Research On Key Techniques For H.264

In order to meet the increasing demands on video applications, International Telecommunication Union-Telecommunication (ITU-T) and International Organization for Standardization (ISO) have published a series of video coding standards in recent 20 years, such as H.261, MPEG-1, MPEG-2, and H.264. Due to excellent performance and network-friendly design, the latest video coding standard H.264 has made a great success and occupied a leading position in video applications. This thesis mainly focuses on the key technologies of H.264, including rate control, mode decision and multiple reference frames motion estimation (MRF-ME).Rate control, the target of which is to maximize the coding quality at a certain channel bandwidth and buffer constrains, plays an important role in the applications of video coding standards. Compared with previous coding standards, rate control is more difficult to implement in H.264 due to its high coding complexity. To overcome the drawbacks of the H.264 rate control algorithm recommended by Joint Video Team (JVT), this thesis proposed several optimizing algorithms as follows:An initial QP selection algorithm is presented. The computing formula of the best initial QP for sequences with different complexity at different target bit rate is obtained by fitting extensive experimental results. The algorithm has low complexity and is simple to use. High performance of this algorithm is proved by experiment.Firstly, an accurate Intra-Frame R-Q model is proposed; Secondly, based on the proposed R-Q model, a rate control scheme for I Frame is presented; because most macroblocks (MBs) in abrupt scene change frame will be encoded using intra mode, the proposed R-Q model is utilized with a bits allocation scheme for scene change frame to largely reduce the fluctuation of both Peak Signal to Noise Ratio (PSNR) and output bits.When rate control is applied to P frames, the algorithm recommended by JVT has many drawbacks:the linear Mean Absolute Difference (MAD) prediction model is not accurate enough, so is the updating procedure of quadratic R-Q model. In order to overcome these drawbacks, an improved algorithm based on sobel edge operator is proposed. Utilizing the Sobel edge operator, the prediction accuracy of MAD can be improved; meanwhile, a more suitable R-Q model history data point selection method is also developed to enhance the accuracy of R-Q model. Experiment demonstrates that compared the original algorithm, the proposed one can achieve high PSNR increment, while the rate control result is more accurate.The coding structure of hierarchal B frames is introduced to improve coding efficiency. However, the existing rate control algorithm can not support this coding structure well. A rate control algorithm for hierarchal B frames is proposed in this thesis. The major idea is to allocate more bits for the frames at lower coding level to guarantee the entire coding efficiency. Compared with other algorithms for hierarchal B frames, the proposed one can get higher control accuracy and better coding performance.Plenty of coding modes can guarantee high coding efficiency for H.264, but this technique induces high coding complexity. Aiming to reduce the computational load of encoder, a fast intra mode decision algorithm and a fast mode decision algorithm for P frames are proposed:In this algorithm, the suitable Intra coding mode (Intra₄×4 or Intra₁6×16) is selected according to the complexity of current MB. If the suitable coding mode is Intra₄×4, the coding information of neighboring blocks is applied to speed up the selection of Intra₄×4 prediction modes; otherwise the complexity information is reused to simplify the selection of Intra₁6×6 prediction modes. Experiment shows that the proposed algorithm can efficiently reduce the computational complexity, only yielding little loss in coding efficiency.In this algorithm, if current coding MB is temporal stationary, only large block size modes (Inter₁6×16 and SKIP) are checked, otherwise spatial correlation and temporal correlation are used to predict the best mode of current MB. If certain condition is satisfied, Intra mode selection can be skipped; otherwise the Intra mode decision algorithm proposed before is launched to further speedup the selection procedure. Experimental results prove that when applying the proposed algorithm, lots of encoding time can be saved, with negligible loss in PSNR and little increment in bit rate.Generally speaking, MRF-ME can improve the accuracy of motion estimation, thus further increase the entire coding performance. However, the high complexity introduced by MRF-ME limit the application of this technique. This thesis put forward a fast multiple reference frames selection method to reduce the coding complexity. In this algorithm, if current coding block is temporal stationary, only the nearest reference frame is checked; otherwise spatial correlation and correlation within different coding modes are used to predict the best reference frame; finally, the motion vectors (MVs) and the cost information of current block on different reference frames are applied to accelerate the selection procedure. The experiment demonstrates that with the proposed algorithm, encoding time is largely saved, while the Rate-Distortion (R-D) performance is very similar to that of full search.