基于中心对齐多核学习的稀疏多元逻辑回归算法

稀疏多元逻辑回归(SMLR)作为一种广义的线性模型被广泛地应用于各种多分类任务场景中。SMLR通过将拉普拉斯先验引入多元逻辑回归(MLR)中使其解具有稀疏性,这使得该分类器可以在进行分类的过程中嵌入特征选择。为了使分类器能够解决非线性数据分类的问题,该文通过核技巧对SMLR进行核化扩充后得到了核稀疏多元逻辑回归(KSMLR)。KSMLR能够将非线性特征数据通过核函数映射到高维甚至无穷维的特征空间中,使其特征能够充分地表达并最终能进行有效的分类。此外,该文还利用了基于中心对齐的多核学习算法,通过不同的核函数对数据进行不同维度的映射,并用中心对齐相似度来灵活地选取多核学习权重系数,使得分类器具有更好的泛化能力。实验结果表明,该文提出的基于中心对齐多核学习的稀疏多元逻辑回归算法在分类的准确率指标上都优于目前常规的分类算法。     

一种新的两分类器融合的空谱联合高光谱分类方法

本文提出一种两分类器融合的高光谱空谱联合分类方法,首先利用子空间多项式逻辑回归在图像的特征子空间中分类,得到满概率图;根据满概率将每个像元分至概率最大的两个最可信类别,并在原始空间中构建最可信类别字典,利用稀疏解混对每个像元在最可信类别字典下进行稀疏表示,得到稀疏概率图;最后将满概率图和稀疏概率图线性融合,并利用边缘保持的马尔可夫正则项挖掘图像空间信息,得到具有边缘保持的空谱分类模型.实验表明,提出的两分类器融合方法即使在训练样本较少时也比现有方法得到更好的分类结果.     

复合正则化联合稀疏贝叶斯学习的高光谱稀疏解混算法

将稀疏贝叶斯学习引入线性混合像元分解中,提出一种基于复合正则化联合稀疏贝叶斯学习的高光谱稀疏解混算法.在多观测向量的稀疏贝叶斯框架下,对各参数建立概率模型,经贝叶斯推断得到基于L2,1正则化的联合稀疏贝叶斯解混模型,并将丰度向量的非负与和为一约束加入到凸优化的目标函数中,通过变量分离法将复合正则化问题分解成多个单一正则化问题交替迭代求解,并利用参数自适应算法对正则化参数进行更新.模拟数据和真实数据的实验结果表明,该算法比贪婪算法和凸优化算法能获得更高的解混精度,并且适用于端元个数较多和信噪比较低的高光谱数据.     

近似稀疏约束的多层非负矩阵分解高光谱解混

稀疏正则化函数的选取直接影响到稀疏非负矩阵分解高光谱解混的效果。目前,主要采用L0或L1范数作为稀疏度量。L0稀疏性好,但求解困难;L1求解方便,但稀疏性差。提出一种近似稀疏模型,并将其引入到多层非负矩阵分解（AL0-MLNMF）的高光谱解混中,将观测矩阵进行多层次稀疏分解,提高非负矩阵分解高光谱解混的精度,提升算法的收敛性。仿真数据和真实数据实验表明:该算法能够避免陷入局部极值,提高非负矩阵分解高光谱解混性能,算法精度上比其他几种算法都有较大的提升效果,RMSE降低0.001~1.676 7,SAD降低0.002~0.244 3。     

光谱加权协同稀疏和全变差正则化高光谱图像解混

针对传统稀疏解混方法对丰度的稀疏性表征不充分及空间信息利用率低等问题,本文在分析迭代加权稀疏解混方法的基础上,提出了一种基于光谱加权协同稀疏和全变差正则化的高光谱解混方法.该方法一方面在协同稀疏解混的基础上引入光谱加权因子进一步刻画丰度系数的行稀疏性,以促进所有像元之间的联合稀疏性;另一方面引入各向异性全变差空间正则化促进图像同质区域的平滑性,以提高解混的准确性.通过交替方向乘子法求解该模型,通过迭代,利用内外部双循环迭代方法对光谱加权因子和丰度系数进行优化.模拟和真实的高光谱数据实验结果均表明本文提出的算法与现有同类算法相比能大幅提高混合像元分解的精度,在稀疏解混方面展现出了巨大的潜力.     

