Radar Target Recognition Based On Matching Pursuit And Manifold Learning Research

Radar target recognition has become an important research topic in modern radar technology, and it goes from only recognizing the categories of targets with low-resolution radar to judging the model of targets using the current high-resolution radar. Matching pursuit, which was proposed in1993, is a nonlinear adaptive decomposition method and can be used to feature extraction of signals and time-frequency analysis. Manifold learning is a kind of recently popular machine learning method and aims to find the low-dimensional manifold embedded in the high-dimensionality data. It is widely used for feature extraction and data visualization.In the thesis, based on the previous work, both matching pursuit and manifold learning are applied to radar target recognition. Moreover, in order to overcome the weakness of existing methods, several improved algorithms are proposed. The main work of the paper is as follows:1. The feature extraction method of the time-domain scattering field based on wave atoms is studied, which are consistent with the underlying wave phenomenology. Different from the traditional Gabor atoms, a new over-complete dictionary is constructed, which is composed of wave-front atoms and resonant atoms. The experiment result shows that even a small number of wave atoms can express the major feature of scattering field data. At the same time, the algorithms have good noise-suppression ability.2. The wave-based matching pursuit is used to extract the feature of scattering data, and the feature is subsequently used to construct the codebook via the vector quantization. There are the angular sectors which the scattering physics is slowly varying, and can be used to construct the hidden Markov models (HMM) in order to classify the targets. The experiment result shows that the algorithm has better performance of identification.3. Manifold learning alogrithm is studied for feature extraction from radar range profiles. By modifying the neighborhood preserving projection (NPP) with label information, neighborhood preserving discriminant projection (NPDP) algorithm is proposed, which not only keeps the local topological structure, but also maximizes the inter-class scattering matrix and reduces the intra-class scattering matrix. At the same time, in order to better solve nonlinear problem, by introducing the kernel function, kernel neighborhood preserving projection (KNPP) is obtained. By considering two distinct sets of reconstruction coefficients that can be computed from the data of same and different identities:intra-class weight matrix and inter-class weight matrix, kernel discriminant embedding (KDE) algorithm is proposed. The experiment shows that the proposed algorithm has better performance of classification.