Kernels Evaluation of SVM-Based Estimators for Inverse Scattering Problems

Buried object detection by means of microwave-based sensing techniques is faced in biomedical imaging, mine detection, and many other practical tasks. Whereas conventional methods used for such a problem consist in solving nonlinear integral equations, this article considers a recently proposed learning by examples approach [1] based on Support Vector Machines, the techniques that proved to be theoretically justified and effective in real world domains. The article considers the approach performance for two different kernel functions: Gaussian and polynomial. The obtained results demonstrate that using polynomial kernels along with slightly sophisticated model selection criterion allow to outperform the Gaussian kernels. Simulations have been carried out for synthetic data generated by Finite Element code and a PML technique; noisy environments are considered as well. The results obtained by means of polynomial and Gaussian kernels are presented and discussed.