本論文首先提出一個綜合3DRS與 block-based 技術以及量測運動向量信賴度 (Confidence level) 的 True motion 演算法。此一方法可適用於各種不同型態的視訊 (如靜態、動態、縮放、推移等)。一般而言、將三度空間的真實運動以二維空間區塊平移來表示必然帶來各式失真,因此針對各個block估算的運動向量需做正確性評估,分析其可靠度等級 (Confidence level)。由於視訊影像空間及時間上的連續性,將可靠度較高的運動向量保留與傳遞使用。應用方面將真實運動向量套入動態補償概念,設計與實作一個Noise reduction演算法。實驗結果顯示,此方法可以達到降低噪訊的目的,同時保留影片原有的細節效果。
In this thesis, a novel true motion estimation method and its application on noise reduction for video sequences are proposed. This estimation method is based on the mixture of 3DRS and block-based searching, along with a confidence model of true motion. It can deal with various kinds of true motion (stationary, moving, camera zooming, panning, etc.) in versatile video sequences. Because of the inherent nature of distortion in projecting 3D motion to a series of 2D motion frames, we deliberately evaluate the reliability of each estimated vector so as to ascertain its fidelity. Once the reliability (or called the true motion confidence) secured, the motion vector having the better confidence level will be retained and propagated to the neighboring blocks in the recursive search procedure. Moreover, in this thesis, a compensation-based noise reduction algorithm for video sequences is proposed, which applies the above technique and other capabilities as well to demonstrate the fidelity performance. Experimental results show that this noise reduction algorithm cleanup the noises to an acceptable level while preserving the details in texture.