Research On Algorithms Of HRR Target Recognition Based On Kernel Method

The technology of radar target recognition has become an important branch of radar area. However, due to the complex and variable application environments, strict requirements on the recognition speed, difficulties in acquiring samples of the non-cooperative targets and low efficiency in employing the available data, this technology is in fact faced with many challenges. As is known, high range resolution profile (HRRP) can reflect the target's geometry information along the radial line. In addition, it is relatively easy to obtain and process HRRP. Therefore, research on radar target recognition based on HRRP is currently extensively carried out. In order to improve the stability and recognition performance of the existing algorithms, the dissertation carries out a research from such aspects as feature extraction, classifier design and parameter selection, with the expectation to solve the problem of learning and recognition, sequential information extraction and usage and parameter selection under the condition of huge volume samples with high dimension.In chapter one, the background and relative research is introduced. Also, the main work of this dissertation is summarized.In chapter two, the characteristics of both kernel perceptrons and time-delay neural network are analyzed, based on which a combinational kernel perceptron classifier and a fast algorithm of time-delay neural network are proposed.Chapter three concentrates on the effort to solve such typical problems in radar target recognition as learning in samples with high-dimension and huge volume, sequential information extraction and parameters selection. A kernel time-delay perceptron algorithm based on the kernel method and time-delay units is proposed. Firstly, Secondly, in order to better tradeoff the contradiction between the sufficiency and efficiency in utilizing information, a kind of kernel time-delay perceptron network based on kernel method and time-delay units is proposed. In addition, the learning algorithm of kernel time-delay perceptron is constructed. The algorithm can not only solve the learning problem conditioned on samples with high dimension and huge volume, but also extract and employ the sequential information without any increase in the calculation burden. Thirdly, a Gaussian kernel parameter selection method based on the distribution characteristics of the samples is proposed, so that the speed of parameter optimization is improved a lot. Moreover, under different conditions, the influence of parameters in kernel time-delay perceptron to the classification rate, generalization ability and complexity is analyzed. It is demonstrated by simulation results that the algorithm performs well on both classification rate and noise suppression ability.Chapter four investigates mainly on two problems, which are target recognition with the deficiency of training samples of one kind of target and multi-classifier fusion, respectively. In order to solve the problem of over large rejection region in SVDD multi-class recognition algorithm, two kinds of partition scheme of rejection region are proposed. Based on this work, two kinds of corresponding SVDD multi-class classifier are constructed and also the application characteristics of each classifier are analyzed. In order to improve the performance of radar target recognition, a new classifier based on multi-algorithm decision fusion is constructed. Computer simulations show that the classifier outperforms the single classifier both in classification rate and noise suppression ability.