摘要
近年来,机器学习在计算机视觉中取得了许多突破性的研究进展.然而,已训练好的学习模型难以直接应用于相似但具有不同数据分布特征的其它学习任务中.域自适应技术通过抽取源域与目标域数据之间的公共特征,来实现把源域中学习到的知识迁移至目标域,从而避免针对目标域的训练数据收集和模型训练代价.但是,现有的视觉域自适应方法大都无法处理高阶的特征数据,一般都是通过简单的向量化操作将高阶张量特征转换成高维一阶向量特征.这不仅会破坏高阶特征数据内部的结构信息,而且还会增加算法的计算复杂度.为了解决上述问题,本文在保持原有张量特征结构不变的条件下,利用张量乘操作,将视觉域自适应问题抽象为求解源域和目标域的共同张量子空间以及源域和目标域特征在该共同张量子空间上投影的多变量优化问题.然后,利用张量奇异值分解和交替方向乘子法,提出一种基于张量奇异值分解的视觉域自适应方法(Visual domain Adaptation method based on TEnsor Singular value decomposition,VATES),以实现上述多变量优化问题的迭代求解.文中证明了正交张量子空间约束条件下源域与目标域表征误差最小化问题的可解性问题,并求得了相应的解析解.在公开数据集Office-Caltech-10、Office31、ImageNet-VOC2007上与17个基线模型进行对比实验.结果表明本文所提出的方法与经典的机器学习方法、非深度域自适应方法、深度域自适应方法以及张量域自适应方法相比,在无标签目标域上的图像分类精度分别提高了10.6%~43.9%、0.7%~31.1%、0.7%~24.8%以及5.7%~34.9.同时,算法的运行效率也提高了40.5%~74.3%,显著优于所对比的基线方法.实验分析也表明,VATES方法的目标域分类精度会随着所选用神经网络特征抽取能力的增强而逐渐提升.
In recent years,machine learning has made a series of breakthroughs in computer vision applications.However,the well-trained machine learning model cannot be directly applied to other related machine learning tasks which have similar but different data distribution features.Thus,a large number of domain adaptation methods have been proposed in the literature to take full advantage of the well-trained machine learning models as well as to avoid the expensive data collection prices and the time-consuming model training costs.By extracting the common features between the source domain and the target domain simultaneously,these domain adaptation methods can transfer the well-trained model from a source domain to a different but correlated target domain efficiently.However,most existing visual domain adaptation methods cannot deal with high-order tensorial features directly.They can only convert a high-order feature tensor into a one-order high-dimensional feature vector naively by carrying out a simple vectorization operation,which not only destroys the internal structure information within the original high-order feature tensor,but also increases its computational complexity in the subsequent operations.To overcome the above challenges,in this paper,the visual domain adaptation problem is formulated as a multiple variables optimization problem by utilizing the tensor singular value decomposition(t-SVD)and the tensor product(t-product)operator.Specifically,a common tensorial subspace between the source domain and the target domain is built constructively,and the projections of the source features and the target features on this common tensorial subspace are also computed correspondingly.In order to solve the multiple variables optimization problem,a visual domain adaptation method based on tensor singular value decomposition(VATES)is proposed in this paper by adopting the tensor singular value decomposition and the alternating direction method of multipliers(ADMM).Moreover,we prove the solvability of the multiple var
作者
李国瑞
许鹏飞
彭三城
阳爱民
LI Guo-Rui;XU Peng-Fei;PENG San-Cheng;YANG Ai-Min(School of Computer Science and Engineering,Northeastern University,Shenyang 110169;Center for Linguistics and Applied Linguistics,Guangdong University of Foreign Studies,Guangzhou 510006;Laboratory of Language Engineering and Computing,Guangdong University of Foreign Studies,Guangzhou 510006;School of Computer Science and Intelligence Education,Lingnan Normal University,Zhanjiang,Guangdong 524048)
出处
《计算机学报》
EI
CAS
CSCD
北大核心
2023年第10期2084-2096,共13页
Chinese Journal of Computers
基金
国家自然科学基金(61876205)
河北省自然科学基金(F2020501034)
河北省高等学校科学研究项目(ZD2021403)
教育部人文社科一般项目(20YJAZH118、19YJAZH128)
广东外语外贸大学外国语言学及应用语言学研究中心语言与人工智能重点实验室招标课题(LAI202306)的资助.
关键词
机器学习
域自适应
张量
奇异值分解
子空间
machine learning
domain adaptation
tensor
singular value decomposition
subspace
