基于机器学习的HEVC帧内模式快速决策算法

针对高效视频编码(HEVC)计算复杂度过高的情况, 提出了一种基于机器学习的帧内快速决策算法。根据图像内容的平滑 程度将PU划分成3类,对具有一定平滑度的预测单元(PU)不需要遍历完所有的帧内预测模式 ,从而有效降低算法的计算复杂度。首 先,计算各个PU的左边参考像素方差、上边参考像素方差和总参考像素的方差,以及各个P U采用的最优的帧内预测模式, 这些方差反映了参考像素的平滑程度；然后,利用机器学习软件Weka对得到的数据进行分类 处理,得到分类决策树； 最后,根据决策树来判定各个PU需要测试的帧内模式,再对各个PU 遍历这些帧内模式,确 定最优的模式,减少不必要 预测,从而降低编码复杂度。实验结果表明,本文算法相对于标准的HEVC 15.0编码算法,在高码率的情况下,编码时间平 均节省约16.18%,BD-rate平均升高约0.25%,BD-PSNR平均降低约0.02 dB；在低码率的情况下,编码时 间平均节省约20.75%,BD-rate平均升高 约0.04%,BD-PSNR平均降低约0.00dB。

A fast HEVC intra mode decision algorithm based on machine learning

In view of the high computational comp lexity of high efficiency video coding (HEVC) encoding,a fast algorithm based on machine learning is proposed in this paper.According to the smoothness of image content,we divide prediction units (PUs) into th ree classes.The smooth PU has no need to test all the intra prediction modes.Thus,the computational complexity of the algorithm ca n be reduced effectively.First,we calculate the variance of the reference pixels on the left side,the above side of each PU,and the variance of all the reference pixels,as well as the optimal intra prediction mode for each PU.The variances reflect the smoothness of the reference pixels.Then,the machine learning software of Weka is used to classify the obtained data previously,and a decision tree is generated.Finally,according to the decision tree, the intra prediction modes for each PU to be tested are determined,then these i ntra modes are tested for each PU to choose the optimal mode,reducing unnecessary process,thus reducing the encoding complexity. Experimental results show that compared with the standard HEVC 15.0coding algorithm,in the case of high bitrate,the encoding time is reduced by about 16.18% on average with negligible increase of Bjontegaard delta rate (BD-rate) (about 0.25%) and decrease of Bjontegaard delta peak signal-to-noise rate (BD-PS) NR (about 0.02dB).In the case of low bitrate,the encoding time is reduced by about 20.75% on average with negligible increase of BD-rate (about 0.04%) and decrease of BD-PSNR (about 0.00dB).