多视图典型相关分析的理论研究和应用

发布时间：2018-07-29 13:00

【摘要】：多视图数据在实际应用中普遍存在,针对多视图数据的特征抽取也逐渐成为模式识别领域中的研究热点和关键问题。多视图典型相关分析作为一种非常重要的多视图特征抽取方法,在国内外已经受到广泛关注。然而实际采集的数据通常是复杂的非线性数据,多视图典型相关分析难以有效地处理这种实际数据,这会很大程度地限制它的适用范围和实际性能。因此,本文以多块嵌入、图增强方法、标签核技术、模糊投影、核排列以及超分辨率重构等为研究对象,深入探索了多视图典型相关分析理论,并提出了一系列多视图特征抽取方法,为一些实际应用提供了有效的解决方案。本文的主要创新性工作和研究成果如下:(1)提出了多块嵌入的多视图典型相关分析方法。在已有的局部相关分析方法中,存在单类局部近邻关系难以揭示原始高维数据内在流行的问题,以及监督引入导致整体局部信息缺失的问题。为此,本文以局部块为基础,深入研究了多视图相关分析框架下的多源块结构和多源融合理论,并提出了多块嵌入的多视图典型相关分析方法。该方法不仅能够自动地学习更有益于相关特征抽取的整体局部信息,而且能够借助视图内散布结构进一步增强相关特征的类分离性。在近红外图像、人脸图像和多视角汽车图像上的实验结果显示,该方法拥有良好的图像识别性能和参数鲁棒性。(2)提出了图增强多视图鉴别相关分析方法。为了解决多视图相关分析中图难以很好掌握数据本质几何结构的问题,本文提出了基于数据空间分割和监督概率整合模型的图增强方法,该方法构建的增强图能够更好地反映本质几何结构和鉴别分布信息。在增强图的基础上,进一步探索多视图相关分析框架下图的有效嵌入和监督信息的指导,并提出了图增强多视图鉴别相关分析方法。该方法构建了视图间类内和类间的增强相关性,并且考虑了能够更好掌握视图内和视图间散布结构的全局散布。大量实验结果不仅揭示了图增强方法的有效性,而且显示了图增强多视图鉴别相关分析方法在识别任务中的优越性。(3)提出了新的标签核以及多视图标签核相关分析算法。针对经验核方法的监督缺失问题,本文提出了一种标签核方法,该方法借助标签指导的单位超球模型很好地保留了类标签信息的鉴别力,然而该方法的标签依赖性导致其难以实现外样本扩展,为此,利用样本分布的相似准则进一步构建了标签核方法的投影辅助策略,即模糊投影策略。通过将标签核方法和模糊投影策略自然地融入多视图相关分析理论,进一步提出了多视图标签核相关分析算法,该算法借助标签核方法和模糊投影策略的优势能够从多视图数据中学习带有强鉴别力的非线性相关特征。另外,在五个不同的数据集上,分析了近邻参数对模糊投影性能的影响和多视图标签核相关分析算法的识别性能。(4)提出了核排列的多视图典型相关分析方法。针对单核的性能局限性和难以为具体数据选择适应核的困境,本文使用核排列将原始的特征向量转化为二维特征矩阵,并依此提出了核排列的多视图典型相关分析方法。该方法能够同时利用大量的核来揭示每个视图更真实的数据分布信息,并且能够为每个视图自动地学习一个具有数据适应性的混合核。此外,该方法的核扩展技术能够在其他特征抽取方法中直接使用,具有很好的迁移性。在近红外人脸图像、热红外人脸图像、可见光人脸图像、手写体图像和目标图像上的实验表明,提出的方法具有良好的图像识别性能。(5)提出了基于局部多视图一致子空间学习的超分辨率方法。在超分辨率重构中,为了更好地满足相似局部假设,本文提出了一种局部多视图一致子空间学习方法;为了能够同时使用多类低分辨率图像进行超分辨率重构,提出了多相关融合策略;为了确定更准确的近邻域,提出了重构校对策略。然后以此为基础,形成一种新的人脸超分辨率重构方法,即基于局部多视图一致子空间学习的超分辨率方法。该方法解决了现存基于相关分析理论的超分辨率重构方法无法处理的一些关键问题,在相关分析理论框架下首次实现了图像超分辨率的多源重构。此外,针对残差补偿的必要性、多相关融合策略的有效性、局部多视图一致子空间的优越性、参数的影响、训练图像数量的影响以及最终重构图像的质量等设计了大量实验,所有的实验结果可以给出一个合理的观察,即基于局部多视图一致子空间学习的超分辨率方法是一种有效的人脸超分辨率重构方法。
[Abstract]:Multi view data is widely used in practical applications, and feature extraction for multi view data has gradually become a hot and key problem in the field of pattern recognition. As a very important multi view feature extraction method, multi view canonical correlation analysis has received extensive attention at home and abroad. Often complex nonlinear data, multi view canonical correlation analysis is difficult to effectively deal with the actual data, which will greatly limit its scope of application and practical performance. Therefore, this paper is based on multi block embedding, graph enhancement method, label kernel technology, fuzzy projection, kernel row and super-resolution reconstruction, and so on. A series of multi view feature extraction methods are proposed to provide effective solutions for some practical applications. The main innovative work and research results of this paper are as follows: (1) a multi block multi view canonical correlation analysis method is proposed. In the existing local correlation analysis methods, there are some existing methods. The single class local neighborhood relationship is difficult to reveal the inherent problem of the original high dimensional data and the problem of the lack of local information. This paper, based on the local block, deeply studies the multi source block structure and multi source fusion theory under the framework of multi view correlation analysis, and puts forward a multi block multi view typical phase. The method not only can automatically learn the whole local information that is more beneficial to the correlation feature extraction, but also can further enhance the class separability of the related features with the aid of the scatter structure in the view. The results show that the method has good image recognition in the near infrared image, face image and multi view vehicle image. Performance and parameter robustness. (2) a graph enhancement multi view identification correlation analysis method is proposed. In order to solve the problem that the graph is difficult to master the essential geometric structure of the data in the multi view correlation analysis, this paper proposes a graph enhancement method based on the data space segmentation and the supervision probability integration model, and the enhanced graph constructed by this method can be better. To reflect the essential geometric structure and the differential distribution