Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2021, Vol. 47 ›› Issue (2): 285-292. doi: 10.19678/j.issn.1000-3428.0056556

• Graphics and Image Processing • Previous Articles     Next Articles

Multi-Scale Multi-Kernel Gaussian Process Latent Variable Model

ZHOU Peichun1, WU Lan'an2   

  1. 1. School of Computer Science and Engineering, Yulin Normal College, Yulin, Guangxi 537000, China;
    2. School of Computer and Information Engineering, Nanning Normal University, Nanning 530299, China
  • Received:2019-11-10 Revised:2019-12-30 Online:2021-02-15 Published:2020-01-21

多尺度多核高斯过程隐变量模型

周培春1, 吴兰岸2   

  1. 1. 玉林师范学院 计算机科学与工程学院, 广西 玉林 537000;
    2. 南宁师范大学 计算机与信息工程学院, 南宁 530299
  • 作者简介:周培春(1970-),男,讲师、硕士,主研方向为数据挖掘、计算数学及其应用技术;吴兰岸(通信作者),副教授、博士。
  • 基金资助:
    国家自然科学基金(61763010)。

Abstract: As an unsupervised Bayesian non-parameter dimension reduction model,the Gaussian Process Latent Variable Model(GPLVM) fails to efficiently utilize semantic label information of data.Moreover,it just assumes that the features of all observed variables are independent in modeling,and thus ignores the spatial information among the features.To address the two issues,this paper proposes a Multi-Scale Multi-Kernel Gaussian Process Latent Variable Model(MSMK-GPLVM).The model projects the images of different scales into a low-dimensional latent space through linear projection for feature fusion.A MSMK-GPLVM is constructed by taking the fused features as the input and the data labels as the output,and it realizes supervised learning through the connection between image data and data labels,and jointly learns the GPLVM and linear projection weight matrix to improve the classification performance.Experimental results show that MSMK-GPLVM can effectively utilize the spatial structure information of images and the semantic label information.Compared with other latent variable models,it has better performance in dimension reduction and classification.

Key words: Gaussian process, latent variable, dimension reduction, semantic information, spatial information

摘要: 高斯过程隐变量模型(GPLVM)作为一种无监督的贝叶斯非参数降维模型,无法有效利用数据所包含的语义标记信息,同时其建模过程中假设观测变量的各特征相互独立,忽略了特征之间的空间结构信息。为解决上述问题,采用图像池化操作获得不同尺度的特征表示,利用线性投影方式将不同尺度的图像投影到低维隐空间进行特征融合,并将融合特征和数据标记分别作为输入和输出,构建多尺度多核高斯过程隐变量模型(MSMK-GPLVM),通过图像数据与数据标记的关联实现模型监督学习,同时对GPLVM和线性投影权重矩阵进行联合学习以提高分类性能。实验结果表明,MSMK-GPLVM能够有效利用图像空间结构信息和语义标记信息,相比其他隐变量模型具有更强的数据降维和分类能力。

关键词: 高斯过程, 隐变量, 降维, 语义信息, 空间信息

CLC Number: