基于多特征相似度曲线曲率检测的关键帧提取

Key frame extraction based on curvature detection of multi-feature similarity curve

网络多媒体的迅猛发展和普及使得对海量视频信息进行快速和低成本管理的需求日益迫切,而关键帧可以大大减少视频索引的数据量,同时也为查询和检索视频提供了一个组织框架。针对现有关键帧提取算法存在的特征选取单一、阈值选择困难和视频类型局限性等问题,提出了一种基于多特征相似度曲线最大曲率点检测的关键帧提取方法。算法利用多特征融合的相似性度量来捕获视频内容的显著变化,弥补了单一特征对视频内容描述不充分的不足,且基于滑动窗口的检测算法无需阈值选择,可以实时、局部地提取关键帧,解决了传统算法计算量大、通用性差的问题。最后通过实验利用一种保真度评估标准验证了该算法的有效性。

With the development of network multimedia, quick and low-cost management of the massive video information is urgently needed. Key frames can decrease data quantity in video indexing and provide a framework for video retrieval and indexing. A key frame extraction method based on high-curvature points detection was proposed. First, three descriptors of color histogram, edge direction histogram and wavelet statistics were used to describe visual content, and combined to form a frame difference measure. Then the key frames were got by detecting high curvature points. Experimental results show that the proposed algorithm is rapid and efficient. It can capture precise changes dynamically in the content of video sequences, and can extract the key frames on the fly. Finally, a fidelity evaluation proves the effectiveness of the algorithm.