應用於MPEG-4標準協定之高效率視訊物件擷取技術

Abstract

本論文主要提出一個新的視訊物件分割演算技術，應用於MPEG-4標準協定中。這個視訊物件分割演算技術主要分為兩部分：初始前景/背景建立與移動物件追蹤。在第一個部分，首先使用多次的改變偵測(Change detection)步驟建立出可信賴的初始前景/背景資訊。接著針對上述步驟所產生的未遮蔽背景區域，利用每個像素位置在每次改變偵測過程中累積的變化資訊，找出並消除屬於未遮蔽背景的區域，改進了只參考最初與最後移動物體位置資訊所造成判斷上的缺失。在第二個部分，我們使用邊緣運算器追蹤移動物件的邊緣，並透過改變偵測和背景預測的方法移除未遮蔽背景的邊緣，進而擷取出完整的移動物件。實驗證明在改進過未遮蔽背景消除方法之後，我們可以較有效率地偵測出未遮蔽背景區域，進而更精確地分割出移動物件與背景。 之後我們將所提出的視訊物件分割演算法加入現今的MPEG-4標準編/解碼器中，進而建立出以物件輪廓為基礎的編/解碼架構。在編碼器(Encoder)架構部分，首先將影像利用所提出的視訊分割技術分離出移動的視訊物件與不動的背景，接著在編碼過程中只編碼一次最初的背景資訊，不需要如以往每張編碼影像都帶有背景資訊；在解碼器(Decoder)架構部分再將背景資訊填補到每個解碼影像中，由於在填補過程中會造成重複區域(Overlap region)與無資訊區域(Non-information region)的產生，故我們分別使用前景優先性(Foreground priority)與背景補償(Background compensation)方式消除上述的問題。實驗證明當影片內移動物體移動程度越大時，所提出的方式與標準MPEG-4編/解碼架構比較，更能夠有效地減少所需傳輸的資訊量，即能夠提高影片壓縮率。而在解碼影像畫質方面，與使用傳統的解碼方式比較也能維持在一定的水準。

In this thesis, first we propose a new method to eliminate the uncovered-background regions in the proposed video object segmentation algorithm. The video object segmentation algorithm mainly includes two parts: initial foreground/background construction and moving object tracking. In the first part, we construct the reliable initial foreground/background by using several times of change detection step. For the uncovered-background regions, which are produced in the previous procedure. According to the collected change information of each pixel during change detection steps, we can resolve the lacks of previous proposed method, which only makes use of the initially and finally located information of moving object as the reference to detect the uncovered-background region. Finally the initial foreground/background is updated. In the secondary part, we use edge operator to extract the edge of the moving object, the uncovered-background edge can be eliminated by using the change detection and background prediction methods. Finally the complete moving object region can be extracted. The experimental results show that the updated method of uncovered-background elimination can detect these regions more efficiently and the moving object and background are separated more precisely. Next, we combine the segmentation algorithm into current MPEG-4 standard to build an object-shape based coding architecture. In the encoder architecture, firstly the segmentation algorithm separates the encoding frame into moving object and monotonous background. Then, the background information is only encoded once in the first frame of the encoding procedure, and the following encoding frames will only include the moving object information. Therefore, identical background information will not be repeatedly included in each encoded frame. In the decoder architecture, the background addition step is used to add the background information of first decoding frame into each following decoding frames. Then, we use the foreground priority and background compensation methods to resolve the overlap regions and non-information regions problems respectively during the background addition step. The experimental results show that as the moving objects of video contents change more obviously, the proposed coding architecture will provide better compression ratio compared with the traditional MPEG-4 coding architecture, and the qualities of encoded images can also get good performance.