Divide-and-Conquer Variational Bayesian Inference for Multi-task Learning of High-resolution SAR Imagery

Conventional statistical-driven synthetic aperture radar (SAR) imaging algorithms can only encode a single and/or static prior, leading to that limited features can be accessed quantitatively. To this end, a novel multi-task learning framework is proposed by devising a divide-and-conquer variational Bayesian (DC-VB) inference, so that elaborated features of interests can be exploited for a high-resolution SAR imagery. Specifically, a flexible generalized Gaussian distribution (GGD) and a customized hybrid probability distribution are introduced and employed for the priors of features of interests. To resolve the resultant complicated Bayesian inference, splitted random variables are incorporated, so that the joint posterior problem can be decomposed into multiple local problems that are easy to be solved. Simultaneously, dual variables are established for residual errors of the decomposition. To guarantee a global solution for the image of the target of interests, the data augmentation is employed to coordinate multiple local solutions. Therefore, the intended Bayesian inference works in a divide-and-conquer manner, which is superior in quantization of multiple features of high-resolution SAR imagery. It is capable of incorporating multiple priors in a fully-statistical probability and guaranteeing closed-form solutions of posterior distributions. Unavoidable propagation errors can be minimized in the DC-VB process. Raw SAR data is applied to validate the effectiveness of the proposed algorithm. Comparisons with conventions show the superiority in terms of qualitative and quantitative aspects.