轉換域運算於MPEG-2 to H.264 Transcoding上之應用

Abstract

影像訊號在儲存或者傳送至網路之前通常會經過壓縮處理。在許多應用當中，例如轉換編碼，直接對壓縮位元流進行處理的運算成本比經由完全解碼再對訊號進行處理的運算成本低。影像壓縮技術一般使用離散餘弦轉換將影像訊號從像素域轉換至轉換域。一些影像運算，如移動補償、樣本減量等運算在轉換域皆有其對應的實現方法。這些轉換域運算的推導乃基於離散餘弦轉換的正交性與運算本身的線性性質。然而，由於轉換域中的資料是以方塊為單位，一些像素域運算無法很直接的推導出其在轉換域中的等效運算。特別是對於H.264標準，由於其提供了許多新的編碼技術，與其相關的轉換域運算的研究有其需要性。 本論文研究的轉換域運算包括轉換核心變換、畫面內預測編碼、移動補償。這些運算皆能應用於轉換編碼技術之中。除了轉換域運算的推導，我們亦針對這些運算提出有效率的實現方法。此外，我們設計了一個MPEG-2 to H.264的轉換編碼器，並將其效能對串接式像素域之轉換編碼器做比較。 在此論文中，我們亦對轉換域中的進位運算做一研究。許多多媒體系統牽涉到進位運算。以移動補償運算為例，半像素位置的樣本需經由周圍像素內插並經過進位運算之後得到。然而，進位運算是個非線性運算，其轉換域上的等效運算並不存在。針對這個問題，我們提出一個解決方法，並將其實現在一MPEG-2解碼器上。相較於在轉換域系統裡忽略進位這個運算，運用我們的方法能夠獲得較好的視迅品質。

Usually, video signals are compressed before they are stored or transmitted over networks. In many applications, such as transcoding, manipulation of video signals directly on the compressed bit streams is more desirable than the approach where the compressed videos are fully decoded and then manipulated, because of the computational cost issue. For energy compaction purposes, video compression methods use techniques such as discrete cosine transform (DCT) to convert video signals from the pixel domain to the transform domain. Image operations such as motion compensation and down sampling have counterparts in the transform domain. The derivations of these transform-domain counterparts are based on the orthogonality of DCT and the linearity of these operations. However, since the data is organized block by block in the transform domain, the coding operations are not straightforward when applying to the transform-domain data. Therefore, there is a need to investigate the transform-domain operations, especially for H.264, which provides many advanced coding techniques. The thesis investigates several coding operations, including transform kernel conversion, intra prediction and motion compensation for H.264. They are expected to have applications in the transform domain transcoder. In the context, we derive the transform-domain equivalent operations for these coding operations and provide the efficient way to implement the algorithm for each transform-domain operation. A transform-domain MPEG-2 to H.264 transcoder is devised, and its performance is evaluated and compared to the pixel domain approach. In this thesis, we also formulate a general problem for rounding in the transform domain. The rounding operation is commonly performed in multimedia systems. Take half-pixel motion compensation as an example. The generation of new pixels at half-pixel locations requires a rounding operation for better approximation of the pixel values. However, rounding is a nonlinear operation; so its corresponding transform-domain operation does not exist. We propose an approximate solution to this problem and demonstrate that the proposed approach is better than other approaches when applied to an MPEG-2 decoder.