摘要
针对现有的基于无监督聚类的视频关键帧提取方法没有考虑镜头内容的时序性、对初始类的划分较敏感、易陷入局部最优等问题,提出了一种新的基于人工免疫的有序样本聚类算法.在传统人工免疫聚类算法的基础上引入了抗原记忆识别机制及改进了抗体的克隆与超变异机制,并在此基础上给出了基于人工免疫有序聚类的视频关键帧提取方法.该方法将镜头帧序列看成一个入侵机体的抗原序列,然后基于首次应答与再次应答机制依次为每个抗原产生记忆细胞池,最终每个记忆细胞池能识别的邻近抗原对应一个类别并选取距其类中心最近的帧为关键帧.对大量不同类型的视频进行了试验.结果表明,该方法能得到较高的保真度和压缩率,能够十分有效地提取出反映镜头内容变化的关键帧.
In view of sensitivity of initial samples division and easy fall into local optimum in existing un- supervised clustering methods of video key-frame extraction without considering temporal order of shot content, a novel ordered samples clustering algorithm was proposed based on artificial immune. Accor- ding to traditional artificial immune based clustering algorithm, memory recognition mechanism for anti- gens was introduced, and clone and hyper mutation mechanism for antibodies was improved. A video key-frame extraction method of artificial immune based on ordered samples clustering was proposed. The image frame sequence in a shot was regarded as an antigen sequence invading the body. Based on the mechanisms of primary response and second response, the memory cell pools for each antigen were ob- tained. The continuous antigens recognized by the same memory cell pool was formed a cluster, and the frame nearest to the cluster center was extracted as a key-frame. The experiments on various videos were completed. The results show that the proposed method has high fidelity and compression ratio, and the key-frames reflecting real shots contents can be extracted effectively.
出处
《江苏大学学报（自然科学版）》
EI
CAS
北大核心
2012年第2期199-204,共6页
Journal of Jiangsu University：Natural Science Edition
基金
国家自然科学基金资助项目(61170126)
江苏省高校研究生科研创新计划项目(CX08B_097Z)
关键词
关键帧提取
人工免疫
有序样本聚类
视频摘要
视频检索
key-frame extraction
artificial immune
ordered samples clustering
video summarization
video retrieval
