Motion Vector (MV) Search Window Prediction in Memory-Constrained Systems

羅仲成 (南加州大學電機所博士班)

 

http://spie.org/x30491.xml

 

動作估計是影像壓縮過程中，最消耗計算量的步驟。對於一張畫面中的Macroblock，家用電腦可使用相當大的搜尋視窗來尋找參考畫面中的動作向量。然而對於移動式裝置，可使用的快取以及處理器資源相當有限；因此為了避免快取失誤，適當的搜尋視窗大小必須列入考量。為了解決這個問題，我們使用統計方法分析每個Macroblock中搜尋視窗的位置、大小、以及形狀。我們並且基於系統對於畫質衰減的需求，估計一個適當的搜尋視窗位置、大小、及形狀。搜尋視窗的大小形狀以及畫質衰減之間的平衡也會在此討論。這些分析的結果可用來大幅提升再記憶體受限的移動式裝置中，動作估計的的效率。

 

除此之外，此一分析亦可運用在另外一種電腦架構──多核心處理器。在多核心處理器中，每個核心可運用的快取相當有限。除了受限的記憶體之外，為了最佳化多核心處理器的平行處理效能，許多Macroblock會同時在多核心中分散運算，這造成許多常用的空間關聯性在多核心系統中將無法繼續運用。所以，針對動作向量搜尋視窗的精確預測對於系統效能的提升有關鍵性的作用。

 

Motion estimation is the most computationally expensive task in video compression. For a given macroblock in the target frame, the search window of its motion vector (MV) in the reference frame can be quite large in the PC platform. In contrast, the available cache of the CPU is limited in mobile devices. Thus, it is important to fetch a proper search window of MV in the reference frame to avoid cache miss. To solve this problem, statistical analysis of MVs for search window location/size/shape prediction is conducted on each macroblock. A proper search window location, size and shape will be estimated while meeting a certain distortion requirement. Trade-off between the search window size/shape and distortion will be examined. These analytical results can speed up the motion estimation process in memory-constrained mobile devices.

 

There is another architecture that will benefit from the proposed technique, i.e. the multi-core processor. Besides a limited cache size for each core, many macroblocks will be executed simultaneously in multi-core processors to exploit parallelism to the maximum degree. As a result, some of the commonly-used spatial dependency may no longer be utilized as the motion estimation predictor. Thus, accurate prediction of MV search window is also critical to the system performance.