SAR Image Recognition Based On Manifold Learning

Synthetic aperture radar (SAR) has high resolution in both range and azimuth, and has the characteristics of strong penetration and being all-weather. It is widely used in military, geological exploration, environmental monitoring and other fields. In this thesis, we will study SAR image recognition. Specifically include:1. The presence of speckle noise, background and shadow in SAR images will reduce the performance of recognition. Thus accordingly, five filter methods have been studied in this thesis, including mean, median, Lee, Gamma Map and enhanced directional smoothing. And target segmentation, geometric clustering, energy normalization and a series of preprocessing have been made. Simulation results show that: 1) enhanced directional smoothing filter has better performance than others in detail preserving and noise smoothing; 2) the CFAR segmentation based on Weibull distribution outperforms the one based on K distribution, which can effectively mark the target area; 3) geometric clustering can effectively remove the small acnode left by image segmentation.2. Compared with the globally linear feature extraction algorithms, the feature extraction algorithm based on manifold learning has a better ability to retain geometric structure, So the linear feature extraction algorithms based on manifold learning are studied.The simulation results show that: 1) the methods taking the label information into account have a better recognition performance than those without using category information.The use of class information can reduce the probability of the presence of heterogeneous data in the neighborhood of each data; 2) methods with statistical independence constraints outperform its corresponding method without this constraint.This constraints can reduce the redundancy of the extracted feature; 3) two-dimensional methods is less sensitive to training numbers than the one-dimensional methods.The two-dimensional methods avoid vectorization and have lower data dimension, which can reduce the singularity of data; 4) the method by fusion of PCA and CLPP proposed in this thesis achieves better performance than both PCA and CLPP. CLPP only takes into account the local characteristics, and does not consider the global, so if it's combined with the global feature extraction algorithm PCA,it will have a better recognition performance.3. SAR image is affected by noise, azimuth and other factors, so its data distribution can't be completely linear. Nonlinear methods can overcome this problem.The nonlinear feature extraction algorithms based on manifold learning will be studied in this thesis. Experiments show that: 1) compared with the corresponding linear method, nonlinear method has a better recognition effect. Projecting data to high dimensional space through nonlinear transform will enhance the linear separability of data; 2) KPCA and CLPP fusion algorithm proposed in this thesis has a better recognition performance than others. The distribution of data can be regarded as linear globally and nonlinear locally, KPCA is a global nonlinear algorithm and CLPP is a local linear algorithm, so the fused features have a better recognition result.