基于音视频特征的新闻拆条算法

News Video Segmentation Algorithm Based on Audio and Video Features

随着人们生活节奏的加快和网络信息技术的迅猛发展,对新闻视频节目的存储和再利用需求日益剧增,如何将较长的新闻视频节目按其内容拆分成多个新闻条目成为了一个有意义的课题。提出了一种基于音视频特征的新闻拆条算法,仅提取了新闻视频在视觉、音频上的基本特征即主持人特征和音频静音段特征进行分析。通过人脸识别提取主持人特征,使用短时能量和过零率提取静音特征,并对其加以条件筛选,结合这两个特征完成拆条工作。针对总计时长3 000分钟的新闻联播节目进行实验,得到较好的实验结果:召回率0.856 3,准确率0.932 6和F1值0.892 8。且视频边界的准确度精确到帧。同时分析了静音段长度阈值、限制条件和毛刺现象对于新闻拆条结果的影响。

The acceleration of life rhythm and rapid development of the Internet information technology lead to the increasing demand of news video storage and reutilization.It is a meaningful subject about how to segment a long news video into several small news items.A news video segmentation algorithm based on audio and video features is proposed.It only extracts anchor person feature and silence track feature which are the basic visual and audio features,respectively.Anchor person feature is detected via face recognition are combined Short-time energy and zero-crossing rate are used to detect silence feature and filter them with condition.Two features are combined to analyze the borders of different news item.The experiments are performed with news videos which are 3000 minutes in total,and the result is great：recall rate is 0.856 3,precision 0.932 6 and f1 measure 0.892 8.The accuracy of the border detected is frame.Analysis has been done about the influence of the length,restrictions and redundancy of silence track to the result.