適應性的自動回歸模型的增加畫面更新率轉換方法

Abstract

增加畫面更新率是視訊處理中眾多議題的其中之一。本篇論文提出了一種適應性的自動回歸模型，使其產生的畫面有更好的視覺品質及更少的計算負擔。在傳統的自動回歸模型中，每個像素被建模為時間上像素點或空間上像素點的線性組合。而在本論文中，我們提出了一個利用視訊資料的特性來選擇回歸模型的機制。選擇適當的回歸模型，可以在回歸運算當中減少不必要的變數，在計算複雜度上得到了相當程度的改善。實驗結果顯示出在運算時間上得到了顯著的進步，並且在內插出的畫面中，視覺效果也得到了改善。

An adaptive auto-regressive model is proposed in this thesis for frame rate up-conversion. In conventional AR model, each pixel in the to-be-interpolated frame is modeled as a linear combination of temporal neighborhood, spatial neighborhood, or joint temporal-spatial neighborhood pixels. This thesis proposed a temporal AR model (called TAR) utilizing temporal neighborhood; and a spatial AR model (called SAR) utilizing spatial neighborhood. Besides that this thesis also proposed a scheme which selects TAR or SAR adaptively according to motion information in the video sequence. By selecting appropriate AR model, unnecessary variables can be eliminated from regression process. Compared to STAR model [2] which utilizes joint temporal-spatial neighborhood for each pixel, computational cost can be greatly reduced with the proposed method. In addition, the experiment results show that visual quality can also be improved by adaptively adopting appropriate AR models for frame interpolation. The results demonstrate the superiority of the proposed method in regarding to improved visual quality and reduced computational cost.