information. On the basis of the enhancement graph, we further explore the guidance of the effective embedding and monitoring information under the multi view correlation analysis framework, and propose a graph enhanced multi view identification correlation analysis method. This method constructs the enhanced correlation between classes and classes between views, and considers it possible A large number of experimental results not only reveal the effectiveness of the image enhancement method, but also show the superiority of the image enhancement multi view identification correlation analysis method in the recognition task. (3) a new label kernel and a multi view sign kernel correlation analysis algorithm are proposed. A label kernel method is proposed in this paper. This method preserves the discriminative ability of the class label information with the help of the unit super ball model guided by the label. However, the label dependence of the method makes it difficult to realize the expansion of the outer sample. Therefore, the label kernel method is further constructed by using the similarity criterion of the sample distribution. The projection assistant strategy, that is, the fuzzy projection strategy. By integrating the label kernel method and the fuzzy projection strategy naturally into the multi view correlation analysis theory, the algorithm of multi view sign kernel correlation analysis is further proposed. The algorithm can learn from the multi view data with the advantage of the advantage of the label kernel method and the fuzzy projection strategy. In addition, on five different data sets, the influence of near neighbor parameters on the fuzzy projection performance and the recognition performance of the multi view sign kernel correlation analysis algorithm are analyzed. (4) a multi view canonical correlation analysis method for kernel arrangement is proposed. In this paper, a kernel arrangement is used to transform the original eigenvector into a two-dimensional feature matrix, and a multi view canonical correlation analysis method is proposed. This method can simultaneously use a large number of kernel to reveal the more real data distribution information of each view, and can automatically learn a data adaptation for each view. In addition, the kernel expansion technique of this method can be used directly in other feature extraction methods and has good mobility. Experiments on near infrared face image, hot infrared face image, visible light face image, handwritten image and target image show that the proposed method has good image recognition performance. (5) A super-resolution method based on local multi view uniform subspace learning is proposed. In order to better satisfy similar local assumptions in super-resolution reconstruction, a local multi view uniform subspace learning method is proposed in this paper. In order to be able to use multiple class of low resolution images at the same time for super-resolution reconstruction, a multi correlation fusion is proposed. In order to determine a more accurate near neighbourhood, the reconstructing proofreading strategy is proposed. On the basis of this, a new method of face superresolution reconstruction is formed, which is a super-resolution method based on local multi view uniform subspace learning. This method solves the problem that the existing superresolution reconstruction method based on the correlation analysis theory can not be processed. For the first time, the multi source reconstruction of image superresolution is realized in the framework of correlation analysis theory. In addition, in view of the necessity of the residual compensation, the validity of the multi correlation fusion strategy, the superiority of the local multi view uniform subspace, the influence of the parameters, the influence of the number of the training images and the quality of the final reconstructed image In a large number of experiments, all the experimental results can give a reasonable observation, that is, the super-resolution method based on local multi view uniform subspace learning is an effective method of face super-resolution reconstruction.
【学位授予单位】：江南大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】