快速分數像素之二元動量估測演算法

Abstract

隨著通訊與多媒體技術的發展，視訊應用變得越來越普遍。動量估測在視訊編碼中扮演一個很重要的角色，不僅因為編碼過程裡它佔了很多的運算時間，也因為它會直接影響到編碼效率。近代的視訊編碼標準多使用分數像素精確度的動量補償與雙向動量補償等預測技術以增進編碼效率，但付出的代價是動量估測中運算複雜度與所需記憶體頻寬的增加。因此，為了實現即時視訊通訊，快速的動量估測是必須的。 在本文中，我們發展了一個可提供分數像素精確度的快速的二元動量估測演算法。在以單一位元代表八個位元之像素資料與全二元金字塔形架構的使用下，動量可經由階層式搜尋得到，同時降低了運算複雜度與所需記憶體頻寬。本演算法實現在MPEG-4視訊編碼標準中P-VOP與B-VOP所需的動量估測部份。經由較佳的量化參數設定，實驗結果顯示，快速分數的像素之二元動量估測演算法加速了動量估測並且維持可接受的視覺品質。

Video applications become more and more popular with the developments of communications and multimedia technologies. Motion estimation that takes much computational time and directly affects the coding efficiency and visual quality of video signals plays an important role in video coding. Modern video coding standards use the techniques of motion compensated prediction at subpixel accuracy and bidirectional motion compensated prediction to improve the coding efficiency with drastically increasing computational complexity and required memory bandwidth in motion estimation. The faster motion estimation methods become essential for realizing real-time video communication.

In this thesis, we develop a fast binary motion estimation algorithm providing fractional pixel motion accuracy. With the pyramidal structure and one-bit representation of pixel data, motion vectors at the accuracy of half or quarter pixel are got via hierarchical search, which can reduce both the computational complexity and required memory bandwidth. The algorithm is realized in the MPEG-4 reference software and evaluated for fractional-pel motion estimation of both P-VOPs and B-VOPs in MPEG-4. With the better quantization parameter setup, experimental results show that the fast fractional pixel binary motion estimation can speed up the encoding process with an acceptable visual quality.