非迭代收缩多方向预测的空域错误隐藏算法

Non-iterative shrinkage multi-directional prediction for spatial error concealment

现有的空域错误隐藏算法通常以高复杂度的迭代逼近机制换取恢复质量的轻微提升,且一些算法只适合于特定的丢失模式.为此,本文提出了一种非迭代收缩多方向预测(Non-iterative Shrinkage Multi-directional,NSM)的空域错误隐藏算法,以便更好地平衡计算复杂度与恢复质量等性能指标,且能处理各种各样的丢失模式.对于当前受损块的错误隐藏,NSM算法首先通过各向同性梯度检测器学习当前延拓区域的梯度特征;随后,基于具有16邻域像素和8预测方向的基本隐藏单元,多方向预测器按照收缩填充次序逐一地恢复受损块的每一个像素,根据该像素的邻域像素可用情况调整预测器的加权系数.在每个受损块的内部,不同的像素组根据它们的邻域级可用度一组接一组地进行恢复;在一个像素组的内部,不同丢失像素根据先验填充准则一个接一个地进行预测,从而实现低复杂度的非迭代重构.相比于其他空域错误隐藏算法,实验结果表明,所提NSM算法在各种丢失模式下均能够取得良好的综合性能,在通用性、计算复杂度和恢复质量之间达到了一种具有竞争力的性能折衷.

At the cost of high complexity,the existing spatial error concealment algorithms may slightly improve the recovery quality by iterative approximation,and some algorithms are optimized only for certain loss patterns.To achieve a trade-off among performance indices,this paper proposes a non-iterative shrinkage multi-directional(NSM)prediction algorithm aiming at handling various types of loss patterns.For the error concealment of the current missing block,the proposed NSM algorithm firstly adopts an isotropic gradient detector to learn the local feature information of the current extrapolation region.Based on a basic concealment unit with 16-pixel neighbor and eight prediction directions,the multi-directional predictor recovers each pixel of the missing block one by one in a shrinkage filling order,and adjusts the weighting coefficients of the predictor according to the availability of adjacent pixels.During the shrinkage filling of a block,different pixel groups are recovered group by group according to their neighbor-level availabilities,and different missing pixels in a pixel group are predicted one by one according to a priori filling rule.Thus,non-iterative reconstruction with low complexity can be realized.Compared with other spatial error concealment algorithms,experimental results show that the proposed NSM algorithm achieved better overall reconstruction performance under various loss patterns,and realized a competitive performance trade-off among versatility,computational complexity,and recovery quality.