基于卷积神经网络和频率域特征的视频拷贝检测方法  

Video copy detection based on deep learning features and DCT coefficients

作　　者：石慧杰 Shi Huijie(School of Computer and Information Engineering, Beijing Technology and Business University, Beijing 100048)

机构地区：[1]北京工商大学计算机与信息工程学院

出　　处：《高技术通讯》2019年第12期1201-1205,共5页Chinese High Technology Letters

基　　金：国家重点研发计划(2017YFB1401000)资助项目

摘　　要：为了解决视频特征鲁棒性差、计算复杂度高等问题,提出一种新的视频拷贝检测方法。该算法将深度卷积网络特征和传统手工特征相结合,提升特征检测的维数,提升检测准确度。方法首先使用密集连接卷积网络(DenseNet)提取关键帧的深度特征,并对关键帧进行离散余弦变换(DCT)提取系数特征,然后使用基于典型相关分析(CCA)的特征融合算法将2种特征进行有效融合,最后使用融合特征进行特征匹配。在标准数据集上的实验表明,本文提出的算法检测效果较好,在常见的拷贝变化下可以得到更高的检测精度。该算法可以作为一种有效的数字视频版权保护技术应用于数字视频的监管领域。Many content-based video copy detection(CBCD)techniques have been proposed to identify the copies of a copyrighted video.Early algorithms using traditional features has the problems of weak robustness to viewing angle changes and high computational complexity.In recent years,deep learning has been used to get better detection results.Due to the poor performance from single visual features,an algorithm combining deep convolutional network features with traditional features is proposed.The deep features with dense convolutional network(DenseNet)and the ordinal measures of the coefficients of its discrete cosine transform(DCT)from the sampled video frames are first extracted,then the fusion features of these two features based on the canonical correlation analysis(CCA)algorithm are obtained.With the matching of fusion features,the copied videos are detected.Experiments on standard datasets show that the proposed method obtains a significant performance against common geometric attacks.

关 键 词：视频拷贝检测 特征表示 卷积神经网络(CNN) 典型相关分析(CCA) 离散余弦变换(DCT) 密集连接卷积网络(DenseNet) 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]