Computational complexity reduction for HEVC intra prediction with SVM

Unlike most researches focus on computation reduction, the fast algorithm proposed in this paper aims at maintaining the coding efficiency as high as possible. In the proposed algorithm we employ support vector machine (SVM) that uses three parameters as features: variances, low-frequency AC components of DCT and spatially neighboring CU levels for fast CU size decision. In addition, based upon RMD cost we propose an adaptive mode candidates method for further RDO computation. Experimental results demonstrate that average 22% reduction of computational complexity can be achieved but only with 0.09% BD bit rate increase.