基于联合子空间与多源适应学习的多标签视觉分类

Multi-label visual classification based on joint subspace and multi-source adaptation learning

传统的视觉分类方法普遍忽视了多标签间的相关性和大量相关源数据的判别信息.为此,基于共享子空间和领域适应学习方法,针对多标签视觉分类问题,提出了一种新的联合子空间和多源适应学习的多标签视觉分类方法,简称为多源适应多标签学习(Multi-Source adaptation Multi-Label learning,MSML).MSML将综合考虑多标签相关性、灵活的特征相似性嵌入和多源模型的适应学习等目标,并将其融为一个统一的学习模型,其全局最优解只需通过一个广义特征分解问题的求解便可获得.在视频概念识别、自动图像标注等实际应用中进行比较分析,结果显示了本文方法的有效性和优越性.

Traditional visual classification methods usually ignore the correlations among different tags, and the discriminative information existed in lots of related auxiliary source domains.In this paper on the basis of the advances of shared subspace and multi-source adaptation learning research, a novel joint shared subspace learning and multi-source adaptation multi-label （MSML）visual classification method is proposed.Specifically,MSML simultaneously considers the label correlation,flexible feature similarity embedding,and multi-source model adaptation,and integrates them into a unified learning model.The results show that the globally optimal solution of the proposed method can be obtained by performing generalized eigen-decomposition.We evaluate the proposed method on two real-world visual classification tasks such as video concept detection and automatic image annotation.The experimental results show that the proposed algorithm is effective and can obtain comparable or even superior to related works.