一种基于音频匹配的广告检测算法

摘要

Commercials detection is to extract advertising sequence out of TV shows automatically .In traditional ways , it adopted algorithm based on computer vision technology framework to detect advertisement in terms of video content .But it wouldn ’ t meet the demands on performance or efficiency for commercial usage .The algorithm described here used only audio information for de-tection.It first extracted original audio information from video , then used short-time Fourier transform to transform it into a spec-trogram .And it applied binary threshold with a set of pre-boosting filters to get local feature descriptors .During the detect phase , it obtained descriptors in the same methods described above , and detected results in offline descriptor library .Audio-based detec-tion algorithm for commercials detection had smaller storage , higher accuracy , and better real-time.Experiments showed that this algorithm could significantly improve robustness and performance of commercials detection systems , and can be applied in reality .%广告检测是指从电视节目中自动地检测出广告序列。传统方法使用基于计算机视觉算法的技术框架对视频内容进行广告检测，无法在性能和效率上满足商业化需求。本算法仅利用广告音频信息完成广告检测工作。首先抽取广告库的原始音频信息，使用短时傅里叶变换将其转化为声谱图，再应用预筛选的滤波器集合进行二值化，得到局部特征描述子，组成广告音频库；其次在广告检测过程中，以同样方法提取其特征描述子，在广告音频库中检索得到检测结果。基于音频匹配的广告检测算法具有存储小、准确度高、实时性强等诸多优点。实验表明该算法可以显著提高广告检测系统的鲁棒性和性能，可应用于现实场景中。