快速動量量測演算法之研究

Abstract

近年來隨著軟硬體以及網路技術的進步，多媒體的應用亦突飛猛進。諸如聲音、靜態影像以及動態影像等等的處理、存取、傳輸乃至於即時傳輸逐一的在現實生活中得以實現。其中聲音及靜態影像在網路上的處理與傳輸技術已經趨於成熟，唯動態影像的即時壓縮以及傳輸尚不能突破硬體的限制。動態影像的壓縮技術中，動量量測扮演了很重要的角色。它不僅佔了大部分的壓縮時間，同時量測的精準度也直接影響到壓縮的品質。 本篇論文著眼於提出一個既快速又能維持量測精準度的演算法。從結合了階層搜尋與選擇性預測的動量量測演算法開始，逐漸地減少運算的複雜度以及資料頻寬，之後便達成一個以單一位元表示八位元寬資料的快速階層性動量量測演算法雛形。緊接著我們試著將此演算法對C語言做最佳化處理，讓此演算法的速度更加的提昇以達到實際應用的價值。待執行速度的問題解決之後，我們進一步測試各種不同將八位元寬的資料單一位元化的方法，並證明了這些方法對於單一位元動量量測演算法的壓縮品質有著絕對的影響。基於上述的實驗結果，最後我們提出一個適應性的快速動量量測演算法。此演算法除了結合了階層搜尋以及單一位元化的特性外，系統設計者更可以依照該系統的特性來調整所需的複雜度。

Recently multimedia applications flourished with the progress of software, hardware and network. Processing and transmission of audio, still images and videos over networks have been realized in our daily life. Among them the audio and still image applications are nearly mature, however the real-time video compression and transmission is still limited by the hardware capability. In the video compression, motion estimation plays an essential role. It not only occupies lots of computation time but also influences the encoding efficiency and visual quality of the encoded bitstream. In this thesis, we focus on a robust and fast motion estimation algorithm. First of all we start with combining the hierarchical motion search and selective pre- diction of motion vector field. Then we try to reduce the complexity and bit-depth. Later the fast motion estimation based on hierarchical motion search and binary representation of 8-bit wide data is developed and we spend lots of efforts to optimize the algorithm. After the problem of speed is solved, we further attempt to improve the binarization efficiency by testing various binarization methods. Finally we will show that binarization methods are closely coupled to the accuracy of binary motion estimation algorithms. All in all we provide a flexible solution of the fast motion estimation based on binary pyramid.