에러 표면 모델 기반 부화소 움직임 예측 알고리듬

Abstract

There have been many significant advances in video compression and communication methods recently. Thanks to these innovative advancements, we are now able to deliver high-quality video at low bit rates. To be more specific, these state-of-art technologies enable many networked multimedia to become a reality. Some of the examples are digital TV, video over the internet protocol, wireless and mobile video, third generation (3G) cell phones with video capabilities, and future generation of internet protocol TV (IPTV). H.264/AVC (Advanced Video Coding) offers a solution to high compression technology. It can reduce up to 50% bit-rate compared to previous MPEG-4 simple profile at the same video quality level. However, a larger amount of computation is required. Profiling report shows that ME (5 reference frames, ±16 search range) consumes more than 90% of the encoding time. There have been many fast algorithms and efficient architectures for integer motion estimation (IME). But, relatively few algorithms have been proposed for sub-pixel ME, which accounts for about 20% of ME time (in full search algorithms). Half pixels (a.k.a. in-between pixels) are obtained by interpolating integer pixels, which increases computational complexity. There are half-pixel accuracy searching methods [38-43] to decrease complexity caused by interpolation, but they commonly yield large errors due to oversimplified MC-error models. In this dissertation, sub-pixel ME algorithms based on mathematical models are proposed. Unlike conventional hierarchical ME techniques, the mathematical modeling methods avoid sub-pixel interpolation and subsequent secondary search after the integer-precision ME. In order to decide the coefficients of the models, the MC prediction errors of the neighboring pixels around the integer-pixel MV are utilized. Three types of sub-pixel modeling algorithms are presented in this dissertation, ones that decrease the amount of computation with minimal loss of image quality compared to the conventional sub-pixel ME. Therefore, this dissertation consists of three proposed parts: model-based quarter-pixel motion estimation (MBQME [71]), hierarchical model-based quarter-pixel motion estimation (HMBQME [75]), and surface-considered modeling algorithm for sub-pixel motion estimation [78]. In MBQME algorithm, the conventional half-pixel mathematical modeling algorithm is expanded to quarter-pixel mathematical model. This algorithm presents an interpolation-free quarter-pixel ME method based on a mathematical model. In this method, only three shift operators and four compare operators are required in each horizontal and vertical direction to find a quarter-pixel resolution motion vector. The proposed method does not have to classify a half-pixel resolution mode or a quarter-pixel resolution mode in the encoding process because both resolution modes are integrated within the proposed method. In the HMBQME process, the limitation of previous MBQME process is solved by hierarchical modeling method. The previous MBQME algorithm cannot find the quarter-pixel points outside of -1/2 and 1/2 (outer-quarter-pixel region). To solve this limitation, a hierarchical flow model is designed, and extended-MBHME algorithm is also defined. Hierarchical flow model is selective interpolation method according to MBQME method. Through the MBQME algorithm, a candidate position of the quarter-pixel motion vector can be found. And the extended-MBHME model is the extended mathematical model for applying at left or right position of the center position. From the proposed HMBQME method, both the computational efficiency and the quarter-pixel full-search comparable accuracy can be accomplished. This stems from the intelligent combination of two advantages: the mathematical model of MBQME and the adequate quarter-pixel search strategy. And finally, this dissertation shows that the mathematical models must base on error surface tendency. If the error surface is close to the convex function, previous modeling algorithms are sufficient. However, if the error surface is irregular, more accurate modeling and further interpolations are needed. Therefore in surface-considered modeling, the classification process of the error surface as plain or textured (complex) is needed. The proposed method analyzes the characteristics of the sub-pixel ME error surface and apply case-by-case modeling. Therefore, sub-pixel ME algorithms based on error surface models are presented at the end of this dissertation. For the experiments, all proposed algorithms have been tested on various sequences. Experimental results show that the proposed error surface-considered modeling provides clearly better performance in searching quarter-pixel MV than previous algorithms with negligible PSNR loss.