基于稀疏贝叶斯的流形学习

针对当前监督学习算法在流形数据集上分类性能的缺陷,如分类精度低且稀疏性有限,本文在稀疏贝叶斯方法和流行正则化框架的基础上,提出一种稀疏流形学习算法（Manifold Learning Based on Sparse Bayesian Approach,MLSBA）.该算法是对稀疏贝叶斯模型的扩展,通过在模型的权值上定义稀疏流形先验,有效利用了样本数据的流形信息,提高了算法的分类准确率.在多种数据集上进行实验,结果表明：MLSBA不仅在流形数据集上取得良好的分类性能,而且在非流形数据集上效果也比较好;同时算法在两类数据集上均具有良好的稀疏性能.     

引入负相似的高光谱图像半监督分类

高光谱图像数据体现为波段多、地物标签获取困难大、谱信息抗干扰能力弱等特征,容易引起维数灾难、光谱空间变异性等问题,从而影响分类器的分类精度。针对这些问题,本文将负相似信息引入到拉普拉斯支持向量机（Laplacian Support Vector Machine, LapSVM）的流形正则化项中,提出了一种引入负相似的拉普拉斯支持向量机（Dissimilarity in Laplacian Support Vector Machine, Diss-LapSVM）分类算法,抑制光谱空间变异对分类结果的影响。同时,本文提出利用线性近邻传播（Linear Neighborhood Propagation, LNP）算法构造图的拉普拉斯矩阵,更有效地引入无标签样本的信息。实验结果表明,本文算法的分类精度得到了提高,特别是对光谱特征相似的地物。      

基于无监督特征选择的高光谱图像分类方法

在高光谱图像分类中,丰富的数据提升了其地物 识别能力。然而,由于样本特 征数大且有标记训练样本点少,导致“维度灾难”问题。本文提出一种基于无监督特征选择 的高光谱图像分类方 法,该方法同时考虑数据的流形嵌入映射和稀疏表达,将特征选择问题转化为一个优 化问题,数据的流形嵌入和稀疏表达作为约束项加入目标函数。设计了三个目标函 数,第一个目标函数描述流形学习的局部性原则,第二个目标函数将原始样本点回归 到低维嵌入空间,第三个目标函数对回归系数进行正则化。针对目标函数非凸的问 题,用迭代的方法来解这个约束优化问题,给出了解该优化问题的算法。优选特征用 于参与后续的分类识别任务。在真实的高光谱数据集上的实验表明,新方法能够提高 分类的精度。     

一种低秩和图正则化的协同稀疏高光谱解混方法

针对经典协同稀疏解混方法中稀疏性表征不足以及丰度矩阵过平滑等问题,提出一种低秩和图正则化的协同稀疏高光谱解混方法。引入加权因子,进一步促进丰度矩阵的稀疏性;引入了图正则化项,获取图像的空间信息,以促进图像的平滑性;在模型中增加低秩项,进而挖掘高光谱数据的细节结构,进一步提高解混的精度。利用2个模拟和1个真实高光谱数据进行实验,结果表明,提出方法的解混精度与经典解混方法相比得到显著提升。     

变形L1正则化的高光谱图像稀疏解混

如何准确地刻画易于求解的稀疏正则化函数是高光谱图像稀疏解混的难点。变形L₁正则化函数是一类由绝对值函数组成的双线性变换的单参数族,类似于L_pp∈0,1范数,通过调整参数a∈0,可以准确表征L₀和L₁之间的任意范数,并具有无偏、稀疏和Lipschitz连续性。论文首先研究变形L₁正则化函数,然后提出变形L₁正则化的高光谱稀疏解混变分模型,最后提出变形L₁正则化高光谱稀疏解混模型的凸函数差分求解算法。通过模拟和真实的高光谱数据实验,与经典的SUnSAL算法相比,表明提出的算法能够更准确地刻画丰度系数的稀疏性,并获得更高的解混精度。     

