面向广告监测的视频指纹提取及匹配技术研究

发布时间：2018-10-17 09:01

【摘要】： 视频广告为商业繁荣发挥了积极作用。电视广告中也存在着虚假违法、内容庸俗、夸大宣传等现象,造成了恶劣的社会影响,损害了媒体的社会公信力。同时,违规网络视频广告也呈泛滥趋势。为了保障消费者利益,营造一个积极的广告产业氛围,迫切需要对电视及网络的视频广告进行监测的有效技术手段。视频广告监测要求具有准确性和实时性。因视频指纹具有区分性和鲁棒性的性质,它可以确保监测结果准确可靠,同时视频指纹的易提取性与匹配算法的高效性又能实现实时性,因此,本文选择使用视频指纹进行视频广告监测的研究,重点探索鲁棒的视频指纹的提取以及高效匹配算法。本文对已有方法进行了综合分析,并针对现有指纹技术鲁棒性较差、匹配方法效率不高或适用度不广等缺点,在以下方面展开探索: 1.在视频指纹提取阶段,使用了一种对特征赋以不同权重方式结合的多特征视频指纹,得到了区分性更好、鲁棒性更强的视频指纹。该视频指纹由图像角点特征和时域特征构成。本文改进了时域特征提取方法提取时域特征,并与改进的图像角点特征进行结合。实验表明,在亮度变换和尺寸变化以及常见的噪声攻击中,本文提取的视频指纹有较强的鲁棒性。 2.在指纹匹配阶段,考虑到子指纹序列的相似性由子指纹集合相似性和位置相似性构成,我们设计了一个两层相似性匹配框架。第一层将子指纹序列匹配的问题首先弱化为子指纹集合相似的问题,使用改进的二层位置敏感哈希加速子指纹匹配,相似的子指纹会尽可能早的进行匹配,因此节省了大量不匹配指纹之间的比较。只有当集合相似后,才在第二层使用全搜索策略。该方案结合了全搜索策略适用度广、位置敏感哈希LSH(Locality Sensitive Hashing)高效性的优点。实验表明,二层匹配方法能够进行实时匹配。
[Abstract]:Video advertising has played a positive role in the commercial boom. There are also false illegal, vulgar content, exaggerated publicity and other phenomena in television advertising, which has caused bad social impact and damaged the social credibility of the media. At the same time, illegal network video advertising is also flooding trend. In order to protect the interests of consumers and create a positive advertising industry atmosphere, it is urgent to monitor the video advertising on television and network with effective technical means. Video advertising monitoring requires accuracy and real-time. Because video fingerprint has the property of distinction and robustness, it can ensure the accuracy and reliability of the monitoring results, at the same time, the easy extraction of video fingerprint and the high efficiency of matching algorithm can achieve real-time, so, In this paper, we choose to use video fingerprint to monitor video advertisement, focusing on robust video fingerprint extraction and efficient matching algorithm. In this paper, the existing methods are comprehensively analyzed, and aiming at the shortcomings of the existing fingerprint technology, such as poor robustness, low efficiency of matching methods and low applicability, the following aspects are explored: 1. In the stage of video fingerprint extraction, a multi-feature video fingerprint with different weights is used to obtain a more discriminative and robust video fingerprint. The video fingerprint consists of corner feature and time domain feature. In this paper, an improved time domain feature extraction method is proposed and combined with the improved image corner feature. The experimental results show that the extracted video fingerprint is robust in luminance transform, size change and common noise attack. 2. In the phase of fingerprint matching, we design a two-layer similarity matching framework considering that the similarity of sub-fingerprint sequences is composed of sub-fingerprint set similarity and location similarity. In the first layer, the problem of subfingerprint sequence matching is reduced to the similarity problem of subfingerprint set, and the improved two-layer position sensitive hash is used to accelerate the sub-fingerprint matching, and the similar sub-fingerprint will be matched as early as possible. As a result, a large number of mismatched fingerprint comparisons are saved. The full search strategy is used at the second level only when the collection is similar. This scheme combines the advantages of wide applicability of full search strategy and high efficiency of position sensitive hash LSH (Locality Sensitive Hashing). Experiments show that the two-layer matching method can be used in real-time matching.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2010
【分类号】：TP391.41

【相似文献】