頻域成分位移估測與變形網目視訊編碼技術之研究

Abstract

物體位移估測(motion estimation)已廣泛地使用在現有的視訊處理系統 上，目前最常用的方法有區塊比對法(block matching method)及微分法( differential method)。在本論文的第一部分中，我們利用頻域成分的分 析方式，對物體位移估測的理論做定性與定量的探討。除了研究影響估測 的兩大原因－準確性(accuracy)與混淆性(ambiguity)二者之間的相互關 係以外，並推導不同估測方法在理論上的極限。最後藉由此分析設計一種 新的頻域成分估測演算法(frequency component algorithm)，並針對現 有的演算法提出可行的改良之道。在傳統的位移補償式壓縮編碼(motion- compensated coding)系統中，通常利用規則的區塊做為編碼的基本單位 。由於影像內容的時空結構以及人類視覺系統的特徵均未列入考慮，使得 區塊式位移補償受到這些限制而降低其效率和品質。此外，這類的編碼架 構亦無法滿足目前視訊壓縮編碼的新趨勢－物體萃取(object scalability)。因此，在本論文的第二部分□，我們提出了若干新技術適 用於以影像形變(image warping)位移補償為基礎的變形網目視訊編碼( deformable mesh video coding)系統中。這些新技術包括：前景╱背景 分離術(foreground/background separation)，經過重疊區域消除( overlapped patch elimination)處理之參考點追蹤術(nodal point tracking)以及非矩形離散餘弦編碼法(nonrectangular DCT coding)。模 擬結果顯示採用這些技術的視訊壓縮架構十分適合應用在低速率之物體萃 取式編碼上。 Motion estimation techniques are widely used in today's video processing systems. The most frequently used techniques are the block matching and the differential method. In the first part of this thesis, we have studied this topic form a viewpoint different from the above to explore the fundamental limits and tradeoffs in image motion estimation. The underlying principles behind two conflict requirements in motion estimation, accuracy and ambiguity, become clear when they are analyzed using this tool --- frequency component analysis. This analysis also suggests new motion estimation algorithm and ways to improve the existing algorithms. Ths so-called frequency component motion estimation algorithm is thus proposed. Comparing to the conventional block matching and phase correlation algorithms, this approach provides more reliable displacement estimates particularly for the noisy pictures. In a conventional motion- compensated coding scheme, images are often partitioned into artificial units such as rectangular blocks. Without considering the natural spatio-temporal structure of images and the characteristics of human visual system, the performance of block based motion compensation is often limited by this fundamental restriction. In addition, this coding structure does not match the object scalability requirement which is a new trend in video coding. Thus, in the second part of this thesis, we develop several techniques to form a new deformable mesh video coding algorithm. These techniques include foreground/ background separation, nodal point tracking with overlapped patch elimination, and nonrectangular DCT coding schemes. The simulation results indicate that this coding structure is quite suitable for object scalable coding at low bit rates.