结合显性与隐性空间光滑的高效二维图像判别特征抽取

发布时间：2019-05-17 10:26

【摘要】：图像具有固有的二维空间结构,空间上邻近的像素点通常具有相近的灰度值,意味着图像具有局部光滑性.为对其特征抽取,传统方法常将原始图像拉成向量,造成空间结构的破坏,由此直接基于图像的2D特征抽取法应运而生.典型的如2DLDA,2DPCA,相比向量方法,计算复杂度显著降低,但其操作针对的是图像整行(或整列),导致空间光滑度过粗.为此,空间正则化通过在向量化空间中显式地施加局部空间光滑弥补这一不足,由此获得了比2D抽取法更优的分类性能,但其遗传了向量法的高计算代价.最近,隐性空间正则化方法(implicit spatial regularization,ISR)提出利用图像划分与重组隐性地体现图像局部光滑性,而后再利用现有2D方法抽取特征,使典型双边2DLDA性能优于SSSL(一种典型的显性空间正则化方法),但是,仅隐性地光滑缺乏显式的强制约束力,其特征空间依然欠光滑,同时双边2DLDA由非凸问题获得,计算耗时却不能保证解的全局最优性.鉴于此,提出一种结合显性与隐性空间光滑的高效二维图像判别特征抽取框架(2D-CISSE).其关键步骤是预先对图像显性地全局光滑,紧接着进行ISR,既继承了ISR的隐性光滑又强化了图像局部光滑的显式约束力,不仅可直接获得全局最优投影,同时该框架具有一般性,即现有大部分图像光滑方法与2D特征抽取法均可嵌入其中.最后,通过在人脸数据集Yale,ORL,CMU PIE,AR以及手写数字数据集MNIST和USPS上的对比实验验证了2D-CISSE框架性能的优越性与计算的高效性.
[Abstract]:The image has an inherent two-dimensional spatial structure, and the adjacent pixel points in space usually have similar gray values, which means that the image is locally smooth. In order to extract the features, the traditional methods often pull the original image into a vector, resulting in the destruction of the spatial structure, so the 2D feature extraction method based on the image emerges as the times require. Typical such as 2DLDA and 2DPCA, compared with vector method, the computational complexity is significantly reduced, but the operation is aimed at the whole row (or column) of the image, which leads to the smoothness of the space. For this reason, spatial regularization makes up for this deficiency by explicitly applying local space smoothness in vector space, and thus obtains better classification performance than 2D extraction method, but it inherits the high computational cost of vector method. Recently, the implicit spatial regularization method (implicit spatial regularization,ISR) proposed to use image partition and reconstruction to implicitly reflect the local smoothness of the image, and then use the existing 2D method to extract features. The performance of typical bilateral 2DLDA is better than that of SSSL (a typical explicit spatial regularization method). However, only implicit smoothness lacks explicit mandatory binding, and its feature space is still not smooth. At the same time, bilateral 2DLDA is obtained by non-convex problem. The calculation time is time consuming, but the global optimality of the solution can not be guaranteed. In view of this, an efficient two-dimensional image discriminant feature extraction framework (2D-CISSE) combining dominant and implicit spatial smoothness is proposed. The key step is to predominate the global smoothness of the image, and then ISR, inherits the implicit smoothness of ISR and strengthens the explicit binding force of local smoothness of the image, which can not only obtain the global optimal projection directly. At the same time, the framework is general, that is, most of the existing image smoothing methods and 2D feature extraction methods can be embedded in it. Finally, the advantages of 2D-CISSE framework and the computational efficiency of 2D-CISSE framework are verified by comparative experiments on face dataset Yale,ORL,CMU PIE,AR and handwritten digital dataset MNIST and USPS.
【作者单位】：南京航空航天大学计算机科学与技术学院;
【基金】：国家自然科学基金项目(61472186) 高等学校博士学科点专项科研基金项目(20133218110032)~~
【分类号】：TP391.41

【相似文献】