四叉樹結構變動大小自動回歸模型的增加畫面更新率轉換方法

Abstract

在視訊壓縮處理的眾多議題之中，增加畫面的更新率是經常被拿出來探討的一項。本篇論文提出了一種以四叉樹結構調整窗格大小的自動回歸模型來增加畫面更新率轉換的方法。在傳統的自動回歸模型裡，在同一影片中使用的訓練窗格均採取統一固定的大小，因此，有較差的插補結果。在本論文中，我們提出了以自動回歸模型去萃取出最適合用來插補畫面的權重，並且預先在壓縮端針對各別序列計算出其在自動回歸模型中所需最佳的視窗範圍大小，編碼成為二進制文件，傳送至解壓縮端待命。當解壓縮端的自動回歸模型運作時，便可以取用精準的訓練窗格大小，來獲得更完善的插補畫面。實驗結果顯示出縱然運算時間增加了，但內插出的畫面中，視覺效果也得到了改善，兩者相權衡之下依舊是值得的。

In numerous issues for video compression processing, a frame rate up-conversion is discussed all the time. An auto-regressive model with variable size and quad-tree structure training windows is presented in this thesis for frame rate up-conversion. In conventional auto-regressive model, fixed training window size is adopted within a video sequence. Therefore, it may result in worse interpolated frames. This thesis proposed a scheme which selects the best training window sizes in the encoder side, encodes the best training window sizes to binary file and transfers the binary file to the decoder side. When the decoder side performs frame-rate up conversion, this binary file provides better training window sizes to improve the quality of the interpolated frame. The experiments show that, with variable training window size, auto-regression model can achieve better interpolated frame quality thought it has more computational time in the encoder side.