基于树形原子字典的匹配跟踪算法

Matching Pursuit Based on Tree-Structured Dictionary

为了降低匹配跟踪视频编码算法的运算复杂度,提出了一种原子字典的树形组织方法,并相应地提出了一种改进的树形字典搜索策略。此算法通过在树中寻找最优路径来搜索最佳原子,并在第一级搜索中引入了一种改进的全搜索策略,进一步降低了计算量。为了有效地捕捉预测误差帧中的曲线特征,冗余原子字典通过对各向异性的生成函数进行真正二维意义上的几何变换来构造,并尽可能地去除了性能不佳的原于以减小字典尺寸。实验结果表明,本文算法具有较高的搜索正确性及较快的收敛速度,而其计算量明显降低。

To reduce the complexity of the matching pursuit used in video coders, dictionary of atoms is organized in a hierarchical tree structure, and a new tree-based pursuit method is proposed. It finds the best path through the tree to get the optimal atom, and an improved full searching scheme is introduced in the first-level searching. In order to capture curves in the inter-frame predicted differences,redundant dictionary is built by applying true 2-D geometric transformations to the anisotrupic generating function. Many pathological atoms are removed to decrease the size of the dictionary. Experimental results show that this method efficiently decreases the complexity of the signal decomposition, while the resulting approximation is kept quite satisfactory.