张量主成分分析及在图像序列识别中的应用

发布时间：2018-08-23 12:55

【摘要】：图像序列,如视频图像、医学图像、高光谱遥感影像等都属于三维张量。张量本质上是多维数组,它是矩阵的多线性推广。图像序列不仅成为人类活动中最常用的信息载体,而且在张量模式下对图像序列的识别也成为近几年来模式识别领域研究的热点问题。在特征提取中,最常用的方法是主成分分析法(Principal Component Analysis,PCA)。由于图像序列属于三维张量,本文考虑使用张量主成分分析(Tensor Principal Component Analysis,TPCA)的方法对图像序列进行识别。然而现有的张量主成分分析的方法在特征提取时无法在张量模式下确定一个合适的奇异值阈值,从而无法确定图像序列的特征保存率。在分类识别中,现有的张量型分类器仅能处理二维数据,无法对多维特征进行直接分类。这种所谓的张量型分类器目前只能局限于二维,本质上是二维张量型分类器。针对以上问题,本文在张量模式下对图像序列的识别进行研究,具体内容如下:首先,现有的TPCA方法存在缺陷:无法在张量模式下确定一个合适的奇异值阈值,即在张量模式下无法找出去除噪声和保留细节之间的平衡点。本文在TPCA的基础上提出了截断张量主成分分析(Truncated Tensor Principal Component Analysis,TTPCA)的方法来确定一个合适的奇异值阈值。奇异值阈值的确定用来滤除较小的奇异值、保留较大的奇异值,从而找到去除噪声和保留细节的平衡点,完成对图像序列的特征提取。然后,为了有效地提高图像序列的识别精度,考虑到经过特征提取后的图像序列依然是张量模式,提出张量模式下的三维支持张量机(Three Dimensional Support Tensor Machine,3DSTM)对张量型数据进行直接分类,避免了将张量数据矢量化。关于张量模式的分类器主要做了如下工作:第一,在3DSTM算法中,利用张量乘法的运算规则,改进传统的支持向量机和二维的支持张量机,把它们扩展到理论上的N维,使其可以直接处理张量模式的输入;第二,在3DSTM分类器模型的基础上,根据SVM和3DSTM两者的优缺点,引入多秩的思想,提出基于多秩三维支持张量机(Multiple Rank Three Dimensional Support Tensor Machine,MR3DSTM)分类器的设计,使图像序列的识别率更高。用本文提出的TTPCA和3DSTM,以及TTPCA和MR3DSTM这两种张量形式的分类识别算法与目前流行的两种算法进行实验对比。实验结果表明,本文提出的算法在图像序列的识别精度和速度上有明显提高;并且本文提出的MR3DSTM和3DSTM这两种分类器方法相比,MR3DSTM的识别精度更高。
[Abstract]:Image sequences, such as video images, medical images and hyperspectral remote sensing images, all belong to 3D Zhang Liang. Zhang Liang is essentially a multidimensional array, which is a multilinear generalization of a matrix. Image sequences have not only become the most commonly used information carrier in human activities, but also the recognition of image sequences in Zhang Liang mode has become a hot topic in the field of pattern recognition in recent years. In feature extraction, principal component analysis (Principal Component) is the most commonly used method. Because the image sequence belongs to three dimensional Zhang Liang, the method of Zhang Liang principal component analysis (Zhang Liang) is considered to recognize the image sequence. However, the existing Zhang Liang principal component analysis method can not determine a suitable singular value threshold in Zhang Liang mode when feature extraction, thus can not determine the feature preservation rate of image sequence. In classification and recognition, the existing Zhang Liang classifier can only deal with two-dimensional data, and can not directly classify multidimensional features. The so-called Zhang Liang classifier can only be confined to two-dimensional, essentially two-dimensional Zhang Liang-type classifier. Aiming at the above problems, this paper studies the recognition of image sequences in Zhang Liang mode. The main contents are as follows: firstly, the existing TPCA methods have some defects: it is impossible to determine a suitable singular value threshold in Zhang Liang mode. That is, the balance between noise removal and detail retention can not be found in Zhang Liang mode. In this paper, we propose a truncated Zhang Liang principal component analysis (Zhang Liang) method based on TPCA to determine an appropriate singular value threshold. The determination of the singular value threshold is used to filter the smaller singular value and retain the larger singular value, so as to find the equilibrium point to remove noise and preserve the details, and to complete the feature extraction of the image sequence. Then, in order to improve the recognition accuracy of image sequence effectively, considering that the image sequence after feature extraction is still Zhang Liang pattern, a 3D support Zhang Liang machine (Three Dimensional Support Tensor Machine 3DSTM) based on Zhang Liang mode is proposed to classify Zhang Liang data directly. Avoid vectorization of Zhang Liang data. The main work of the classifier about Zhang Liang pattern is as follows: first, in the 3DSTM algorithm, the traditional support vector machine and the two-dimensional support Zhang Liang machine are improved by using the operation rules of Zhang Liang multiplication, and they are extended to the N-dimension in theory. It can directly handle the input of Zhang Liang mode. Secondly, on the basis of 3DSTM classifier model, according to the advantages and disadvantages of SVM and 3DSTM, the idea of multi-rank is introduced. A new classifier based on multi-rank Zhang Liang machine (Multiple Rank Three Dimensional Support Tensor Machine MR3DSTM is proposed, which makes the recognition rate of image sequence higher. The classification and recognition algorithms of TTPCA and 3DSTM, as well as TTPCA and MR3DSTM, which are proposed in this paper, are compared with the two popular algorithms. The experimental results show that the proposed algorithm improves the recognition accuracy and speed of image sequences obviously, and the two classifiers, MR3DSTM and 3DSTM, have higher recognition accuracy than MR3DSTM.
【学位授予单位】：东北电力大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】