非参数Bayesian字典学习的遥感影像融合方法

Remote Sensing Images Fusion Based on Bayesian Nonparametric Dictionary Learning Approach

针对现有高光谱成像技术因硬件条件限制而较难获取兼具高光谱及高空间分辨率高光谱影像的问题,提出一种基于非参数Bayesian字典学习的高光谱与多光谱影像空谱融合方法.该方法将目标影像的融合问题投影转换至较低维度的子空间内:首先,利用Beta-Bernoulli process的非参数Bayesian方法对观测影像进行字典学习,建立各隐变量的概率分布模型,使用Gibbs抽样方法来计算字典元素的后验分布;然后,采用正交匹配追踪算法进行稀疏系数学习;最后,采用交替方向乘子对目标图像及对应的稀疏系数进行交替优化更新,通过最小化目标函数来最大化目标影像的估计值.实验结果表明,方法因加入了更多的先验信息而获得较高的定量评价指标及目视效果,具有一定的普适性.

Aiming at the problem that it is difficult to obtain hyperspectral images with both high hyperspectral and high spatial resolution due to hardware limitations of existing hyperspectral imaging technology,an image space-spectrum fusion method based on Bayesian nonparametric dictionary learning is proposed.The algorithm performs image fusion in the low-dimensional subspace,and firstly uses the Bayesian nonparametric approach with the Beta-Bernoulli process to learn the dictionary of the observed images,which establishes the probability distribution models for each latent variable and calculates the posterior distributions by Gibbs sampling.Secondly,the orthogonal matching pursuit approach is adopted to learn the sparse coefficients.Finally,the alternating direction multiplier is used to alternately optimize the target image and the corresponding sparse coefficient,and the estimated value of the target image is maximized by minimizing the objective function.Experiments show that the algorithm can obtain higher quantitative evaluation index and visual effect with more prior experience added,which can be performed universally.