随着内容分享网络的发展,网络上的视频数据急剧增长,出现了大量的非法拷贝。为了减少版权侵犯纠纷,需要检测出网络上的非法拷贝。视频指纹是实现拷贝检测的关键技术,能够将视频感知内容表示为简短摘要。利用降噪自编码器(DAE)稳健性强的优点,通过逐层训练DAE构建独立提取各帧特征的深度网络,设计了一种基于时空神经网络的视频指纹算法。在此基础上,采用长短时记忆网络提取视频时序特征,并根据慢变特征分析理论设计了网络训练算法。实验结果表明:基于时空神经网络的视频指纹在视频拷贝检测中能够表现出较高的准确率,性能指标优于对比算法。

With the development of content-sharing networks, the on-line video data have grown dramatically and a large number of illegal copies have been appeared. To reduce any copyright infringement disputes, it is necessary to detect illegal copies on-line internet. Video fingerprint, which can express the video perceptual content as a compact description, is a key technology for copy detection. The video fingerprint algorithm based on spatio-temporal deep neural network is designed by the use of the excellent robustness of denoising auto-encoder (DAE) and building a deep neural network to extract features on frame level through greedily training DAE. Consequently, a long short-term memory network is adopted to extract each frame features of the deep network, and the training algorithm is designed on the basis of the theory of slow-feature analysis. Experimental results show that the proposed algorithm can reveal a high accuracy in video copy detection and outperform a number of the comparative algorithms.