Research On Key Techniques For High Efficient Implementation Of Video Encoder

As the original video data is so voluminous for real applications, video coding has been a focus of research and applications for saving transport bandwidth and storage space. For now, many video coding standards have been established for different applications. In the latest, the new generation of video coding standards, such as H.264/AVC and AVS standards, have achieved significant improvement compared with previous video coding standards. Due to the continuous development of video coding techniques, the computational complexity of video coding algorithms is rapidly increasing, especially in high resolution applications. Thus, how to efficiently implement a high performance video encoding system becomes an important issue.In this dissertation, we focus on the techniques for efficient design and implementation of video encoder, and try to achieve both high encoding speed and high coding efficiency. To reach this goal, two key units of the encoder which is in-loop deblocking filter and motion estimation are carefully studied. A series of pipeline schemes and adaptive algorithms are proposed. And parallel encoding structures are also discussed in this paper. The detailed descriptions for these techniques are as follows.Firstly, the in-loop deblocking filter (loop filter for short) contains highly adaptive processing which inevitably appears in the inner most loops of the algorithm, so it is a challenge for parallel processing on a DSP platform. To resolve this problem, two pipelined solutions to the loop filter in H.264/AVC and AVS are respectively presented. As for the loop filter in H.264/AVC, the global filter control is improved at first. Then masks and conditional storing are used together in order to overcome complex conditional jumping in edge filter. Two symmetrical samples, conditions or masks are packed together for"pair processing"to increase the parallelism of pipelines. Two-pass pipelines are designed to overcome the multi-nest conditional branch in boundary strength decision. Moreover, two-level internal memory organization is presented as well to keep pipelining. As for the loop filter in AVS, the whole process is divided into several sub-processes at first, so that the global filter structure can be improved to achieve regular processing flow. Then pipelines are carefully designed for these sub-processes with elaborately allocating functional units, so better performance is achieved.Secondly, most of the video coding efficiency is derived from motion estimation, while motion estimation also contributes the heaviest computational burden to the encoder. Especially in high resolution video coding, large search range and variable block size are adopted to achieve high coding efficiency, and become the most serious bottlenecks in real-time applications. So index search technique is presented to resolve this problem. In this scheme, large search-range motion estimation is optimized by accelerating the global search, and the temporary replacement strategy is used to remove the correlations in variable block-size motion estimation while keeping almost the same coding performance. Then three highly parallel pipelines are designed based on DSP platform and bring a great time reduction without decreasing the search points. Simulated results also show that the proposed scheme is faster than other schemes implemented by fast search strategies. Index search technique provides a good solution to achieve both high coding efficiency and high encoding speed.Thirdly, search range control is an effective method to reduce the computational complexity of motion estimation. An adaptive search range strategy for B picture coding, ASR, is proposed. There are two parts in ASR, one is frame-level adaptive search range scaling algorithm, F-ASRS, and the other is macroblock-level adaptive search range algorithm, MB-ASR. Different from the idea of conventional search range control, the basic scaling rule of F-ASRS is derived from the linear relation between the motion feature in P and B pictures. And a group of adaptive thresholds obtained from intra mode and motion vector statistics are used to prejudge the motion degree of P pictures, so that the search range can be scaled at frame level. Then the motion information of adjacent blocks is used to adjust the search range at macroblock level. Through joint control on both frame and macroblock level, ASR can effectively avoid the redundant search in B picture coding.Fourthly, the parallel video encoding structure is analyzed. Parallel encoding with multi processing units is a basic way to increase the computational performance. Based on the hybrid coding framework, MB-level parallel encoding structure can perfectly support a video encoder, but its small-data communication frequency is extremely high. To overcome this disadvantage, GMB-level and SMB-level parallel encoding structure are proposed by recombining the macroblocks and rearranging their coding order. The communication frequency is greatly reduced while the coding time is almost the same. Furthermore, the effectiveness of MB-level parallel encoding is discussed to show how the coding flow, parallel structure, picture difference and platform factors affect the parallel video encoder. Finally, an example of a parallel video encoder product is presented which integrates all the techniques proposed in this paper.