JPEG2000編碼器之設計與實現

Abstract

JPEG2000標準中，最吸引人的莫過於它對於各階級影像解析度及影像品質的自由調整能力。其中數位小波轉換（DWT）與EBCOT尤其是兩大功臣，然而，它們也是整個系統當中最花費記憶體、最消耗運算力的部分。 為了減少記憶體需求，分別組合五種DWT運算順序與兩種DWT分解層級順序，來考量它們對緩衝存儲器（buffer）大小的影響。最後，我們選擇level-by-level與optimal-z san的運算順序組合，它不但減少DWT本身所需的buffer大小，同時也考量其與EBCOT之間的buffer。再者，將以stripe為基本單位的EBCOT運算順序運用到整個系統上，則能使DWT與EBCOT間的buffer再減少93.8%，而表現在整個系統上，則是總記憶體需求減少為原來的66%。然而，由於此方法會造成運算時間增加原來的14%，我們提出一個 “全零stripe略過方法”，預先跳過全零的位元平面不做，不但能補救此缺點，還能多減少0.22%的運算時間。 至於減少運算需求上，在DWT方面，兩個方向的一維DWT核心共用同一組加法及乘法器，這使得整個二維DWT的gate count減少三分之一。在EBCOT，則設計為pass-層級的平行處理，能達到傳統作法的三倍速度，並減少三分之二的記憶體存取次數。我們所設計的pass-層級平行的context formation架構，gate count約是其他人的6.8%。 最後提出一個JPEG2000編碼器的實作計畫，它使用一個DWT模組與三套embedded block coder。這個編碼器使用較小的面積與記憶體，能達到55.6Msamples/sec。

The ability to have scalability in resolution as well as image quality is the main attractiveness of JPEG2000. DWT (Discrete Wavelet Transform) and EBCOT (Embedded Block Coding with Optimal Truncation) which are two major technologies enable it, however, are also the parts that demand huge storage and computations. To reduce memory requirement, we combine five different computing orders of DWT with level-by-level or mixed-level and find that level-by-level optimal-z scan can reduce the temporal buffer in DWT as well as the buffer between DWT and EBCOT. We also adopt the new stripe-based computation order of EBCOT to further reduce 93.8% buffer size between DWT and EBCOT. The total buffer for the JPEG2000 encoder can be reduced to 66% of the original design. However, the stripe-based computing order will increase 14% more computation time. Thus, we proposed the zero-stripe skipping technique to skip the all-zero-bitplane. With this approach, we can eliminate this overhead and reduce 0.22% computation time further. To reduce the computation complexity, we share the multipliers and adders of the two directional DWT kernels, so that 1/3 of the area of DWT module can be saved. For EBCOT, a pass-level parallelism is adopted to speed up 3 times of the traditional processing time and to reduce 2/3 memory accesses. The gate count of proposed context formation is 6.8% of others. Finally, we proposed a plan to use one DWT module with three embedded block coders to integrate our JPEG2000 encoding system. It can achieve a throughput of 55.6 Msamples/sec at 100 MHz clock rate with lower cost and less memory requirement.