Non-parametric Bayesian Learning for Geometric Reconstruction of Scattering Primitives from Multi-Dimensional SAR Images

Target geometrical reconstruction from multi-dimensional SAR(synthetic aperture radar) images is very important for target recognition and scene understanding. This paper proposes a novel geometric reconstruction method for man-made targets based on scattering primitives, e.g., plate, cylinder, dihedral, trihedral and so on. Non-parametric Bayesian learning is employed to optimize the scatterer parameters into an integral computation. The Bayesian learning can extract scattering primitives of targets via a reversible jump Markov chain Monte Carlo strategy. Experimental results of two models demonstrate good feasibility of the proposed method.