基于视觉特征提取的压缩域在线视频摘要快速提取

Online video abstract extraction based on visual features in compressed domain

为了在有限的时间内产生质量可接受的视频摘要以达到在线使用的要求,提出一种基于视觉特征提取(visual features extraction,VFE)的压缩域视频摘要快速提取方法。从每帧输入视频中提取视觉特征,采用零均值归一化交叉相关(zero mean normalized cross correlation,ZNCC)指标检测有相似内容的视频帧组,为每组选择代表性帧,运用2个量化直方图过滤所选择的帧,从而避免视频摘要中可能的冗余或无意义帧。在视频检索国际权威评测(TREC video retrieval evaluation,TRECVID)2007数据集上的实验结果表明,与基于聚类的高斯混合模型、基于熵的模糊C均值聚类和关键帧提取方法相比,该方法提取的视频摘要质量更高,且在时间和空间复杂度上具有明显优势,适合在线实时处理。

In order to produce acceptable quality of video abstract in a limited period of time,and achieve the online requirement,a fast video summarization method in compressed domain based on visual features extraction（ VFE） is proposed. Firstly,visual features from each frame of the input video are extracted. Then,the zero mean normalized cross correlation（ ZNCC） is used to detect similar content video frames,and the representative frames for each group are selected.Finally,two quantized-histogram filters are used to select frames to avoid possible video redundant or meaningless frames.The experimental results on the TREC video retrieval evaluation（ TRECVID） 2007 data show that proposed method has higher video summary quality than the Gaussian mixture model based on clustering,entropy based fuzzy C means clustering and key frame extraction method,it has obvious dominance in the time and space complexity,and it is suitable for online real-time processing.