【作者】 石梅；

【导师】 郭军；

【作者基本信息】 西北大学 ， 计算机应用技术， 2021， 硕士

【摘要】 多视图分析能够较好地利用不同视图之间信息的兼容性和互补性,比单视图分析具有更好的决策性能,目前广泛应用于计算机视觉、目标检测、行为识别等领域。然而传统的多视图分析大多只研究成对视图间的关系,忽略了所有视图间关系,导致得到的公共空间判别性不强。尽管一些非成对多视图方法得到了判别性公共空间,却使用广义特征值分解方法对其近似求解,使得最终结果偏离全局最优解。因此,人们希望既能得到判别性公共空间,还能得到全局最优解的多视图分析模型。本文重点研究了两种多视图学习方法:TRCMv DA模型解决多视图判别分析中的迹比(Trace Ratio,以下简称为TR)问题;HMMv DA解决一些多视图分析对所有的类间距离赋予相同权重,较小类间距离在公共空间上重叠的问题。主要工作如下:(1)提出基于迹比准则的多视图公共空间判别分析(TRCMv DA),该有监督模型同时研究了视图内和视图间的关系,得到判别性公共空间。针对多视图判别分析中的TR问题,我们提出一个可以放宽半正定矩阵约束条件的策略,使模型可以通过Newton-Raphson迭代方法求解。最后,该模型与7个多视图主流算法在5个常用公开数据集上进行对比实验,结果说明该模型解决了TR问题,且识别精度提升3%以上。(2)提出基于加权调和均值的多视图判别分析(HMMv DA),解决一些多视图算法对所有的类间距离赋予相同权重,造成较小类间距离在公共空间上发生重叠的问题。该有监督模型优先考虑成对类间距离,使所有类在公共空间上尽可能的区分开,得到更具判别性的公共空间。并且利用基于调和均值的类间散射矩阵增加了模型的鲁棒性,引入正则化参数平衡类内距离和类间距离对目标函数的影响,最后通过拉格朗日乘子法和KKT条件对模型进行迭代求解。实验证明,HMMv DA模型的识别精度略高于TRCMv DA,说明了本模型相比于后者可得到更具判别性的公共空间。本文提出的两个多视图监督学习模型,它们同时研究了视图内与视图间的关系,能够得到判别性较强的公共特征空间,并能求得全局最优解。最后,我们还成功地将两个模型应用于人脸识别和多模态人体动作识别并取得了良好的效果,证明模型具有较好的普适性。更多还原

【Abstract】 Nowdays multi-view analysis is widely used in the fields of computer vision,target detection,human behavior recognition and so on.Multi-view learning method utilizes the compatibility and complementarity of information between different views,which has the higher decision-making performance than single view analysis.However,many conventional multi-view methods just care about the relationship between the pairwise views,but ignore the relationship among all views.This leads to the common space obtained is not discriminative enough.Although some non-pairwise multi-view algorithms gain the discriminative feature space,the global optimal solution may not gain due to the common used generalized eigenvalue decomposition(GEVD)method.Therefore,the research community pays a lot of attentions on developing more efficient multi-view analysis models,which not only can get the discriminative common space,but also have the global optimal solution.This dissertation focuses on 2 multi-view discriminant analysis algorithms:Multi-view Discriminant Analysis for Trace Ratio problem(TRCMv DA)and Harmonic Mean Weighted based Multi-view Discriminant Analysis(HMMv DA).The main contributions are summarized as follows:(1)We propose Trace Ratio Criterion for Multi-view Common Space Discriminant Analysis(TRCMVDA).This supervised method aims to project the samples from multiple views into a common space by considering all views.Because of the TR problem in Multi-view Discriminant Analysis(Mv DA),we present a practical strategy to relieve the restraint condition of positive semi-definite matrix,thus we can obtain the closed form solution by using Newton-Raphson method directly.Compared with the experimental results of 7 mainstream multi-view algorithms on 5 public datasets,the accuracy is improves by more than 3%,and the TR problem can be solved.(2)Some methods give the equal consideration to all between-classes distances,and the close pairwise between-classes tend to overlap in the common space,it will reduce the performance.To avoid this problem,we present Weighted Harmonic Mean Based Multi-view Discriminant Analysis(HMMv DA).This supervised model gives priority to the pairwise between-class distance,then we can get a more discriminative common space,so the different class can be distinguished as much as possible.Moreover,we employ the harmonic mean between-class scatter matrix to increase the robustness of our proposed.A regularization term is introduced to balance the influence of between-class scatter and within-class scatter on the objective function.In the end,we use the Lagrange multiplier method and KKT condition to solve the model iteratively.The experimental results illustrate the accuracy of HMMv DA is slightly higher than TRCMv DA,which demonstrates this model can get more discriminative common space than latter.We present 2 multi-view learning methods,they consider the relationship inter-view and intra-view to gain the discriminant common space,then we can obtain the global optimal solution.This paper also successfully apply our two models to the fields of face recognition and multi-modal human action recognition.The results of our 2 methods prove the universal.更多还原