Efficient inter prediction for 2D and 3D video coding

In the recent several years, the popularity of High Definition and Three-Dimensional (3D) videos are creating significant challenge for the video compression techniques. There is an increasing demand for video coding tools with higher efficiency and lower complexity. Using advanced inter prediction technique the state-of-the-art video coding standard H.264/ A VC has achieved significant compression gains. However, the complexity of the inter prediction is extremely high, which limits the application of real-time video service. When applying H.264/ A VC to colour plus depth represented 3D videos, the inter prediction designed for Two-Dimensional (2D) videos is not efficient enough for depth map coding, since the depth map is only a 2D signal to aid virtual view rendering. The work described in this thesis speeds up the inter prediction by proposing several fast approaches. A novel inter prediction scheme for 3D video coding is developed to maximise rate distortion performance of rendering image and depth map bit rates. The thesis proposes four methods to reduce complexity of inter prediction. Firstly, the motion estimation searching points are reduced by proposing a novel Adaptive Direction Search algorithm. In this algorithm, the intensive search and coarse search regions are adaptively detected using predicted motion vector direction as an indicator. Secondly, an Adaptive Simulated Annealing Search algorithm is developed to further adjust the search pattern not only in directional regions but also in different distances. Thirdly, the thesis presents a hybrid fast mode decision algorithm, which is mainly based on Coded Block Pattern to infer block's homogeneity characteristic so that the unnecessary inter modes can be eliminated. Fourthly, an improved residual homogeneous based fast mode decision algorithm is proposed. Computational complexity can be further reduced at low Quantisation Parameter conditions. Experimental results indicate that using the proposed algorithms up to 70% of encoding time can be saved with negligible Rate Distortion performance losses. For depth map coding, a novel inter prediction scheme is proposed in this thesis. By employing rendered view distortion in Rate-Distortion Optimisation, the optimal motion vector and inter mode are determined to maximise the quality of the rendered views. An effective Lagrange multiplier is also derived. Finally, fast approaches have been implemented to reduce the complexity. Experimental results illustrate that up to 61 % of bitrate can be saved for depth map coding.EThOS - Electronic Theses Online ServiceGBUnited Kingdo