基于深度学习的无局部结构矩阵序列预测

Ordinary Matrix Sequence Prediction Based on Deep Learning

由于对图像局部近邻结构的敏感性,卷积神经网络在计算机视觉领域有着广泛深远的应用.然而对于无局部结构的一般性矩阵,由于无法处理其中的大量全局依赖,卷积神经网络通常对其不够有效.鉴于此,提出了一种基于深度学习的网络结构,用来解决这种无局部结构的一般性矩阵序列预测问题.通过采用提出的对偶线性连接结构,使网络学习到了广泛存在于无局部结构矩阵序列中的全局依赖,从而得到了更好的预测结果.将提出算法与几种主流视频预测算法进行对比,实验结果表明:该算法对于无局部结构矩阵序列具有很好的特征提取能力,可以有效地学习到给定输入数据中的全局依赖,从而生成更好的预测结果.

Because of the sensitivity to the local neighbor structure of image,convolutional neural network has a wide and profound application in the field of computer vision.However,for the ordinary matrix processing without local structure,convolutional neural network is usually not effective for it.For this specific situation,a network framework based on a novel dual linearized connecting is proposed to deal with the ordinary matrix sequence prediction task.By processing of ordinary matrix sequence through dual linearized connecting,the network owns the ability to learn inherent feature and global dependence from data,and better prediction results are obtained.The experimental results show that the proposed algorithm owns robust feature extraction ability for the ordinary matrix sequence without local structure,and can generate better prediction results than mainstream video prediction methods.