GBST: Separable transforms based on line graphs for predictive video coding
Hilmi E. Egilmez, Yung‐Hsuan Chao, Antonio J. Ortega, Bumshik Lee, Sehoon Yea · 2016
This paper introduces a novel class of transforms, called graph-based separable transforms (GBSTs), based on two line graphs with optimized weights. For the optimal GBST construction, we formulate a graph learning problem to design two separate line graphs using row-wise and column-wise residual block statistics, respectively. We also analyze the optimality of resulting separable transforms for both intra and inter predicted residual block models. Moreover, we show that separable DCT and ADST (DST-7) are special cases of the GBSTs. Our experimental results demonstrate that the proposed optimized transforms outperform 2-D DCT/ADST and separable KLT.