基于非线性可鉴别的稀疏表示视频语义分析方法

Video semantics analysis based on nonlinear identifiable sparse representation

为了提高视频语义分析的准确性,提出一种基于非线性可鉴别的稀疏表示视频语义分析方法.该方法在K-SVD稀疏表示字典优化算法中引入了核函数和类别矢量,将稀疏表示特征映射到高维空间并使之满足类内离散度小、类间离散度大的Fisher准则,建立了可鉴别模型.使用该模型对字典的优化求解产生约束,形成了优化字典.将该字典用于求解视频特征的稀疏表示,同时提出了视频特征稀疏表示的分类鉴别准则来分析视频语义.在TRECVID 2007的新闻视频库上进行了视频语义概念分析.试验结果表明,该方法显著提高了视频特征稀疏表示的鉴别性,从而提高了视频语义分析准确性.

In order to improve the accuracy of the video semantic analysis,the video semantic analysis was proposed based on nonlinear identifiable sparse representation.Kernel function and category vector were introduced into the K-SVD dictionary optimization for the sparse representation.The sparse representation features were mapped into a high dimensional space to establish label identifiable model by Fisher criterion.An optimization dictionary was generated according to the constraint of the proposed model.The sparse representation codes of the video features were calculated by the dictionary.An identification criterion was proposed for the classification of the video sparse representation features to analyze the video semantic by the criterion.The video semantic concept was analyzed in the news video library of TRECVID 2007.The experimental results show that the discriminability of the sparse representation of video features can be markedly improved with improved accuracy of video semantic analysis.