【作者】 邹长忠；

【导师】 夏又生；

【作者基本信息】 福州大学 ， 应用数学， 2018， 博士

【摘要】 高光谱图像是遥感图像处理领域非常重要的一类图像。相对于多光谱图像或者全色图像,目前设备所获取的高光谱图像空间分辨率较低,这限制了高光谱图像的一·些实际应用。因此提高高光谱图像的空间分辨率具有重要意义,已经成为近几年的研究热点。现有的高光谱图像超分辨技术和方法对噪声不鲁棒,致使其超分辨的效果不佳。另外有关泊松噪声或者高斯-泊松混合噪声情形高光谱图像超分辨的研究工作甚少。本文围绕不同噪声干扰的高光谱图像超分辨问题,展开了深入研究。主要的工作和创新点如下:1.在高斯噪声环境下,提出了一种基于双约束光谱解混技术的高光谱图像超分辨方法。通过建立端元矩阵的互距离约束正则项和丰度矩阵的拉普拉斯图模型正则项,解决了现有的高光谱图像超分辨方法对强噪声不鲁棒问题。另外,在计算拉普拉斯图结构的权重矩阵中,引入基于学习的权重计算方法。将权重更新和图像估计统一在一个优化问题中,从而能自适应地更新丰度矩阵,提高了估计精度。实验结果表明本文提出的方法优于目前的其他方法。2.在泊松噪声环境下,提出了一种基于稀疏表示和非局部双正则化的高光谱图像超分辨方法。而且利用主成分分析方法对高光谱图像降维处理,减少需要优化的变量数,降低了计算复杂性和提高了估计精度。提出符合泊松噪声分布的数据保真项,利用表示图像稀疏特性的稀疏表示正则作为其中一个正则项,同时引入非局部正则表示图像的空间非局部相似性,从而有效地解决稀疏表示含有的伪信息问题,进一步提高了估计精度。最后利用增广的拉格朗日交替法求解,实验结果表明本文提出的方法能有效地处理泊松噪声的高光谱图像超分辨问题。3.在高斯-泊松混合噪声环境下,提出了一种处理混合噪声情况的高光谱图像超分辨方法。首先建立混合噪声情形下的数据保真项,接着采用针对高光谱图像的基于非参数贝叶斯理论的字典学习方法,克服了传统的字典学习方法的不足,最后将非参数贝叶斯字典学习融合到图像估计优化式中。在学习字典时,采用变分贝叶斯方法推导相关变量的后验概率分布,利用增广拉格朗日优化方法估计图像。本文提出的字典学习方法与传统的字典学习方法如K-SVD和OnLineDL等方法相比,在图像恢复效果上有显著的提高;本文提出的方法和其他它高光谱图像超分辨方法比较,能有效地去除重构后的噪声,并使重建的图像有较清晰的边界和纹理。4.在未知模糊核环境下,提出了高光谱图像盲超分辨方法,解决了未知模糊核的高光谱图像超分辨问题。引入了全变分正则作为模糊核约束,并基于光谱解混技术,构建向量型全变分和稀疏表示的双重丰度矩阵正则项,将模糊核和丰度矩阵的估计统一到一个优化式子,最后采用交替优化方法求解。实验表明本文提出的方法能有效处理高光谱图像盲超分辨问题。更多还原

【Abstract】 Hyperspectral image has many applications.Due to various hardware limitations,the acquired hyperspectral images unfortunately have lower spatial resolution than mul-tispectral image or panchromatic image,which influences its real applications.So,it is desirable to develop software techniques to enhance the spatial resolution of hyperspec-tral image.The present methods for hyperspectral image super resolution have less satis-factory performance in robustness against heavy noise.Furthermore,there is not almost literatures about hyperspectral image super resolution in the case of Poisson noise or the mixed Gaussian-Poisson noise.This dissertation forcuscs on hyperspectral image super resolution in various noise.Our major works are summarized as follows:1.A novel method is proposed for hyperspectral image super resolution by a nov-el double regularization unmixing-based technique.The widely used linear observation model is combined with the linear spectral mixture model to form the likelihoods of the observations.The structure properties of the endmember and abundance are introduced as prior information to regularize this ill-posed problem.In detail,the mutual distance be-tween the endmember elements is proposed as the regularization for endmember matrix,and the based on graph laplacian regularization term is presented as the abundance ma-trix regularization.Furthermore,the weight matrix of the graph laplacian is computed by graph learning model.Finally,the designed optimization problem is effectively solved by an alternating direction optimization algorithm.Simulation results illustrate that the pro-posed method has a better performance than several well-known methods,both in terms of quality indexes and reconstruction visual effect.2.A novel method based on sparse representation and nonlocal regularization is p-resented for Poissonian hyperspectral image super resolution.We use dimensionality re-duction technique for hyperspectral image by PCA,which brings the more efficient com-putation and the more accurate estimate since the number of variables to be estimated is significantly reduced.The super resolution scheme is designed as an optimization prob-lem whose cost function consists of the two data-fidelity terms about Poisson distribution,the sparse representation term,and the nonlocal regularization term.The sparse represen-tation term is used for enhancing the quality of sparsity-based signal reconstruction,and the nonlocal regularization term exploits the spatial similarity of hyperspectral image and reduces the artificial information brought by sparse representation.Finally,the designed optimization problem is effectively solved by an alternating direction optimization algo-rithm.Simulation results illustrate that the proposed method has a better performance than several well-known methods,both in terms of quality indexes and reconstruction visual effect.3.A novel method is developed for hyperspectral image super resolution using a Bayesian nonparametric dictionary learning method in the presence of mixed Gaussian-Poisson noise.An optimization model is introduced,including the data-fidelity term cap-turing the statistics of mixed Gaussian-Poisson noise,and a beta process analysis-based sparse representation regularization term.Compared with conventional dictionary learn-ing methods,such as K-SVD and OnLineDL methods,the introduced dictionary learning method is based on a popular beta process factor analysis(BPFA)for an adaptive learning performance.In order to implement the proposed method,we use alternating direction algorithm of multipliers(ADMM)for simultaneous Bayesian nonparametric dictionary learning and image estimation.Variational Bayesian inferring is used for the dictionary learning.Simulation results illustrate that the proposed method has a better performance than several well-known methods in terms of quality indices and reconstruction visual ef-fect.4.A novel hyperspectral image blind super resolution method is proposed in the case of unknown blur kernel.The simultaneous total variation and sparse representation are proposed as abundance regularization terms,while a total variation regularization term is selected for the blur kernel.Because the image and blur kernel are simultaneously estimat-ed,the estimated result error is minimized.Finally,the proposed optimization formulation is effectively solved by block coordinate descent method.Simulation results show the fea-sibility of the proposed method and its advantages over existing approach from two aspects of visual effectiveness and quality indices.更多还原