基于音频的视频拷贝检测

Audio-Based Duplicate Video Detection

提出了一个兼具实时性和鲁棒性的短视频拷贝检测框架,并以长视频中的广告检测为应用进行研究。框架主要分为两个部分:1.基于音频匹配的短视频拷贝检测;2.基于图像的边界精确定位。第一阶段从原始视频中获得声谱图,并进一步地提取出音频特征并在音频库中寻找匹配;在第二阶段,对检测到的视频片段进行合并,然后利用视频片段在边界处附近的图像特征,对边界片段进行精确定位。实验结果显示,准确率高达99%,取得了很好的效果。

This paper proposed an algorithm framework,which is both real-time and robust,for short video copy detection,and researched the application in terms of long video commercials detection.The framework was consisted of two parts：1.Audio-Based short video copy detection; 2.Vision-based precise video border location.Stage 1 of study,which extracted spectrogram from original video,and further extracted audio features,then searched for the match in the sample library.In stage 2,it merged candidate video clips,then precisely pinpointed the video border using image feature along with the candidate border area.Experimental results show accuracy rate of up to 99％,has yielded very good results.