Study On Dnowledge-Based Radar Adaptive Processing

As an important sensor, radar gains information through interaction with target and scene which it illuminates. In complex battle field, it is hard to have good performance with fixed signal processing procedure and only the information about target and environment contained in radar echoes. This is the problem that conventional radar faces and needs to be solved for further development. An effective strategy to circumvent this problem is to exploit some prior information about the scene illuminated by radar, namely to perform a cognitive knowledge-based processing. It is also regarded as an important part of the next generation radar. Therefore, how to exploit prior knowledge efficiently in radar adaptive processing has recently received intensive attention from radar signal processing community. This dissertation mainly focuses on knowledge-based radar signal processing with application on radar adaptive detection and image processing. The main research work is summarized as follows:1. Although General Inner Product (GIP) and Sample Matrix Inverse (SMI) are two effective non-homogeneous detectors (NHD), they show notable performance degradation when disturbance covariance matrix (CM) is estimated with secondary data contaminated by target-like signals. To solve this problem, a knowledge-based (KB) NHD scheme is presented, which incorporates prior knowledge sources of clutter when selecting secondary data for radar adaptive processing. Meanwhile, as an example, the proposed KB NHD is used to select radar adaptive filtering algorithms intelligently. In the example, returned data that radar collects are first separated into homogeneous and non-homogeneous sectors. For different sectors, different algorithms are accordingly applied. The performance of disturbance suppression and target detection are improved via the algorithm selection approach.2. It is well known that for airborne radar, the location of the ground clutter spectrum in angle-Doppler plane is dependent mainly on the platform velocity and radar parameters. A robust non-homogeneous clutter suppression algorithm is proposed, which makes use of such prior information. Similar to Generalized Multiple Beams (GMB) and Joint Domain Localized (JDL) algorithms, the proposed algorithm also localizes clutter in angle-doppler plane. The main distinction is that the characteristic structure of clutter ridge is considered here. Two-dimension Gaussian power spectral density model is incorporated and the parameter of the model can be obtained by exploring the sensed environment and returned data. The open-loop algorithm has low computation cost and do not require iteration convergence.3. Bayesian radar adaptive detection is discussed. When designing conventional adaptive radar receiver, disturbance CM is modeled as deterministic and unknown parameter, meanwhile, homogeneous secondary data are used for its maximum likelihood (ML) estimation. In this way, information about target and environment is only obtained from secondary data. Enough homogeneous secondary data can result in nearly optimal performance; otherwise, notable performance degradation will be observed. To solve this problem, a knowledge-based Bayesian framework is proposed, which considers disturbance CM as random variable with appropriate prior distribution. Meanwhile, the prior knowledge about environment is used in the prior distribution. Within this framework, Bayesian General Likelihood Ratio Test (GLRT), Adaptive Mached Filter (AMF), Rao and Wald tests are devised and a scale variable is introduced to tune the amount of prior knowledge used. These Bayesian detectors exploit information from both secondary data and environment illuminated, thus improve the detection performance when only a small number of secondary data is available.4. A statistical model on heterogeneity of clutter is proposed, which makes it possible to incorporate heterogeneity in the design of radar adaptive detectors. The model takes into consideration of both the scale and structure difference between the diturbance CMs of cell under test (CUT) and secondary data. It assumes the two covariance matrixes are not same with probability one but have some similarity which can be tuned by a scale variable. With proper prior distribution on CM, information about environment is also considered when design adaptive detectors. Finally, a close form of Bayesian GLRT detector is proposed based on the model, which can improve the detection performance.5. SAR image pre-processing plays an important role in the application of image based radar detection and identification. The conventional image filters can not result in satisfied performance, since they have fixed structure and only use image data within the sliding window. To solve this problem, over-complete dictionary learning based on non-parametric Bayesian method is proposed and applied to image (including SAR image) processing. Specifically, dictionary learning is cast as a factor analysis (FA) problem. Utilizing nonparametric Bayesian methods, such as the beta process (BP) and the Indian buffet process (IBP), one may infer the dictionary with the data form the whole image and information obtained from previous processing stagea. With all these information, the performance of image denoising and interpolation is improved.