双层稀疏贝叶斯学习ISAR超分辨成像算法

Super-resolution ISAR imagery algorithm based on bi-sparsity Bayesian learning

传统贝叶斯成像常采用拉普拉斯分布进行成像特征表征,易使得成像结果过稀疏而容易丢失部分弱散射的结构特征,进而影响逆合成孔径雷达(inverse synthetic aperture radar,ISAR)成像精度提升。为实现高精度ISAR超分辨成像,本文采用伯努利拉普拉斯混合稀疏先验对目标统计特性进行概率建模,利用双层稀疏对目标先验进行统计约束,从而有效模拟目标散射场统计先验。并在贝叶斯层级模型下,通过引入隐变量建模的方式对先验进行分层构建,在解决先验分布与高斯似然函数不共轭问题的同时简化贝叶斯推断,降低模型复杂度。为避免繁琐的手动参数调整,实现超参数的自调节,本文对各随机变量建立条件概率依赖模型,并利用马尔可夫链蒙特卡罗随机模拟估计算法解决高维积分和后验分布难以求解的问题,实现相关超参数的统计估计,提升算法自学习能力。仿真和实测数据均证明本文所提方法具有有效性和优越性。

The Laplacian distribution is often used to characterize imaging features in the conventional Bayesian imaging,which makes the image over-sparse.It is easy to lose the weak scattering characteristics of some structural features,which in turn affects the improvement of inverse synthetic aperture radar(ISAR)imaging accuracy.In order to effectively achieve ISAR super-resolution imaging,Bernoulli-Laplace(BL)mixed sparsity priori is adopted in this paper to formulate the statistical characteristics of the target,and the bi-sparsity model is applied to constraint the imaging target prior.Under the Bayesian hierarchical model,the prior is hierarchically constructed by introducing latent variables to simplify the Bayesian inference and reduce the complexity of the model.The problem that the prior distribution and the Gaussian likelihood are not conjugated is solved.In order to avoid tedious manual parameter adjustment,conditional probability functions are established in this paper for random variables,and the Markov chain Monte Carlo(MCMC)sampling algorithm is used for the solution,so that high-dimensional integration can be avoided and analytical posteriors can be obtained.All the hyper-parameters can be fixed automatically,and the proposed algorithm can be performed without too much manual interventions.Both simulated and measured ISAR data validate the effectiveness and superiority of the proposed algorithm.