基于卷积神经网络的视频软广播

Convolutional Neural Networks based soft video broadcast

随着信息技术和互联网技术的不断发展,无线视频广播越来越受到人们的欢迎,成为流行的多媒体应用之一。然而,传统的数字编码和传输方法很难适应于向多个具有不同信道质量的用户同时发送视频的场景,通常会遭遇悬崖效应。近期,一种新颖的无线视频广播方法称为SoftCast被提出,其保存在信道中传输的信号与视频像素值之间所具有线性关系并利用有效的能量分配方法,使得视频重构质量随着信道噪声的增加而平缓下降。在本文中,提出了一种新型的无线视频广播方法,其利用深度卷积网络和基于图像组的稀疏表示模型,通过解码端估计的信道质量,优化视频的解码过程并减轻多种由信源编码和信道噪声造成的视觉失真。通过视频软传输技术,本文提出的方法具有出色的视频广播质量可伸缩性并避免了悬崖效应的发生,同时还能提供视觉友好的主客观重构质量。实验结果表明,本文提出的方法在视频广播场景下能够获得优于传统SoftCast最高1.2 dB的重建质量。

With the continuous development of information technology and Internet technology,video broadcasting is becoming more and more popular in wireless networks.However,the existing digital coding and transmission approaches can hardly accommodate users with diverse channel conditions,which is called the cliff effect.Recently,a novel video broadcasting method called SoftCast has been proposed.It achieves graceful degradation with increasing noise by making the magnitude of the transmitted signal proportional to the pixel value and using a novel power allocation scheme.This paper proposes a novel video broadcast method that exploits deep convolutional networks and group based sparse representation.They utilize the channel condition information generated from decoder to optimize the decoding process and reduce the various artifacts caused by source and channel coding.By utilizing soft video broadcast transmission,it achieves good broadcasting performance,avoids the cliff effect,and also can provide visually friendly subjective and objective reconstruction quality.The experimental results show that the proposed scheme provides better performance compared with the traditional SoftCast with up to 1.2 dB coding gain.