基于交叉重构的领域自适应算法

Cross Reconstruction-Based Domain Adaptation

摘要: 领域自适应是解决跨域识别的有效方法，它是迁移学习在计算机视觉方面的有效应用，将源域学到的知识迁移到目标域的识别任务中，有效解决目标域标注数据不足的问题。本文提出了一种新的基于交叉重构的领域自适应方法(Cross Reconstruction-based Domain Adaptation, CRDA)，通过对原始源域和目标域的交叉重构来构造新的源域与目标域，使得同类数据相互交织，缩短同类数据间的距离。并通过对重构矩阵施加低秩约束，将两个域的同类数据对齐，以此来充分挖掘源域和目标域同类数据之间的内在结构信息，并利用该结构信息来学习分类器，从而取得更好的跨域识别效果。在五个公开数据集上的实验结果表明CRDA有着较高的跨域识别准确率。

Abstract: Domain adaptation is an effective method to solve the problem of cross-domain recognition. It is an effective application of transfer learning in computer vision, which transfers the knowledge learned in the source domain to the recognition task of the target domain, and effectively solves the problem of insufficient labeled data in the target domain. In this paper, a new Cross Reconstruction-based Domain Adaptation (CRDA) method is proposed, which constructs a new source domain and target domain through the cross reconstruction of the original source domain and target domain, so as to make the same kind of data intersect with each other and shorten the distance between the same kind of data. By applying low-rank constraints to the reconstruction matrix, the same kind of data in the two domains are aligned, so as to fully mine the internal structure information between the same kind of data in the source domain and the target domain, and use the structure information to learn the classifier, so as to achieve better cross-domain recognition effect. The experimental results on five open datasets show that CRDA has a high cross-domain recognition accuracy.