摘要
视频和图像中的人脸蕴涵了丰富的语义信息,可以使用人脸对视频内容进行分析与标注,尤其是视频新闻节目。而要达到这样的目的,就必须先将对视频新闻具有语义价值的人脸从视频流中检测出来。本文提出基于语义人脸检测的视频新闻语义聚类与标注算法:在这个算法中,首先使用肤色模型检测人脸可能出现区域,然后提取人脸可能区域的独立成分特征,用训练好的支持向量机检测出所有人脸,套用语义人脸模板过滤出最终的语义人脸集合,最后通过高斯混合聚类,将视频新闻标注为主持人镜头、访谈类新闻镜头和其他新闻故事镜头三类。实验表明,该算法在视频新闻结构化中可以得到较好的应用。
The human faces in video and image imply lost of semantic contents , thus we can use faces to index and analyze video contents,especially for video news. In order to realize such goal,semantic human faces must be detected and recognized from video stream. This paper presents a new algorithm for semantic clustering and indexing of video news based on semantic-face: in this algorithm,complexion model is first used to detect possible face area; then pre-trained face/non-face support vector machine is used coarse-grained to recognize face and non-face respectively based on face independent component features from possible face area; third, the semantic-face template is used to filter out non-semantic-faces and we get legible and obverse semantic-faces; in the end,video news is segmented and classified into anchorperson shot, interview shot and other news story shot through mixture Gaussian clustering of semantic-faces. Structure used to index and explorer the video news is established. Experiment shows this algorithm works well for video news indexing.
出处
《计算机科学》
CSCD
北大核心
2004年第5期187-192,共6页
Computer Science