Enhanced intra prediction and transform for video coding

Intra-coding plays an important role in video coding schemes and it has become a hot research topic of the next generation video coding standard. In this paper, we propose two techniques to improve the efficiency of the prediction and the transform, respectively, for intra-coding. First, we introduce an overlapped block intra-prediction (OBIP) method, which makes use of the intra directional prediction modes of neighboring blocks as well as the mode of the current block. A position dependent weighted sum of several possible predictors will be treated as the final prediction. Second, we propose a multiple-model Karhunen-Loève transform (MMKLT) technique to further improve the mode dependent directional transform (MDDT) method. More than one KLT will be trained offline for each intra-prediction mode based on residual variances. Experimental results indicate that the proposed two methods together can achieve an average bit-rate reduction about 11% compared with H.264 or 5% compared with MDDT in all-intra coding.