一种面向稀疏表示的最大间隔字典学习算法  

Max-margin dictionary learning algorithm for sparse representation

作　　者：段菲[1] 章毓晋 

机构地区：[1]清华大学电子工程系 [2]信息科学与技术国家重点实验室,北京100084

出　　处：《清华大学学报（自然科学版）》2012年第4期566-570,共5页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(61171118);高等学校博士学科点专项科研基金(SRFDP-20110002110057)

摘　　要：近年来,基于稀疏表示的分类技术(SRC)在图像分类和目标识别中取得了巨大的成功。在该框架中,过完备基的学习和多类分类器(通常为支持向量机SVM)的训练是最关键的两个步骤。但在目前的许多方法中,这两个模块的构建过程都是相互独立的。该文针对以上问题,提出了一种用于稀疏表示的最大间隔字典学习算法,将两类SVM分类器的损失函数项的平方及分类间隔作为正则项与稀疏字典的学习过程进行了整合,并提出相应的坐标轮换优化算法对目标函数进行优化,实现了字典和分类器的同步学习。所提出的框架能够增强多类分类器中两类分类器的推广性能,并减少多类分类器的误差界。为了对所提出算法的性能进行评价,在2个常用标准库上进行了分类实验。结果表明,所提出的算法的与SRC相比识别率提升均超过3%。Sparse representation based classification （SRC） has been successfully used for image classification and object recognition. The two critical steps in this framework are the learning of an overcomplete basis and the training of the multi-class classifier （usually support vector machine, SVM）. However, these two modules are usually independent in existing approaches. This paper presents a max-margin dictionary learning algorithm for sparse coding that incorporates these two aspects. The square of the hinge-loss term of the base binary SVM is introduced into the sparse dictionary learning with a numerical optimization algorithm based on a coordinate transformation to optimize the dictionary and the binary SVM parameters. This approach enhances the generalization ability of the binary SVM and, consequently, reduces the error bound of the aggregated multi-class SVM. Experimental results demonstrate that this algorithm has at least 3% improvements over SRC on two benchmark datasets.

关 键 词：最大间隔 稀疏表示 字典学习 支持向量机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]