摘要
动态场景的外形或表观在很大程度上往往受到一个潜在低维动态过程的控制。基于视频序列之间的时间相干特性,引入一种称为自编码(autoencoder)的特殊双向深层神经网络,采用CRBM(continuous restricted Boltzmann machine)的网络结构,用来学习序列图像的低维流形结构。将autoencoder用于人体步态序列的实验表明,该方法能提供从高维视频帧到具有一定物理意义过程的低维序列的映射,并能从低维描述中恢复高维图像序列。
The shape/appearance of dynamic scenes is often largely governed by a latent low-dimensional dynamic process.Based on the temporal coherence between video frames,this paper introduced a special bi-directional deep neural network called autoencoder which used CRBM to learn the low-dimensional manifold structure of image sequences.The experiments on the video sequences of human gait show that the algorithm not only can provide a map from the frames of video to a low-dimensional sequence which represents a phy...
出处
《计算机应用研究》
CSCD
北大核心
2009年第3期1183-1185,共3页
Application Research of Computers
关键词
视频序列
流形学习
自编码网络
降维
重构
video sequence
manifold learning
autoencoder network
dimensionality reduction
reconstruction
