利用方向小波轉換分析進行視訊編碼之位元分配

Abstract

本論文主旨在於探討人類視覺特性於視訊編碼中之位元配置方式，並根據論文中提出的方法利用方向小波轉換分析視訊影片中的原始圖像及殘值圖像的結構性特徵，且依此分析結果設計一個位元分配的策略，使得在視覺上較為重要的區域可得到較多的位元，以期達到較佳的視覺品質。本論文中使用方向小波轉換分析的主因為它可在不同方向上做子頻帶分解，所以相較於離散餘弦轉換或其他基於可分離轉換的小波轉換，方向小波轉換能顯現出更多的結構性資訊。論文中提出的位元分配的策略嘗試在非結構性區域節省位元及在結構性區域增加視覺品質。在MPEG-4簡易版類別編碼器上的實驗結果顯示，提案方法在所有測試案例中皆有較好的表現，能在人眼視覺較為重視的區域得到較佳的影像品質。

This paper proposes a video bit allocation scheme based on Curvelet domain analysis. The proposed algorithm analyzes the structural characteristics of the intensity and motion-compensated residual images of a video sequence in curvelet domain to determine a bit-allocation policy so that visually important regions will be allocated with more bits. Curvelet transform is adopted in this thesis for such analysis because it performs sub-band decomposition in various directions so that more structure information is revealed in curvelet domain than in DCT or other wavelet domains based on separable transforms. The proposed bit-allocation policy tries to save bits in unstructured regions and increase quality in structured regions. Experiments using standard test sequences coded with an MPEG-4 simple profile video encoder show that the proposed bit allocation method has better performance (achieves higher PSNR’s) in the regions most human observers care about in all test cases.