摘要： 针对视觉词袋模型的量化误差与视觉词含糊性，提出一种基于视觉词模糊权重的视频语义标注方案。该方案在训练样本集的预聚类基础上，逐个聚类训练单类支持向量机OC-SVM。根据样本特征与聚类超球球心的距离函数及聚类超球的空间分布确定视觉词映射及权重，以提高视觉词的表达力、区别力。实验结果表明，基于该方案的视频语义标注精度分别比TF方案和VWA方案提高34%和16%。

关键词: 视频语义标注, 视觉词袋模型, 模糊权重方案, 单类支持向量机, 聚类超球, 模糊隶属度

Abstract: This paper proposes a formulation of visual word weighting scheme Fuzzy Weighting Scheme(FWS) aiming at the Bag of Visual Word(BoVW) model vector quantization loss and visual word ambiguity. Based on K-Nearest Neighbors(KNN) pre-clustering results, One-Class Support Vector Machine(OC-SVM) on each clustering samples subset is trained. Visual words corresponding to a single local visual feature vector are determined according to the spatial distribution information of clustering-hyperspheres and fuzzy weights are evaluated according to the distance function between sample feature and center of clustering-hypersphere. FWS is designed to boost the visual word expressiveness and discriminativeness. Experimental results show that the scheme outperforms TF scheme and VWA scheme by 34% and 16% respectively on video semantic annotation precision.

Key words: video semantic annotation, Bag of Visual Word(BoVW) model, Fuzzy Weighting Scheme(FWS), One-Class Support Vector Machine (OC-SVM), clustering hypersphere, fuzzy membership degree

中图分类号: 

• TP391