面向视频语义分析的局部敏感的可鉴别稀疏表示

Locality-sensitive Discriminant Sparse Representation for Video Semantic Analysis

视频语义分析已经成为人们研究的热点。在传统稀疏表示方法中,相似视频特征未必能产生相近稀疏表示结果。在基于稀疏表示的视频语义分析中,假定相似的视频数据样本的稀疏表示也相似,即两个相似视频特征的稀疏系数之间的距离较小。为了提高视频语义分析的准确性,基于该假设提出一种面向视频语义分析的局部敏感的可鉴别稀疏表示方法。该方法在局部敏感稀疏表示中引入基于稀疏系数的鉴别损失函数,优化构建稀疏表示的字典,使稀疏表示特征满足类内离散度小、类间离散度大的Fisher准则,并建立可鉴别稀疏模型。为验证所提方法的有效性,在相关视频数据库中将其与多种算法进行对比,实验结果表明,该方法显著地提高了视频特征稀疏表示的鉴别性,有效地提高了视频语义分析的准确性。

Video semantic analysis has been a research hotspot. Traditional sparse representation methods cannot pro- duce similar coding result when the input video features are close to each other. We assumed that similar video features should he encoded as similar sparse codes in the process of video semantic analysis based on sparse representation. In other words, the similar video features should have smaller distance between their sparse codes. In order to improve the accuracy of video semantic analysis, locality-sensitive discriminant sparse representation（LSDSR） based on the hypothe- sis for video semantic analysis was developed. In proposed method, discriminant loss function based on sparse coefficient is introduced into the locality-sensitive sparse representation. An optimization dictionary is generated with the con- straint. In the process, the sparse coding coefficients have both small within-class scatter and large between-class scatter using Fisher criterion, so as to build the discriminant sparse model in the LSDSR. The proposed method was extensively evaluated on related video databases in comparison with existing sparse representation methods. The experimental re- sults show that this method significantly enhances the power of discrimination of sparse representation features, and consequently improves the accuracy of video semantic analysis.