Studies On The Sub-pixel Motion Compensation In Video Coding

The video applications with high resolution, high sampling bit depth, or high fidelity chroma format increase the video source data dramatically. Accordingly, compressing video source data more efficiently becomes very necessary. Meanwhile, with the coding efficiency improvement, the implementation complexity of decoder and encoder should be in consideration. For this purpose, ITU-T video coding expert group (VCEG) has established a software experiment platform and has been doing some studies towards the next generation video coding standard. Based on this platform, this thesis sets its research direction on the technique of sub-pixel motion compensation.This thesis establishes a new mathematical model for sub-pixel motion compensation and indicates that, during the process of motion compensation, aliasing effects, motion vector errors and quantization errors deteriorate the prediction accuracy. Using wiener filter to interpolate the reconstructed reference pictures can reduce the prediction errors. In an extreme case, when the aliasing effects are very severe, using wiener filter can reduce the prediction errors'energy by about 50% in comparison with using the ideal low pass filter.In order to implement wiener filter more efficiently and with less complexity, for luma component, this thesis studies the techniques of frame level adaptive interpolation filter and macroblock level adaptive interpolation filter respectively.During the research in the frame level adaptive interpolation filter for luma component, this thesis proposes a combined adaptive-fixed interpolation (CAFI) scheme. Test results show that the CAFI scheme can reduce the overall bit rates by about 27% at the same PSNR compared with H.264/AVC 6 tap fixed coefficient interpolation filter for 4:2:0 video sequences. Compared with other adaptive interpolation schemes in the software platform, the CAFI scheme shows a better balance between coding performance improvement and implementation complexity increment.During the research in the macroblock level adaptive interpolation filter for luma component, considering that the memory access bandwidth is a major bottleneck in the high end real time applications, this thesis proposes a low memory bandwidth macroblock level adaptive interpolation (LMMAIF) scheme. Tests results show that the LMMAIF scheme can reduce the memory bandwidth for luma component by about 40%, meanwhile,achieve the similar coding efficiency compared with frame level CAFI scheme.During the research in the sub-pixel motion compensation for 4:4:4 video sequences, this thesis proposes the method of multiple components adaptive interpolation filter (MCAIF). In MCAIF, the frame level CAFI scheme or LMMAIF scheme can be used for luma component. While for chroma components, 2D 2x2 tap adaptive interpolation scheme is proposed in this thesis. Test results show that, for 4:4:4 video sequences, compared with H.264/AVC high4:4:4 profile, MCAIF can reduce the overall bit rates by about 29.46% at the same PSNR, and it reduces the memory bandwidth by about 60%. Moreover, it reduces the computation complexity by about 69% in DSP implementation for chroma components.