【作者】 王领；

【导师】 李豪杰；

【作者基本信息】 大连理工大学 ， 软件工程， 2017， 硕士

【摘要】 随着互联网技术的发展,人们可以随时随地进行视频拍摄并上传,造成了网络视频数据的大规模增长。同时,盗版侵权等违法行为也制约着网络视频的发展。为了快速地对相似性视频进行检索,基于内容的视频拷贝检测应运而生。传统的基于内容的视频检测方法,使用局部特征点或者图像特征来描述视频。为了减少内存占用与时间消耗,传统方法只能对视频关键帧进行稀疏采样。然而被抛弃的视频帧拥有大量视觉信息,缺失的信息造成了精度上的不足,制约了传统方法的发展。为了改善传统方法在计算速度上的缺陷,本文在其基础上提出了一系列改进。引入大字典模型,并使用快速汉明距离计算策略加速匹配。使用输入数据修改汉明嵌入算法,并将弱几何一致性从约束图像匹配扩展到约束视频序列匹配,提高了准确率。最终在精度不输于前人的情况下极大提升了检索速度。此外,为了从根本上解决匹配精度和匹配速度的问题,本文提出了一种基于紧凑视频表征的视频拷贝检测方法。这种方法建立在稠密采集的关键帧之上,得以保留更多关键帧的有效信息。使用卷积神经网络来提取关键帧中的视觉信息,并用稀疏编码保留其中的关键信息。为了保证特征紧凑性,减轻内存和CPU负担,使用融合的方式将一段短视频内的所有帧特征融合为一个视频表征。与其他方法相比,该方法在VCDB数据集上使召回率和准确率都有了很大的提升。更多还原

【Abstract】 With the development of the Internet technology,people can record and upload videos in anywhere and anytime,which results in large-scale growth of online video data.At the same time,piracy and patent infringement of videos also constrain the growth of video.In order to detect the similar videos efficiently,content-based video copy detection came into being.Traditional content-based video copy detection methods use local features or image features to describe videos.Traditional methods sparsely sample the key frame to decrease the memory usage and time consuming.But the abandoned video frames also include a lot of visual information,which will make an inaccurate retrieval result.In order to improve the shortcomings of the traditional method in speed,we propose a series of improvements to make it better.We use the big vocabulary method and quick strategy of hamming distance computation to speed up the retrieval stage.Then we use the weak geometric consistency to constraint the matched video sequence to improve the accuracy rate.Finally,we get a method with super-fast speed and competitive accuracy rate.In addition,we try to fundamentally solve the problem of matching accuracy and speed.We propose a video copy detection method based on compact video representation.The convolution neural network is used to extract the visual information and the sparse coding is used to retain the key information.And we fuse the features of frame into a compact video feature to reduce the usage of memory and CPU.The compact video feature represents a short time interval video and keep compact enough.Compared with other methods,our method increases both the recall rate and accuracy rate.更多还原