融合自适应阈值与αML核函数的双稀疏空域错误隐藏

Double-sparse spatial error concealment with adaptive threshold and α-ML kernel functions

关键词:  错误隐藏  自适应阈值  核函数  区域匹配  相关模型

Key words:  error concealment  adaptive threshold  kernel function  region matching  correlation model

摘要:

在实时视频流的解码端,恢复压缩视频时经常出现块丢失现象。空域错误隐藏利用在当前帧中块与块之间的相关性进行受损图像的恢复,无需其他帧的信息。在众多空域错误隐藏算法中,稀疏表达机制进一步利用了图像的稀疏性,与逐点插值机制相比恢复质量更好。当前的稀疏表达算法仍面临候选子区域选取不准、相关模型参数较敏感的难题。为此,对基于稀疏表达的对偶空间正则框架进行研究,重点对其中的局部区域匹配和局部线性相关建模两个阶段进行优化,提出一种融合自适应阈值与α-ML核函数的双稀疏空域错误隐藏算法。在局部区域匹配阶段,所提算法设计了一种基于自适应阈值的局部区域匹配方法,能够灵活地适应特征各异的丢失子区域,为字典构建和局部线性相关建模提供更准确的观测空间和潜在空间。在局部线性相关建模阶段,所提算法使用了一种基于α-ML核函数的核岭回归方法作为局部线性相关模型, 与现有的高斯核函数相比,α-ML核函数参数敏感性更低、灵活性更好。实验结果表明,在典型的块丢失模式下,所提算法在恢复质量上高于其他现有的空域错误隐藏算法。

Abstract:

In the decoder of live video streaming, compressed video often suffers from block loss during recovery. The spatial error concealment utilizes the correlation between blocks in the current frame for the recovery of error image, without requiring information from other frames. Among many spatial error concealment algorithms, sparse representation mechanism further utilizes the sparsity of an image and achieve better recovery quality than pixel-wise interpolation mechanism. The current sparse representation algorithms still face challenges such as inaccurate selection of candidate subregions and parameter sensitivity of correlation model. Therefore, this paper studies the dual space regularization framework according to sparse representation, and focuses on optimizing such two stages as local region matching and local linear correlation modeling in this framework. We proposes a double-sparse spatial error concealment algorithm with adaptive threshold and α-ML kernel function. During the stage of local region matching, the proposed algorithm designs a local region matching method with adaptive threshold, which can flexibly adapt to the missing subregions with different characteristics, and provide more accurate observation space and potential space for dictionary construction and local linear correlation modeling. During the stage of local linear correlation modeling, the proposed algorithm utilizes a kernel ridge regression method with α-ML kernel function as the local linear correlation model. Compared with the Gaussian kernel function, the α-ML kernel function has low parameter sensitivity and good flexibility. Experimental results show that in typical block loss modes, the proposed algorithm outperforms other existing spatial error concealment algorithms in terms of recovery quality.