Research On SAR Automatic Target Recognition Based On Curvelet Transform And Kernel Support Vector Machine

Kernel Machines (KM) and Multi-scale Geometric Analysis (MGA) have already come in many disciplines and have achieved plentiful fruits in diversified fields, which include signal processing, image processing, pattern recognition, information retrieval, data mining and automatic controlling, etc. From the techniques of Curvelet Transform and KM Learning methods, this dissertation studied Synthetic Aperture Radar automatic target recognition (ATR) and occluded target recognition relevant techniques. Based on feature extraction and machine learning algorithm in SAR ATR system, three algorithms have been proposed, which can be summarized as follows:1. Aiming at extracting feature of SAR target, a kind of SAR ATR method based on Curvelet transform is developed. SAR target consist of the content information and contour information feature by Curvelet Transform, which contour information is to compensate content information feature error brought by segmentation. The final features consisting of content information and contour information feature are further classified by SVM. Experiments prove that our proposed algorithm extract a more effective target feature and improve the recognition rate. Besides, we also use this algorithm in SAR occluded target recognition problem, experiments also show our algorithm the stability and efficiency of SAR occluded target recognition.2. Because of multi-resolution property in frequency and time domain and approximation ability of wavelet, we constructed Meyer wavelet kernel by combination of wavelet and kernel support vector machines. Wavelet kernel compensate for the disadvantages of Gaussian kernel on approximating functions with specific singularities, and the approximation ability of wavelet kernel is similar to Gaussian kernel for smooth function. Experiments on regression estimation and SAR target recognition illustrate efficiency of the approximation ability of our algorithm and a better recognition ability of SAR target. The construction of wavelet kernel function is to provide many choices of ML kernel function.3. A kernel function constructs the relevant feature space. Wavelet kernel is consi- dered the best base in 1-D space, which can not be extended into high dimension. Therefore, we improved a kernel support vector machines algorithm based on curvelet principle, which depends on quite approximation ability of high dimension information of MGA. This method consist of optimized mapping direction by using Immune Clone Selection algorithm and wavelet kernel mapping on the optimized direction, which can achieve a more efficient approximation. Experiments on regression estimation and SAR target recognition show the efficiency and generalization.