Network-aware perceptual error concealment method for H.264 video with side information

In order to improve the video quality of transmission with data loss,a spatial and temporal error concealment method was proposed,which considered both the state information of the network and the perceptual weight of the video content.The proposed method dynamically changed the reliability weight of the neighboring macroblock,which was used to conceal the lost macroblocks according to the packet loss rate of the current channel state.The perceptual weight map was utilized as side information to do weighted pixel interpolation and side-match based motion compensation for spatial and temporal error concealment,respectively.And the perceptual weight of the neighboring macroblocks was adaptively modified according to the perceptual weight of the lost macroblocks.Compared with the method used in H.264 joint model,experiment results show that the proposed method performs well both in subjective video quality and objective video quality,and increases the average peak signal-to-noise ratio(PSNR) of the whole frame by about 0.4 dB when the video bitstreams are transmitted with packets loss.

变换域重叠隐蔽通信系统的隐蔽性度量方法

提出了一种基于频域加权的均方误差(MSE)感知失真度量方法以评价变换域重叠隐蔽通信系统的隐蔽性能。该方法针对Q因子度量方法应用于变换域重叠隐蔽通信的局限,根据变换域通信系统的基本原理,将信号的感知失真从时域转到频域进行度量,根据不同宿主信号频谱幅值对于叠加的隐蔽信号的感知失真程度不同,以宿主信号频谱幅值的倒数为加权因子,分别对宿主信号各频点的感知失真做加权求和,从而得到系统的感知失真度量结果。仿真结果表明,该方法不仅可以提高系统感知失真能力,与重叠隐蔽通信的基本机理保持一致,符合实际感知效果,而且还统一了系统感知能力与抗干扰能力,可以为隐蔽系统的设计提供有效参考。

基于感知相关代价函数的深度学习语音增强

当前大多数基于深度神经网络(DNN)的语音增强算法使用的均方误差代价函数没有充分利用人耳听觉感知特性,且语音可懂度没有必要关联性,为此提出一种端到端的基于DNN的语音增强框架,并将基于频域加权分段信噪比的感知相关代价函数作为优化目标来训练DNN;在此基础上,将频域加权分段信噪比和感知加权均方误差相结合,提出一种联合优化代价函数用于训练DNN,改善人耳对含噪语音的听觉感知。实验结果表明,采用深度神经网络对含噪语音进行去噪时,通过将感知相关代价函数整合到数据驱动的模型学习中,语音质量和语音可懂度显著提高。

基于感知加权线谱对距离的最小生成误差语音合成模型训练方法

提出一种基于感知加权线谱对(Line Spectral Pair,LSP)距离的最小生成误差(Minimum Generation Error,MGE)模型训练方法,用以改善基于隐马尔科夫模型的参数语音合成系统性能.在采用线谱对参数表征语音频谱特征时,传统MGE训练中使用的欧氏距离生成误差计算方法并不能较好地反映生成频谱与自然频谱之间的真实距离,而采用与谱参数无关的对数谱间距(Log Spectral Distortion,LSD)定义的生成误差函数可改善这一问题,但改进后主观效果不明显,且运算复杂度很高.文中先提出基于加权LSP距离的MGE模型训练方法,并在实验中从主客观对比不同加权方法以及基于LSD的MGE训练.最后,找到一种感知加权方法,不但具有较好的主观表现,而且在运算复杂度上与传统MGE训练相比几乎没有增加.