一种基于稀疏编码的多核学习图像分类方法

 本文提出一种基于稀疏编码的多核学习图像分类方法.传统稀疏编码方法对图像进行分类时,损失了空间信息,本文采用对图像进行空间金字塔多划分方式为特征加入空间信息限制.在利用非线性SVM方法进行图像分类时,空间金字塔的各层分别形成一个核矩阵,本文使用多核学习方法求解各个核矩阵的权重,通过核矩阵的线性组合来获取能够对整个分类集区分能力最强的核矩阵.实验结果表明了本文所提出图像分类方法的有效性和鲁棒性.对Scene Categories场景数据集可以达到83.10%的分类准确率,这是当前该数据集上能达到的最高分类准确率.     

Hyperspectral image classification using an extended Auto-Encoder method

This article proposes a spectral–spatial method for classification of hyperspectral images (HSIs) by modifying traditional Auto-Encoder based on Majorization Minimization (MM) technique. The proposed method consists of suggesting three main modifications. First, to construct weights of Auto-Encoder, similarity angle map(SAM) criterion is used as regularization term. It is useful to extract spectral similarity of initial features. Second, to enhance the classification accuracy, fuzzy mode is used to estimate parameters. These modifications lead to create an extended Auto-Encoder based on MM (EAEMM). Third, to improve the performance of Auto-Encoder, multi-scale features (MSF) are extracted. In comparison with some of the state-of-the-art methods, the experimental results obtained using the proposed method (MSF-EAEMM) show that it significantly improves the classification accuracy of HSI classification.     

Kernel eigenmaps based multiscale sparse model for hyperspectral image classification

Hyperspectral imaging (HSI) is the emerging method that combines traditional imaging and spectroscopy to provide the image with both the spatial and spectral information of the object present in the image. The major challenges of the existing techniques for HSI classification are the high dimensionality of data and its complexity in classification. This paper devises a new technique to classify the HSI named Spatial–Spectral Schroedinger Eigen Maps based Multi-scale adaptive sparse representation (S²SEMASR). In this, two different phases are employed for the accurate classification of the HSI, namely, Schroedinger Eigen maps (SE) based spatial–spectral feature extraction and multi-scale adaptive sparse classification for the feature extracted image. SE makes use of spatial–spectral cluster potentials which allows the extraction of features that best describes the characteristics of different classes of HSI. The multiscale adaptive sparse representation (MASR) applied over the SE features provides the sparse coefficients that includes distinct scale level sparsity with same class level sparsity. With the obtained coefficients, the class label of each pixel is determined. The proposed HSI classifier well utilizes the spectral and spatial characteristics to exploit the within-class variability and thus reduces the misclassification of similar test pixels Experimental results demonstrated that the proposed S²SEMASR approach outperforms the traditional results both qualitatively and quantitatively with an overall accuracy of 98.3%.     

基于多融合多尺度特征的高光谱图像分类研究

针对高光谱图像(HSI)波段之间的冗余性给高光谱图像分类结果产生的不利影响,研究基于多融合多尺度特征的高光谱图像分类方法。将采用于主成分分析降维处理的HSI数据作为多尺度特征多融合残差网络输入,利用多尺度特征多融合残差块提取HSI中的光谱特征和空间特征,并组成若干组光谱-空间特征;采用支持向量机展开分类处理,获取各光谱-空间特征的概率输出结果和权重,建立多特征加权概率融合模型,利用最大后验概率获取高光谱图像分类结果。实验结果表明:光谱-空间多尺度特征融合残差块数量为2+2模式、空间输入尺寸大小为9×9,可获取最佳多尺度特征融合残差网络;所提方法抗噪能力较好,可较好体现地物细节信息;且具备较高的高光谱图像分类精度。     

Band selection using variational mode decomposition applied in sparsity-based hyperspectral unmixing algorithms

