Abstract: Multi-class image classification is a fundamental problem in computer vision research area.The existing approaches solving this problem mainly focus on how to construct a one-vs-rest multi-class classifier.The important intrinsic connections among different classes are completely ignored by such a strategy,and consequently the image features cannot be utilized sufficiently.To solve this issue,this paper proposes a classifiers construction method based on truncated nuclear norm.In principle and practice,such truncated nuclear norm is able to capture the intrinsic connections among different classes,and meanwhile overcome the drawback of traditional nuclear norm for matrix rank approximation.Experimental results show that the proposed method can remarkably improve the image classification performance on the benchmark datasets.

Key words: image classification, truncated kernel function, convex optimization, class correlation, matrix rank, Support Vector Machine(SVM)

摘要： 多类别图像分类是计算机视觉领域的一个基本问题,现有分类方法大多是根据一对多的原则构建一个多类别分类器,在构建分类器时忽视了类与类之间的本质关联,难以较好地利用样本特征。为此,提出一种基于截断核函数的分类器构建方法。利用截断核函数捕捉图像类别之间的关联,同时避免传统核函数在逼近矩阵秩时的偏差问题,并针对建立的截断核函数优化模型,设计一种有效的交叉迭代算法。实验结果表明,该截断核函数方法能够提高图像分类的精确度。

关键词: 图像分类, 截断核函数, 凸优化, 类关联, 矩阵秩, 支持向量机

CLC Number: 

• TP242.6+2