面向图像识别的测地局部典型相关分析方法  

A Geodesic Locality Canonical Correlation Analysis Method for Image Recognition

作　　者：许欢 苏树智 颜文婧 邓瀛灏 谢军 XU Huan;SU Shuzhi;YAN Wenjing;DENG Yinghao;XIE Jun(College of Computer Science and Engineering,Anhui University of Science&Technology,Huainan 232001,China;School of Computer and Information Engineering,Beijing Technology and Business University,Beijing 100037,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,淮南232001 [2]北京工商大学计算机与信息工程学院,北京100037

出　　处：《电子与信息学报》2020年第11期2813-2818,共6页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61806006);安徽省高等学校自然科学研究基金(KJ2018A0083);中国博士后科学基金(2019M660149)。

摘　　要：典型相关分析(CCA)是一种经典的多模态特征学习方法,能够从不同模态同时学习相关性最大的低维特征,然而难以发现隐藏在样本空间中的非线性流形结构。该文提出一种基于测地流形的多模态特征学习方法,即测地局部典型相关分析(GeoLCCA)。该方法利用测地距离构建了低维相关特征的测地散布,并进一步通过最大化模态间的相关性和最小化模态内的测地散布学习更具鉴别力的非线性相关特征。该文不仅在理论上对提出的方法进行了分析,而且在真实的图像数据集上验证了方法的有效性。Canonical Correlation Analysis(CCA)is a classic multi-modal feature learning method,which can learn low-dimensional features with the maximum correlation from different modalities.However,it is difficult for CCA to find the nonlinear manifold structures hidden in the sample spaces.This paper proposes a multimodal feature learning method based on geodesic manifolds,namely Geodesic Locality Canonical Correlation Analysis(GeoLCCA).The geodesic distances are used to construct the geodesic scatters of low-dimensional correlation features,and the nonlinear correlation features with better discriminative power are learned by maximizing the between-modal correlation and minimizing the within-modal geodesic scatters.This paper not only analyzes the proposed method in theory,but also verifies the effective of the proposed method on the realworld image datasets.

关 键 词：图像识别 典型相关分析 多模态特征学习 流形学习 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]