摘要
针对互联网相似视频内容检测问题,提出了基于短空时变化的鲁棒视频哈希算法。特征提取和特征量化是该算法的两个关键步骤。在特征提取中,与现有基于时空信息融合的特征提取方法相比,该算法的创新性在于充分利用相邻帧之间局部空域信息的短时变化(简称"短空时变化")来提取特征。该算法首先构造视频内接球,并以球心为起点对内接球进行划分,获取一系列内接球环,从而捕捉相邻帧的空域信息的短时变化,然后将球环非负矩阵分解系数作为视频内容进行特征表示;在特征量化中,该算法采用改进的曼哈顿量化策略将视频特征映射成二进制的哈希序列,更好地保留了原空间中的近邻关系,提高了量化的准确度。实验结果表明,该算法具有良好的性能。
A robust video hashing algorithm based on short-term spatial variations was proposed to detect near-duplicate videos in the Internet.Feature extraction and feature quantization are key steps in this algorithm.In the feature extraction phase,compared to the existing feature extraction methods based on temporal and spatial information fusion,the innovation of the proposed algorithm is to make full use of short-time variations of local spatial information between adjacent frames(referred to"short-term spatial variations").In the proposed algorithm,inscribed spheres of the video are constructed first,and then a series of spherical tori are obtained by partitioning the inscribed spheres with the center of the sphere as the starting point to capture short-term changes in spatial information between adjacent frames.After that,the decomposition coefficients by non-negative matrix factorization of spherical tori are used as the feature representation of the video.In the feature quantization phase,to map the feature representation into binary hash sequences,the optimized Manhattan hashing strategy is adopted which better reserves the neighborhood structure in the original data space,and thus improves the accuracy of quantization.Experiments were carried out on a video dataset to evaluate the performance of the proposed video hashing method.Experimental results show that the proposed algorithm has good performance.
出处
《计算机科学》
CSCD
北大核心
2018年第2期84-89,共6页
Computer Science
基金
国家自然基金项目(61671274)
中国博士后科学基金项目(2016M592190)
山东省高等学校科技计划项目(J17KB161)
山东省高等学校优势学科人才团队培育计划资助
关键词
视频哈希
时空信息
非负矩阵分解
相近视频检测
曼哈顿哈希
Video hashing
Spatio-temporal information
Nonnegative matrix factorization
Near-duplicate video detection
Manhattan hashing