Improved Intra Prediction using a Least Square Method for High-performance Video Coding

Abstract

Intra prediction has been adopted in MPEG-4 Part 10 Advanced Video Coding (AVC) to improve video coding efficiency by reducing spatial redundancy. The Intra prediction scheme in AVC predicts the current block pixels from its neighboring reconstructed blocks in different directional prediction modes. It has nine directional prediction modes for 4x4 and 8x8 Intra prediction coding and four directional prediction modes for 16x16 Intra prediction coding. Since the Intra prediction only uses the boundary pixels in the neighboring reconstructed blocks, the prediction capability is limited by not fully utilizing the other pixels in the neighboring blocks. Thus, in this thesis, we study an improvement on the Intra prediction of AVC by using a Least Square method. The proposed Lease Square based Intra Prediction (LSP) method is trained with the pixels in a causal window constituted by the reconstructed neighboring blocks, then, the coefficients of the LSP model is used to predict the current pixels so that a decoder can obtain exactly the same coefficients, not requiring to transmit them to the decoder. The proposed LSP method is replaced with the existing vertical left mode so that the total number of the directional modes for $4\times4$ and $8\times8$ Intra prediction remains the same. Furthermore, the LSP method includes a bias term into the linear model which is useful to compensate the illumination difference inside the local blocks. The experimental results show that the proposed LSP based Intra prediction yields 0.67% of bit reduction in BDBR with 0.046dB increment in BDPSNR.