基于块对角化表示的多视角字典对学习  

Multi-view Dictionary-pair Learning Based on Block-diagonal Representation

作　　者：张帆 贺文琪[1,3] 姬红兵 李丹萍[4] 王磊 ZHANG Fan;HE Wen-qi;JI Hong-bing;LI Dan-ping;WANG Lei(Xidian University Qingdao Institute of Computing Technology,Qingdao,Shandong 266000,China;Key Laboratory of Marine Intelligent Equipment and System of Ministry of Education,Shanghai Jiao Tong University,Shanghai 200240,China;School of Electronic Engineering,Xidian University,Xi’an 710071,China;School of Telecommunications Engineering,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学青岛计算技术研究院,山东青岛266000 [2]上海交通大学海洋智能装备与系统教育部重点实验室,上海200240 [3]西安电子科技大学电子工程学院,西安710071 [4]西安电子科技大学通信工程学院,西安710071

出　　处：《计算机科学》2021年第1期233-240,共8页Computer Science

基　　金：国家重点研发计划资助项目(2016YFE0207000);国家自然科学基金资助项目(61203137,61401328);陕西省自然科学基础研究计划资助项目(2014JQ8306,2015JM6279)。

摘　　要：字典学习作为一种高效的特征学习技术被广泛应用于多视角分类中。现有的多视角字典学习方法大多只利用多视角数据的部分信息,且只学习一种类型的字典。实际上,多视角数据的相关性信息和多样性信息同样重要,且仅考虑一种合成型字典或解析型字典的学习算法不能同时满足处理速度、可解释性以及应用范围的要求。针对上述问题,提出了一种基于块对角化表示的多视角字典对学习方法(Block-Diagonal Representation based Multi-View Dictionary-Pair Learning,BDR-MVDPL),该方法通过引入字典对学习模型获得包含更多对分类有用的信息的表示系数,并通过显式约束使其具有块对角化结构,保证了编码系数矩阵的判别性;然后采用特征融合的方式将所有视角的编码系数进行串联,并将串联后的编码系数回归到对应的标签向量上,使多视角数据的多样性信息和数据相关性能够同时被利用;最后,该算法将字典学习与分类器学习整合到一个框架中,采用迭代求解的方式,交替更新字典对和分类器,使所提方法能够自动完成分类。3个多特征数据集上的实验结果表明,与主流的多视角字典学习算法相比,所提算法在保持低复杂度的同时具有更高的分类准确率。Dictionary learning is widely used in multi-view classification as an efficient feature learning technology.Most multi-view dictionary learning methods either only use part of the information of multi-view data,or only learn one type of dictionary in their frameworks.However,in practice,both the diversity information of multi-view data and the correlation of multi-view data are equally important.A single synthetic dictionary learning scheme or a single analytic dictionary learning scheme cannot meet the requirements of the processing speed,interpretability and application feasibility at the same time.To solve these issues,a novel block-diagonal representation based multi-view dictionary-pair learning framework(BDR-MVDPL)is proposed in this paper.This algorithm obtains the representation coefficients that contain more useful information for classification by introducing dictionary-pair learning model.Firstly,in order to ensure the discriminant ability of the coding coefficient matrix,the proposed me-thod directly enforces a block-diagonal constraint on the coding coefficients with explicit formulation.Then,it adopts a feature fusion strategy to concatenate the coding coefficients of different views,and regresses the concatenated coding coefficients to the corres ponding label vectors.In this way,both the diversity information of multi-view data and the correlation of multi-view data are considered.Finally,it integrates dictionary learning and classifier learning into a unified framework,so that the dictionary-pair and classifier can update alternately in an iterative manner and the whole classification task can be realized automatically.Experiments on several multi-feature datasets show that,compared to other multi-view dictionary learning algorithms,the proposed method achieves competitive performance in terms of classification accuracy,while enjoying a low computational complexity.

关 键 词：字典学习 字典对学习 多视角学习 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]