基于层次特征融合哈希的近似重复视频检索方法

Hierarchical feature fusion hashing for near-duplicate video retrieval

近年来,由于互联网上视频数量的爆炸式增长,基于视频哈希(Hash)的近似重复视频检索已经吸引了越来越多的研究者关注.在现有方法中,视频的视觉特征,包括单一特征以及多特征融合的方法被广泛应用于近似重复视频检索算法中.而低层视觉特征在表达高层语义方面存在不足,使得近似重复视频检索的性能变低.针对这一问题,本文提出了一种基于层次特征融合的视频哈希方法,用于近似重复视频检索.该方法首先从视频中提取低层人工定义特征,然后利用卷积神经网络提取中间层深度特征以及高层语义特征,最后把不同层级的特征融合起来,利用层次特征和样本之间的全局结构关系以及各特征之间的互补性,学习得到视频哈希,进而进行近似重复视频检索.该方法在CC-WEB-VIDEO数据库上进行了实验,实验结果证明本文方法与现有的方法相比在性能上有较大提升.

In recent years,owing to the rapid growth of the number of videos on the Internet,near-duplicate video retrieval (NDVR)by video hashing has attracted huge attention.In the existing methods,the visual features of videos,including single feature and multiple visual feature fusion,are widely used in the NDVR algorithms. However,low-level visual features have some disadvantages in expressing high-level semantics,which may lead to low performance in NDVR.In this paper,we propose a video hashing method for NDVR based on hierarchical feature fusion to address this issue.In the proposed method,low-level handcrafted features from videos were first extracted;then,the intermediate-level deep features and high-level semantic features extracted from the convolutional neural network are obtained.Finally,these semantic features are combined with low-level visual features,where the global structural relationships and complementarity discovered among the hierarchical features are utilized to learn the hash code for NDVR.Extensive experiments are performed on the CC-WEB-VIDEO dataset;the proposed framework proves to have a better retrieval performance compared with the state-of-the-art approaches.