Research On Sar Image Classification Based On Kernel Learning

In recent years, SAR (Synthetic Aperture Radar) is developed in many fields, such as military and civilian institutions. We mainly discuss SAR image recognition techniques.Dimensionality reduction and classification are crucial issues for feature extraction and SAR image recognition. In this paper, we firstly introduce the most famous algorithm: Principle Component Analysis (PCA) and non-linear dimensionality reduction methods (manifold). Then we introduce Locality Preserving Projections (LPP) classification method. Kernel method is developed from statistical learning theory which is a new method. Kernel learning (kernel trick) is introduced in chapter3. Then we proposed two novel kernel-based SAR image classification methods:AQKPCA and SGKLPP.The main contribution is organized as follows:Firstly, Kernel-based novel method, Adaptive Quasiconformal Kernel Principal Component Analysis (AQKPCA) is proposed in this paper. The main problems of kernel learning are construction of kernel function, optimization of kernel function, and the choice of kernel parameters. In AQKPCA, Maximum margin criterion (MMC) is adopted to optimize the kernel function, which directly avoids the recognition and classification mistakes due to the choice of kernel function. The experimental results indicate AQKPCA performs a better recognition rate than traditional method.Secondly, SGKLPP (Supervised Gabor-based Locality Preserving Projections) is presented in this paper. LPP virtually extended LE (Laplacian Eigenmap) by linear projections, so that we can derive linear classification. In this paper, our method SGKLPP is improved and extended by SKLPP (Supervised Kernel Locality Preserving Projections) is improved and extended. The performances of the experiments illustrate that SGKLPP on ORL and Yale dataset achieve good results.In experiments part, we not only adopt MS TAR SAR dataset but also apply ORL dataset and Yale dataset on experiments. The performances of experiments illustrate the reliability of our methods. Compared with the proposed methods and exist method, our methods are illustrated to be advanced. In this paper, the principle of the classifier and singular value decomposition (SVD) is not explained completely. How to optimize SAR recognition performance and computing cost will be the focal point in future research.