|Block-based motion estimation algorithms for video coding applications
|Hong Kong Polytechnic University -- Dissertations.
Video compression -- Standards.
|Department of Electronic and Information Engineering
|xiv, 103 leaves : ill. (some col.) ; 30 cm.
|Block-based motion estimation (ME) is one of the most important parts of video coding process. Over recent years, variable block size motion estimation (VBS-ME) and multiple reference frames motion estimation (MRF-ME) techniques have been included as parts of the H.264/AVC standard to improve the performance of ME. Although they can produce higher video quality as compared with the traditional fixed block size and single reference frame ME, applying them will suffer from heavy computational load so it is necessary to look for fast implementation approaches. There are already many fast algorithms of VBS-ME and MRF-ME developed, but nearly all of them are lossy methods besides the SAD reuse algorithm which is a lossless method of VBS-ME. It is always desirable that some fast algorithms can be developed to give the optimal searching results. To achieve this purpose, we propose respectively a fast approach of VBS-ME and a fast approach of MRF-ME, and both of them can provide exactly the same video quality as compared with the original exhaustive approach. The proposed VBS-ME makes use of the row-based partial distortion search (PDS) with special scanning orders of the rows of pixels inside the blocks for different block sizes. The partially computed sums of absolute differences (PSAD) of the searched candidate blocks are reused to pre-compute the sum-of-absolute-difference (SAD) values for the unsearched candidates with other block sizes. Therefore, the number of operations can be reduced. In addition, it is found that the locations in the searching region with similar extent of pixel error usually appear in clusters. Since the early rejection of non-best matches in the PDS becomes faster if the pixel errors with larger values are considered in the highest priority for the calculation of distortion values, the special scanning orders are arranged in a dispersed way so that even though a block matching process starts at a low error region, it will not keep processing at that region and will jump to other locations. Our proposed MRF-ME is different from the original approach which will find the best match in each reference frame firstly, and then select the optimal one among those matches. Instead, the location of the predicted block will be found within just one stage. This is achieved by allowing the comparisons between the costs of candidate blocks in one reference frame with those in other reference frames while searching for the best candidate block. By using this method, the tentative minimum value of the costs of searched candidate blocks will be stored and will keep on being updated when the searching process scans through all the candidates in different reference frames. Therefore, once the best match in a certain reference frame which has the global minimum cost among all the candidates in different frames has scanned, all remaining unsearched candidates including those in other reference frames can be rejected without fully computing their costs. To perform a full ME, usually, the motion search is bounded within a search window. Square window is a traditional type of search window and is popularly used. However, it was found after conducting our investigation that nearly all of the motions in video sequences are contained within a star-shaped range but not a square range. This means that it is inefficient to search for motions from positions near the four corners of the square window. We propose a star-shaped search window with about 1/4 of the size of a square window. Using the star-shaped window can largely reduce the number of operations, and also the video quality will not be downgraded. It is because those motion vectors corresponding to the positions near the corners of a square window are too far away from the center, and coding them requires a large amount of bits. That is why excluding those motion vectors by using the star-shaped window instead can still maintain the same level of video quality. On the other hand, when the PDS is applied, usually a spiral search pattern is used, but it is also inadequate according to the star-shaped motion distribution. A star-shaped search pattern is proposed that allows searching the candidate positions along a star-shaped path from the center of the search window to positions far away. According to experimental results, the average coding speed can be increased by using the star-shaped search pattern.
|All rights reserved
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item: