視訊編碼中使用移動向量預測與資料重複使用之移動估計

Abstract

時間預測為視訊編碼標準中最重要的模組。在本篇論文裡，我們將討論及分析在視訊編碼標準中時間預測方法裡的移動估計。討論分析裡包含了移動向量的預測以及資料重複使用策略。 在移動向量預測部份，利用時間上與空間上高度相關的影像區塊進行搜尋移動向量前的預測，是非常重要的時間預測步驟。我們採用支援最新視訊編碼標準中可變式區塊移動估計的方法，利用時間上影像資料的高度相關，與空間上周圍區塊的移動向量多重回歸平面來預測出移動向量。在QCIF (176 × 144) 相素、搜尋範圍為 [-32, 32）的規格下實現，能減少傳統時間預測法壓縮最多13％的時間，並同時減少整體壓縮時間最多13％。 另外，在移動估計中，最常被使用的資料重複使用策略為區塊層級之資料重複使用。在這方面的研究，我們打破全域區塊搜尋法的搜尋限制，發展出新的區塊層級資料重複使用策略，使其適用於快速區塊搜尋法。不同於以往的資料重複使用策略設計皆僅侷限於全域區塊搜尋法，新的區塊層級資料重複使用策略不但具有資料重複使用的效能，同時藉著調整參數可以提供多種不同記憶體頻寬大小的選擇。本篇論文所提及的移動向量預測之改善，恰能應用於快速區塊搜尋法的資料重複使用策略，進而改善傳統。在QCIF (176 × 144) 相素、搜尋範圍為 [-32, 32）的規格下，只需要傳統快速搜尋法約80％的頻寬，並同時減少內部記憶體75％。

The applications of digital video, such as image sensor, video conference, and networks, are coming more popular. Therefore, the development and improvement of video coding techniques are very important in these years. And video compression is made great progress, with continuing improvements in processing performance. Video compression is the key technique in various digital video devices. And the latest international video coding standard, H.264/AVC, which provides better coding efficiency than others, is widely adopted. To reduce the temporal redundancy of the digital video, motion estimation plays an important role in the coding process. However, there takes a lot of load in order to find out the matching motion vectors in motion estimation. Thus, we tend to speed up the coding time to present with the real-time application. In the H.264/AVC standard, motion vector prediction is used to speed up the calculation. That is because with choosing a wise start point, the high temporal redundancy between successive frames will be reduced. In many important coding standards, the prediction will find out by the median calculation. The previous median calculation will form the prediction error. But the motion vector prediction errors exist with the median calculation. For the more accurate statistics of the neighboring partitions, an improved algorithm to estimate the motion vector prediction is proposed. Besides, the proposed ideas are standard compatible for H.264/AVC. The multiple regression analysis for motion vector prediction is proposed. This algorithm is design with the spatial redundancy. And the proposed method is also calculated in the simple way. The simulation result shows that the coding time will be reduce by 5%~13% for all sequences. Furthermore, the PSNR performances compared to the median calculation gains 0.01~0.07 dB by using the proposed multiple regression analysis. For the complexity of image data accessing, conventional data reuse schemes were aimed to reduce the access times of motion estimation. Fast search block matching algorithms are established to speed up motion estimation. However, previous works were focus on the full search block matching algorithm for the regular accessing characteristics. We propose the data reuse schemes with the most commonly used fast search BMA. We use the overlap between every accessing blocks. Experiment result shows that the on-ship memory size and the memory bandwidth will both reduce compared to previous data reuse scheme, level C, and the original fast searching BMA. Proposed scheme is proved to reduce 75% of on-chip memory size is reduce, and save 80% memory access size.