一种基于模型共享的半监督多标签图像学习法  

Model Shared Semi-supervised Learning Approach for Multi-label Image

作　　者：张大鹏[1,2] 闻佳[1] 刘曦[3] 

机构地区：[1]燕山大学信息科学与工程学院,秦皇岛066004 [2]桂林电子科技大学广西可信软件重点实验室,桂林541004 [3]中国科学院计算技术研究所智能信息处理重点实验室,北京100190

出　　处：《系统仿真学报》2012年第9期1826-1830,共5页Journal of System Simulation

基　　金：广西可信软件重点实验室开放基金(KX201212);秦皇岛市科学技术研究与发展计划项目(201001A043)

摘　　要：提出一种快速且有效的半监督多标签学习方法:模型共享半监督推举。该方法能发现、共享并组合多个基模型,每个基模型是在某个标签上利用半监督支持向量机(S3VM)上学习的。通过使用模型共享,标签关联被显示地利用且对于每个标签来说只需要少量的基模型即可生成最后的决策结果。在Corel5k和Mediamill数据集上评估方法,实验结果显示的方法与当前流行的监督和半监督多标签学习方法是可比的。An efficient and effective semi-supervised multi-label learning approach called model-shared semi-supervised boosting was proposed. The approach found, shared and combined a number of base models across multiple labels, where each base model was learned from labeled and unlabeled data by use of a semi-supervised support vector machine （S3VM）. The label correlations were explicitly incorporated by use of model sharing and only a small number of base models are needed to generate final decision function for each label. The proposed method on Corel and Mediamill dataset was evaluated, showing results competitive with the state-of-the-art semi-supervised multi-label learning approach and some supervised techniques.

关 键 词：模型共享 半监督学习 推举 半监督支持向量机 多标签图像 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]