Stable multi-label boosting for image annotation with structural feature selection  

作　　者：ZHUANG YueTin HAN YaHong WU Fei YANG JiaCheng 

机构地区：[1]College of Computer Science, Zhejiang University, Hangzhou 310027, China

出　　处：《Science China(Information Sciences)》2011年第12期2508-2521,共14页中国科学（信息科学）（英文版）

基　　金：supported by National Natural Science Foundation of China(Grant Nos.90920303,61070068);National Basic Research Program of China(Grant No.2009CB320801);Program for Changjiang Scholars and Innovative Research Team in University(Grant Nos.IRT0652,PCSIRT);Han YaHong is supported by Scholarship Award for Excellent Doctoral Student granted by Ministry of Education of China

摘　　要：Automatic annotating images with appropriate multiple tags are very important to image retrieval and image understanding. We can obtain high-dimensional heterogenous visual features from real-world images to describe their various aspects of visual characteristics, such as color, texture, and shape. Different kinds of heterogenous features have different intrinsic discriminative power for image understanding. The selection of groups of discriminative features for certain semantics is hence crucial to make the image understanding more interpretable. This paper proposes an approach, called stable multi-label boosting with structural feature selection （S-MtBFS）, for image annotation. S-MtBFS comprises two steps, namely structural feature selection for each label and stable multi-label boosting by curds and whey. In the first step, a （structural） sparse selection model is learned to identify subgroups of homogenous features for the purpose of predicting a certain label. Moreover, a stable method of multi-label boosting with a re-sampling policy is employed in the second step to utilize the correlations among multiple tags. Extensive experiments on public image datasets show that the proposed approach has better and stable performance of multi-label image annotation and leads to a quite interpretable model for image understanding.Automatic annotating images with appropriate multiple tags are very important to image retrieval and image understanding. We can obtain high-dimensional heterogenous visual features from real-world images to describe their various aspects of visual characteristics, such as color, texture, and shape. Different kinds of heterogenous features have different intrinsic discriminative power for image understanding. The selection of groups of discriminative features for certain semantics is hence crucial to make the image understanding more interpretable. This paper proposes an approach, called stable multi-label boosting with structural feature selection （S-MtBFS）, for image annotation. S-MtBFS comprises two steps, namely structural feature selection for each label and stable multi-label boosting by curds and whey. In the first step, a （structural） sparse selection model is learned to identify subgroups of homogenous features for the purpose of predicting a certain label. Moreover, a stable method of multi-label boosting with a re-sampling policy is employed in the second step to utilize the correlations among multiple tags. Extensive experiments on public image datasets show that the proposed approach has better and stable performance of multi-label image annotation and leads to a quite interpretable model for image understanding.

关 键 词：image annotation structural feature selection multi-label boosting STABILITY 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TH126[自动化与计算机技术—计算机科学与技术]