分布式视频编码中基于改进FCM聚类的相关噪声模型估计

Correlation noise modeling based on improved Fuzzy C-Means clustering in distributed video coding

目的在分布式视频编码中,为了更加准确地描述相关噪声残差子带的变化特性,提出一种基于改进的模糊C均值(FCM)聚类的模型估计方法。方法本文算法针对每一解码子带选取不同的特征矢量;利用改进的模糊C均值进行聚类;采用阈值控制法求取相应的模型参数;然后用重建子带更新下一解码子带的特征矢量,直到一帧中所有子带解码完成。针对模糊C均值对初始聚类中心的敏感性,采用随机生成隶属度矩阵的方法来缓解聚类陷入局部最优的问题。结果从实验效果和算法复杂度角度考虑,将残差样本聚为8类。实验结果表明,本文聚类算法可以更加准确地模拟帧内不同区域的不同信道噪声特性,对于运动越剧烈的序列效果越好,相对于子带级拉普拉斯估计,平均增益达1 dB。结论提出了一种新的相关噪声估计方法,针对不同的子带选取不同的特征矢量,并重建更新。实验结果表明,本文算法能更好地描述相关噪声特性,获得系统性能的提高。

Objective In Wyner-Ziv （WZ） Distributed Video Codec, in order to describe the change characteristics of the correlation noise residual sub-band more accurately, we propose a new correlation noise-modeling algorithm based on im- proved-FCM （Fuzzy C-Means） clustering. Method In our proposed method, for each decoded sub-band, the residual co- efficient eigenvectors are formed with the residual coefficients from adjacent sub-bands, and then they are clustered into dif- ferent categories by an improved fuzzy C-means clustering algorithm. In order to avoid the overflowing problem caused by having only a few samples in one category, a threshold method is adopted to estimate the correlation noise parameter, which is useful to help decode the corresponding sub-band. Then the reconstructed sub-band is used to update the next subband eigenvectors to obtain more accurate eigenvectors. All the sub-bands are decoded using the same process. Fuzzy C-means algorithm is sensitive to the initial clustering centers. A method that is to produce a random membership degree matrix be- fore iteration can solve this problem to some extent. Result Considering both the algorithm performance and complexity, the subband residual coefficients are clustered to eight classes. The experimental results show that this method can simulate accurately the different channel noise characteristics of different region in one frame. What＇s more, the more complex the video motion, the more obvious the performance superiority of the new algorithm. Compared with that of the subband level Laplace method, the average online rate-distortion performance can be improved up to 1 dB. Conclusion A new correlation noise model based on an improved fuzzy C-means clustering algorithm is proposed in this paper. The experi^lent results show that the rate-distortion performance based on the new algorithm is better than that of sub-band Laplacian solution and the Laplacian-Cauchy mixture model, and the more complex video motion, the more obvious performance gain for the new algorithm.