适于PCT编码的低内存通用树状多带滤波器组实现

Low-memory Implementation of Generic Tree-structured Filter Banks for PCT-based Image Coders

基于常用的多带滤波器组的紧支撑性及FIFO(First-in First-out)缓存技术提出了一种具有低内存需求的通用树状多带滤波器组(Tree-structured Filter Bank,TSFB)的实现方法SBFB。该方法具有两大特点:(1)生成与全局变换法相同的子带系数,但是内存需求大大减小且仅与图像宽度及采用的TSFB相关;(2)在FIFO缓存中直接生成父子树(Parent-children Tree,PCT)。基于PCT的编码器可直接对位于缓存中的PCT进行编码,而无须在SBFB与编码器之间引入中间缓存。通过分析TSFB的各分解层中的数据流,给出了样本点与子带系数之间的时序关系,并且使用该关系从数学上严格证明了SBFB的正确性。

Based on the compact support property of popular multi-channel filter banks, a low-memory implementation of generic tree-structured filter banks （TSFBs）, called the stripe-based tree-structured filter bank （SBFB）, is presented by utilizing the first-in first-out （FIFO） buffer technique. The SBFB has two advantages： （1） It produces exactly the same sub-band coefficients as the conventional global implementation of TSFB does, while the memory budget is unrelated to the image height and only dependent on the image width and the TSFB adopted. As a result, the memory budget of SBFB is only a fraction of that of the global implementation; （2） It generates integral parent-children trees （PCTs）, which are completely contained in its buffers. The SBFB lends itself to PCT-based coders in that no intermediate buffeting is needed between the SBFB and the coder. By analyzing the dataflow of different levels of TSFB, the relation between the samples and the sub-band coefficients is also attained, by which the correctness of the SBFB is strictly proved.