Flexible Priori Proximal Bayesian Learning for High-Resolution SAR Imagery

Abstract

The conventional high-resolution synthetic aperture radar(SAR) imagery based on Bayesian learning encounter the problems of static and inflexible of prior. In this paper, a novel Bayesian learning algorithm based on flexible prior (FP-Bayes) is proposed for high-resolution SAR imagery, in which a new flexible prior is introduced to represent the rich features of target scene. However, the model becomes complicated due to the intended prior, which is hard to be solved by the classic Markov Chain Monte Carlo (MCMC) sampling algorithms. Therefore, the Hamiltonian Monte Carlo (HMC) is applied to this problem. Considering the potential function of the posterior distribution is likely to be non-differentiable, the proximal operator is introduced, and the proximal-HMC (P-HMC) is developed, which is capable of solving the sampling problem of this posterior. To this end, the high-resolution SAR images can be obtained. In the experimental part, the raw SAR data is utilized to verify the effectiveness and superiority of the proposed algorithm. The phase transition diagram (PTD) is also adopted to analyze the imaging performance quantitatively.