核可鉴别的特征分块稀疏表示的视频语义分析  

Video Semantic Analysis Based on Kernel Discriminative Features-Blocked Sparse Representation

作　　者：詹永照[1] 田华锋[1] 毛启容[1] 

机构地区：[1]江苏大学计算机科学与通信工程学院,镇江212013

出　　处：《计算机辅助设计与图形学学报》2014年第8期1290-1296,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61170126;61272211)

摘　　要：针对视频特征的多样性和稀疏字典的冗余特点,提出一种基于核可鉴别的特征分块稀疏表示的视频语义分析方法.首先按照实际需求提取视频段多种特征,并根据各种特征的维数大小分别建立其分块稀疏字典,对每个分块字典在K-SVD算法基础上加入核可鉴别准则进行优化,使各种特征的稀疏表示特征具有更好的类别鉴别能力;在对视频段进行语义分析时,使用优化字典求解各种特征的稀疏表示特征,并对各种特征的稀疏表示特征采用加权KNN算法进行类别分类分析,最后依据各种特征对决策分析的支持度进行视频段的语义融合分析.实验结果表明,该方法有效地提高了视频语义分析的准确性和分析速度.For the characteristics of video feature diversity and sparse dictionary redundancy,a method of video semantics based on discriminative features-blocked sparse representation is presented in this paper.Firstly,various features are extracted from video segments as actual requirement.Then each blocked dictionary for the sparse representation is constructed in terms of the dimensional size of each feature.The dictionaries are optimized by embedding discriminative criterion into K-SVD algorithm in order to improve category identification ability of sparse representation feature.In the semantic analysis for video segment,the sparse features of any kinds of primitive features are obtained by the optimized dictionaries.The weighted KNN is adopted for classification analysis of these sparse features.Finally,the video segment semantic is analyzed by fusing analysis of the sparse features according to the decision support degrees.The experimental results show that the proposed method can effectively improve the accuracy and speed of video semantic analysis.

关 键 词：视频语义分析 核可鉴别性 特征分块稀疏表示 融合分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]