使用交疊區塊位移補償之強健適應遞歸解交錯處理

Abstract

電視已經由交錯掃描的裝置(例如 CRT)逐漸轉變為循序掃描的裝置(例如LCD TV, Plasma TV)，所以解交錯處理(de-interlacing)對於改善整體的視訊品質就變得越來越重要。位移補償解交錯處理(motion compensated de-interlacing)比其他的方法可以提供更佳的視訊品質，但是伴隨著而來極龐大的運算量。這篇論文提出一個方法來對付上述的問題。輸入影像區塊首先用強健位移偵測器(robust motion detector)來偵測是否為移動區塊(motion block)。位移偵測器(motion detector)使用經過雜訊消除的畫面差異值(field difference)來精緻的描繪出移動物體(motion object)的外形而且從而提供強健位移偵測(robust motion detection)。如果移動區塊(motion block)進一步被偵測為大區域複雜移動畫面的一部份，且因為人眼對複雜的移動影像沒有辦法清楚辨識，則將會用簡易的邊緣方向性解交錯處理(edge-directional de-interlacing)來處理這個區塊。其他的移動區塊(motion block) 則使用交疊式方塊位移補償(overlapped block motion compensation)來處理去降低估測錯誤(prediction error)。為了進一步降低估測錯誤和傳遞錯誤，採用適應性遞歸式位移補償(adaptive recursive motion compensation)去利用前一個解交錯輸出畫面緩衝存儲器(Previously De-interlaced Frame Buffer)儲存全靜態影像(the full still image)來做為最佳的參考畫面(reference frame)。從最後的實驗結果得知我們提出的方法比以前低複雜性的作品高出2dB。

With the TV changes from interlaced devices (e.g. CRT) to progressive ones (e.g. LCD TV and Plasma TV), de-interlacing becomes more and more important to improve the overall video quality. In which, motion compensated de-interlacing can provide better quality than others but with heavy computational complexity. This thesis proposes a robust adaptive recursive de-interlacing by overlapped block motion compensation to address above issues. The proposed method first use a robust motion detector to detect the input image block as a motion block or not. The motion detector uses the frame difference with noise reduction to finely describe the shape of motion object and thus provides robust motion detection. If the motion block is further detected as part of large-area complex motion image, this block will be processed by simple edge- directional de-interlacing since human eyes are less sensitive to complex motion images. Other motion blocks are processed by overlapped block motion compensation to reduce prediction errors. To further reduce the prediction and propagation errors, the adaptive recursive motion compensated de-interlacing is adopted by using recursive buffer to store the full still images as the best reference frames. The final experiment shows that the proposed method can achieve 2dB higher than previous works with lower complexity.