基于音频语谱图像识别的广告检索  

Advertisement retrieval based on audio spectrogram image recognition

作　　者：姜洪臣[1] 任晓磊[1] 赵耀宏[2] 徐波[1] 

机构地区：[1]中国科学院自动化所数字内容技术与系统研究中心,北京100190 [2]北京交通大学软件学院,北京100044

出　　处：《清华大学学报（自然科学版）》2011年第9期1249-1252,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家科技支撑计划项目(2009BAH40B03)

摘　　要：随着中国广播电视事业的快速发展,对广播电视进行广告监测的重要性越来越突出。该文提出了一种基于音频语谱图像识别的广告检索方法,该方法把每个广告片段的音频转换成语谱图像,并通过boosting训练算法,从中提取可区分的、适合索引的特征以训练分类器。检索时,通过特征对齐,采用基于期望最大化(expectation maximization,EM)的迭代方法识别最匹配的广告。实验结果表明:该方法可以快速准确地从广播电视中检索到广告,而且具有很好的鲁棒性,基于该方法开发的广播电视广告监测系统已经得到了实际应用。With the rapid development of broadcast and television in China,advertisement monitoring and retrieval of broadcast and television images is becoming more and more important.This paper presents an advertisement retrieval approach based on audio spectrogram image recognition.After converting the audio signal into a spectrogram image,the system extracts the discriminative and robust features with training of weak classifiers using a boosting algorithm.Queries use an expectation maximization（EM） based algorithm to recognize advertisement clips that best match the advertisement templates through feature bit alignment.Tests using real television clips show that the approach is accurate,fast and robust.In addition,an advertisement monitoring system was then developed based on this approach.

关 键 词：信息处理 广播电视 广告检索 语谱图像识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]