一种基于多互补分类器的自动视频语义标注方法  

An Automatic Video Annotation Method Based on Multiple Complementary Classifiers

作　　者：宋彦[1] 李厚强[1] 戴礼荣[1] 唐金辉[1] 华先胜[2] 

机构地区：[1]中国科学技术大学电子工程与信息科学系,安徽合肥230027 [2]微软亚洲研究院,北京100081

出　　处：《小型微型计算机系统》2007年第11期2085-2089,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(603333020)资助.

摘　　要：在现有互训练(Co-Training)算法的基础上,提出了一种基于多个互补型分类器的半监督学习(Semi-Supervised Learn-ing)方法,并将其应用到自动视频语义标注框架中.该方法通过构建基于特征互补和模型互补的多个分类器对未标注样本中的隐含信息加以利用,并结合视频序列中概念分布的时间相关性和局部聚集性等特性提高了分类的准确性,相对于有监督学习方法提高了约7%左右.A novel semi-supervised learning method based on multiple complementary classifiers is proposed according to the analysis of the existing co-training algorithms, which is applied to the automatic video semantic annotation. By constructing the classifiers with feature complementarity and model complementarity,the hidden information in unlabeled dataset is effectively exploited. Moreover,the local consistency and temporal relationship of video sequences is further applied to improve the final annotation accuracy, the improvement of annotation accuracy is about 7%.

关 键 词：视频语义标注 半监督学习 互训练算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]