基于鲁棒回归度量学习的图像分类算法  

Robust regression metric learning algorithmfor image classification

作　　者：常冬霞[1] 王舒伟 CHANG Dongxia;WANG Shuwei(School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044

出　　处：《北京交通大学学报》2021年第2期119-126,共8页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：中央高校基本科研业务费专项资金(2019YJS051);国家自然科学基金(61532005)。

摘　　要：度量学习是机器学习中的重要研究问题之一,针对实际应用中的噪声数据,如何建立一个鲁棒的度量仍是一个挑战.本文将稀疏表示、特征学习与分类模型相结合提出了一种新的基于鲁棒回归度量学习(RRML)的算法并将其应用于图像分类.算法对最优特征子空间和稀疏表示进行联合学习,在更具判别性的低维表征空间中,通过稀疏表示有效地编码数据的局部结构信息,进而更好地揭示数据的内在鉴别信息,并以此指导该模型学习到最优的投影矩阵;同时对噪声矩阵和投影矩阵的行稀疏约束,可以极大降低噪声的影响.实验结果表明所提算法在图像分类准确率和鲁棒性方面均优于其他对比算法.Metric learning is one of the important research problems in machine learning.It is still a challenge to build a robust metric,especially for corrupted data in the real-world application.In this paper,a new Robust Regression Metric Learning(RRML)algorithm is proposed and applied to image classification,in which the sparse representation,feature learning and classification model are integrated into a framework.The optimal feature subspace and sparse representation are jointly learned.The sparse representation matrix can effectively encode the local structure information of the data to further reveal the internal discriminative information of the data in a more discriminative low-dimensional representation space,which could guide the model to learn the optimal projection matrix.In order to insignificantly reduce the noise influence,the row sparsity constraint is imposed on the noise matrix and projection matrix.Results show that the proposed algorithm is better than other comparison algorithms in terms of image classification accuracy and robustness.

关 键 词：图像分类 度量学习 鲁棒回归 稀疏表示 特征子空间 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]