In this work, a frequency-based dimensionality reduction technique using variational mode decomposition (VMD) is proposed. Dimensionality reduction is a very important aspect of preprocessing in case of hyperspectral image (HSI) analysis where this step helps in elimination of the lesser informative bands, thereby reducing the size of the data and making its processing computationally less challenging. In contrast to the standard dimensionality reduction methods such as inter-band block correlation (IBBC) where bands are eliminated based on their similarity with the consecutive bands, the proposed method uses frequency information of each band to categorize it as a less or more informative band. In this way, only the topmost informative bands of HSI are selected to form the reduced dataset. In our experiment, in order to verify the efficiency of VMD as a dimensionality reduction technique, the hyperspectral unmixed results obtained for IBBC reduced dataset is compared with those obtained for VMD reduced dataset. From the parametric measures such as classification accuracy, root-mean-square error (RMSE) and visual results obtained after unmixing for both IBBC and VMD reduced datasets, it is noticed that the VMD reduced dataset performs better by achieving higher classification accuracy and lower RMSE than that of the existing IBBC method.     

用于阿尔茨海默症分类的模糊逻辑特征选择和异质集成学习方法

阿尔茨海默症(AD)分类有助于在AD早期阶段及时采取针对性的治疗和干预措施,对降低老年群体的AD发病率和延缓AD疾病进展具有重要意义。该文提出一种改进的高斯模糊逻辑特征选择方法,首先采用互信息量和方差齐性分析两种方法给出特征重要性评分并分别进行归一化,然后使用改进的高斯模糊逻辑方法对其加权得到最终的特征重要性评分,最后依据特征重要性评分选取特征。该文还使用逻辑回归、随机森林、LightGBM、支持向量机和深度前馈网络作为初级分类器,多项式朴素贝叶斯分类器作为次级分类器,构建异质集成分类器,利用选取的特征进行AD分类。在TADPOLE数据集上进行实验,实验结果证实了所提特征选择方法是有效的,且采用所提特征选择方法,基于多项式朴素贝叶斯的异质集成分类器在AD分类上的性能要优于传统分类器。     

Gaussian process regression approach for robust design and yield enhancement of self-assembled nanostructures

Self-assembled nanostructures are increasingly used for nanoelectronic and optoelectronic applications due to their high surface area to volume ratio and their ability to break traditional lithography limits. However, they suffer due to poor yield and repeatability as the growth process is often not well studied or optimized. Gaussian process regression (GPR) is a machine learning technique that can be used for both regression and classification purpose. In the GPR framework, a probability measure is defined according to one prior belief about the response surface and the Bayesian rule is applied to combine the observations with prior beliefs to form a posterior distribution of the response surface, which is known as the “surrogate model”. We propose here the use of GPR as an effective statistical tool to optimize the growth conditions of nanostructures so as to improve their yield, controllability and repeatability ensuring at the same time that the yield is not affected by process variations at the identified optimum process conditions. In effect, we are proposing a design for reliability and robust design strategy for optimization of self-assembled nanostructure growth. We present here a case study of cadmium selenide nanostructures making use of an extensive design of experiment result (available open source) to illustrate the proposed methodology. The prediction accuracy of GPR is compared with two other commonly used statistical models → binomial and multinomial logistic regression. The use of the GPR method resulted in much better accuracy of probabilistic prediction of the different nanostructures with fewer fitting parameters than the logistic regression method.     

空谱结构保持的高光谱图像分类

高光谱遥感图像具有特征（波段）数多、冗余度高等特点,因此特征选择成为高光谱分类的研究热点。针对此问题,提出了空间结构与光谱结构同时保持的高光谱数据分类算法。考虑高光谱图像的物理特性,首先对图像进行加权空谱重构,使图像的空间结构信息自动融入光谱特征,形成空谱特征集;对利用最小二乘回归模型保存数据集的全局相似性结构的基础上,加入局部流形结构正则项,使挑选的特征子集更好地保存数据集的内在本质结构;讨论了窗口大小和正则参数对分类精度的影响。对Indian Pines、PaviaU和Salinas数据集的实验表明,该算法得到的特征子集的总体分类精度达到93.22%、96.01%和95.90%。该算法不仅充分利用了高光谱图像的空间结构信息,而且深入挖掘了数据集的内在本质结构,从而得到更有鉴别性的特征子集,相比传统方法明显提高了分类精度。