SAR Image Target Recognition Research Based On Kernel Function

Synthetic aperture radar(SAR) works in the microwave band and has the function of coherent imaging. It is able to achieve all-time reconnaissance of targets in harsh climate and obtain a large area of two-dimensional high-resolution images. These images provide us with the data source effective for target classification and recognition which has a wide range of application. In recent years, SAR automatic target recognition(ATR) has emerged to enhance the efficiency of information processing and improve the accuracy of target recognition at the same time, and it will be an important research direction in the military field. This paper focuses on the SAR image preprocessing and feature extraction and verifies the applicability of pretreatment and studies influence of various feature extraction methods for target recognition. The main contents are as follows:1. According to the particularity of SAR images, the systemic preprocessing methods for MSTAR data are adopted. Through enhanced Lee filter and median filter, the original SAR image speckle noises are suppressed effectively and edge details are retained. Gray enhancement by power transformation is used to improve the contrast of the image and the identification ability of target. Using two-parameter constant false alarm rate(CFAR) image segmentation, the target region of interest is separated from a complex background clutter. Through the centroid registration and energy normalization, the influence that target scattering echo intensity caused by different distance between target and radar will be overcome.2. The method that kernel principal component analysis(KPCA) is studied, which can effectively solve the problem that traditional PCA is not conducive to extract nonlinear characteristics. In addition, we research kernel two-dimensional PCA(K2D-PCA) method, which directly uses 2D-PCA in kernel space. It solves the problem of nonlinear feature extraction and is less dependent on number of samples and reduces the precision requirement of target azimuth information while two-dimensional image structure information is kept.3. For the feature extraction method based on linear discriminant analysis(LDA) uses one-dimensional vector model which results in curse of dimensionality and small sample size problem, kernel LDA(KLDA) and its improved method kernel weighted maximum margin criterion(KWMMC) are presented and the method preserving image structure information based on kernel two-dimensional LDA(K2D-LDA) is researched. K2D-LDA utilizes the constructed kernel sample matrix directly adopting 2D-LDA method in kernel space, which is robust to the variation of target azimuth under the premise of solving the small sample size problem.