) Radar has, consequentially, attracted more attention due to its potential performance in moving target identification. However, devising reliable identification techniques using HRR signatures is challenging because the signatures are extremely sensitive to radar aspect angles primarily because of scintillation. This aspect sensitivity causes the HRR signatures to exhibit irregular behavior that makes extracting robust target features a challenge. As a result, HRR applications tend to base their processing on the magnitude of complex HRR signatures. We argue that insightful feature selection shoudl be based on a detailed understanding of the properties of the complex signatures. In this paper we focus on studying the fundamental behavior of complex HRR signatures that are generated from a representative HRR model. Our analysis focuses on (1) scintillation effects; (2) the relationship between HRR signatures and aspect angle, and (3) the utility of the phase of complex HRR signatures. In this paper we present a number of observations concerning the redundancy of phase information, the variance of HRR signatures as a function of aspect angle, and the relationship between scattering coefficients and scatterer locations.

Eigen indexing in satellite recognition

In many image analysis problems it is possible to take advantage of the structural relationships between various parts of the objects being imaged in order to index the images of the objects. For example, many satellites consists of a main body and outlying sub-components. Thus, in many circumstances satellites can be indexed in a model database by the distinct structural relationships between their sub- components. However, algorithms based on structured sub- components necessitate the use of robust and reliable 2-D image segmentation techniques to successfully partition images into their sub-components. Unfortunately, this segmentation task can be highly problematic for objects with complex components and under harsh, unfavorable lighting conditions. The research presented here describes a new method to compute indices which can be used for image indexing without image segmentation. We use satellite imagery as a convenient image class for which to demonstrate our method. Our method partitions the image into many small equal-area pieces. We refer to this technique as differentiation. Differentiated images result in a set of sub-images that collectively represent the structural information inherent in the image. We prove that a primitive matrix with at most four non-zero eigenvalues can be constructed from the differentiated image. This property (1) significantly reduces storage requirements for a model database, (2) reduces the computational burden of subsequent recognition processes, and (3) supports an efficient and accurate matching procedure. To evaluate the efficiency of our algorithm for a recognition application, we use boundary methods as a feature set evaluation method to quantify the utility of the eigen-indexes obtained by our method as compared to other existing indexing methods.

Derivation of physics-based HRR moving target models

Although SAR has demonstrated excellent performance in stationary target identification, SAR resolution suffers in moving target scenarios. High Range Resolution (HRR) radar appears to be an attractive alternative in applications to moving target identification because HRR target signature can provide target scatterer information with high range resolution. Since many HRR processing steps, such as feature extraction and clutter suppression, are based on underlying modeling assumptions, devising a reliable physics-based HRR model for moving targets has become an increasingly important topic. In this paper, we derive a scattering-based HRR moving target model. However, the general form of the derived model is quite complex, and this complexity makes subsequent analysis difficult. We therefore simplify the complex model to obtain different simplified versions that facilitate the utility of the models. Simplifications is achieved by instantiating the parameters in this model with radar and target parameters from the real world, and then retaining only those terms with dominant value. A series of reliable, yet theoretically tractable models, are obtained with different degrees of simplification. The contributions of this paper are as follows: (1) Two new physics-based HRR moving target models with different degrees of simplification are presented; (2) These models make no assumptions regarding the distribution of the clutter; (3) Performance boundaries on the subsequent feature extraction algorithms are derived and delineated.