基于深度学习的重复视频检测

发布时间：2018-09-19 18:32

【摘要】：近两年来,随着深度学习的快速发展,传统方法无法解决的问题正在不断被攻克。特别是在图像、语音和文本领域,深度学习的技术结合更高性能的GPU计算,已经逐渐成为一种趋势。在互联网时代,多媒体视频作为一个重要的数据载体,在信息传播中发挥着巨大的作用。然而信息膨胀,给多媒体视频的检索带来了挑战。重复视频检测,不仅可以减少冗余信息的传播,也在保护版权和防止不良视频传播中起到了作用。基于内容的重复视频检索,一般可以分为三个部分,1.关键帧提取2.特征提取和索引3.匹配过滤。当前的方法中,大多都是利用图像的局部特征。随着深度学习的兴起,卷积神经网络在图像识别和检索中取得了很好的效果,但并没有被广泛应用到视频检索的相关领域之中。本文的主要工作有两个,一个是提出了比较快速的视频关键帧提取的方法,适用于时长不超过30分钟的短视频；另一个是比较了卷积神经网络的图像特征和局部点的SIFT特征在重复视频检索中的效果。在关键帧提取中,本文从性能和效果出发,分别提出了基于镜头分割和基于聚类的方法。在重复视频检测中,本文基于CC WEB VIDEO数据集,从关键帧提取、图像特征选取以及匹配排序策略三个角度,设计并完成了实验。本文的实验结果表明,关键帧的选取,能显著降低检索时的时间与空间开销；卷积神经网络特征的检测准确率要明显高于SIFT特征,当查全率为0.8时,查准率不低于0.97；采用关键帧加权匹配策略,能够在一定程度上提高排序的准确度。
[Abstract]:In the past two years, with the rapid development of deep learning, the problems that can not be solved by traditional methods are being solved. Especially in the fields of image, speech and text, the technology of depth learning combined with higher performance GPU computing has gradually become a trend. In the Internet era, multimedia video, as an important data carrier, plays an important role in information dissemination. However, the expansion of information brings challenges to multimedia video retrieval. Repetitive video detection not only reduces the spread of redundant information, but also plays an important role in protecting copyright and preventing bad video transmission. Content based repetitive video retrieval can be divided into three parts. Key frame extraction 2. Feature extraction and index 3. Match filter. Most of the current methods are based on the local features of the image. With the rise of deep learning, convolutional neural networks have achieved good results in image recognition and retrieval, but have not been widely used in video retrieval related fields. There are two main works in this paper. One is to propose a fast method of key frame extraction, which is suitable for short video with less than 30 minutes. The other is to compare the image features of convolutional neural networks and the SIFT features of local points in repetitive video retrieval. In the key frame extraction, based on the performance and effect, this paper proposes shot segmentation and clustering based methods respectively. In repetitive video detection, based on CC WEB VIDEO data set, experiments are designed and completed from three angles: key frame extraction, image feature selection and matching sorting strategy. The experimental results show that the selection of key frames can significantly reduce the time and space overhead of retrieval, and the detection accuracy of convolution neural network features is obviously higher than that of SIFT features, and when the recall rate is 0. 8, the precision rate is not less than 0. 97. The accuracy of sorting can be improved to a certain extent by using key frame weighted matching strategy.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】