低复杂度HEVC帧内编码快速划分算法  

Fast Partition Algorithm of Low Complexity HEVC Intra Coding

作　　者：周帅燃 杨静[1] ZHOU Shuai-ran;YANG Jing(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院,上海201306

出　　处：《小型微型计算机系统》2021年第7期1475-1478,共4页Journal of Chinese Computer Systems

基　　金：国家自然科学资金项目(61902239)资助。

摘　　要：高效视频编码(HEVC)在H.264之上实现了显著的编码性能.但是却以明显的编码复杂度为代价获得了性能上的提高,其中,由于对所有可能的编码单元(CU)进行基于率失真优化的遍历搜索,因此编码树单元(CTU)最耗时的部分就是CU的划分.为了解决此类问题,本文提出了一种基于纹理分类的深度卷积神经网络(CNN)模型来对CTU的划分进行预测.首先通过考虑每个CU的特征,开发了基于阈值的纹理分类模型来识别纹理简单的编码单元.其次,设计并训练一种CNN结构来预测纹理复杂的编码单元.最后,依据实验结果表明,本文提出的算法与原始HM16.5相比该方案可节省60.28%的帧内编码时间,而BD速率损失可忽略不计2.15%.与其他优秀算法相比,本文算法减少帧内编码时间更明显,并且编码质量更优.High efficiency video coding(HEVC)achieves significant coding performance on top of H.264.However,the performance is improved at the cost of obvious coding complexity.Among them,the coding tree unit(CTU)is the most expensive because of the Rate Distortion Optimization(RDO)based traversal search for all possible coding units(CU).The part of time is the division of CU.In order to solve such problems,this paper proposes a deep convolutional neural network(CNN)model based on texture classification to predict CTU partition.First,by considering the characteristics of each CU,a threshold based texture classification model was developed to identify coding units with simple textures.Second,design and train a CNN structure to predict coding units with complex textures.Finally the experimental results show that the proposed algorithm can save 60.28%intra coding time compared with the original HM16.5,while the BD rate loss is negligible by 2.15%.Compared with other excellent algorithms,the proposed algorithm reduces the intra-frame encoding time more obviouslyand has better coding quality.

关 键 词：帧内编码 编码单元划分 深度卷积网络 纹理复杂度 

分 类 号：TN919.81[电子电信—通信与信息系统]