基于深度学习的视频编码单元选择算法研究  

Video coding unit selection algorithms based on deep learning

作　　者：温洁 杨帆 潘旭冉 王晓宇 范海瑞 WEN Jie;YANG Fan;PAN Xuran;WANG Xiaoyu;FAN Hairui(School of Electronics and Information Engineering,Hebei University of Technology,Tianjin 300401,China;Tianjin Key Laboratory of Electronic Materials and Devices,Tianjin 300401,China)

机构地区：[1]河北工业大学电子信息工程学院,天津300401 [2]天津市电子材料与器件重点实验室,天津300401

出　　处：《河北工业大学学报》2020年第6期28-34,39,共8页Journal of Hebei University of Technology

基　　金：天津市自然科学基金(17JCTPJC54500);河北省高等学校科学技术研究重点项目(ZD2016123)。

摘　　要：新一代视频编码标准VP9与上一代标准VP8相比性能提升了近1倍,且它的开源特性使得其在视频编码领域取得了广泛应用,但是编码性能提高的同时带来了编码复杂度的增加,从而对一些实时的视频应用产生很大影响。因此本文通过对编码单元模式划分复杂度过高问题的影响因素进行分析,提出基于深度学习的视频编码单元选择算法,该算法首先选择对编码复杂度很高的块划分进行研究,主要针对超级块划分模式的选择进行了优化。应用深度学习中的全连接神经网络模型作为划分模型,输入特征向量为36个,输出是具体的块划分模式,训练方式选择离线训练。其次,为了进一步的简化模型结构同时提升分类器的性能,将对复杂度很高的四叉树递归划分方式进行优化,并根据具体的量化参数(QP)值和块大小得到不同的结构,以便得到1个4层二分类模型。最后,通过对不同复杂视频图像应用简化版的四叉树进行测试,测试结果与原四叉树递归算法相比编码复杂度降低很多,编码复杂度平均降低比例高达77.84%,编码效率得到了很大的提升。The performance of the new generation video coding standard VP9 is nearly double that of the previous gener⁃ation VP8,and its open source characteristics make it widely used in the field of video coding.However,the improve⁃ment of coding performance brings about an increase in coding complexity,which has a great impact on some real-time video applications.Therefore,this paper proposes a video coding unit selection algorithm based on deep learning after an⁃alyzing the influencing factors of the problem of excessively high coding unit mode partitioning.The algorithm selects the block partition with high coding complexity,mainly for super block partitioning.The choice of mode is optimized.The fully connected neural network model in deep learning is applied as the partition model,and the input feature vector is 36.The output is a specific block partition mode,and the training mode selects offline training.Secondly,in order to fur⁃ther simplify the model structure and improve the performance of the classifier,the recursive partitioning method of the highly complex quadtree is optimized,and different structures are obtained according to the specific QP value and block size,so as to obtain a four-layer-two-classification model.Finally,by applying a simplified version of the quadtree to dif⁃ferent complex video images,the test results are much simpler than the original quadtree recursive algorithm.The aver⁃age coding complexity is reduced by 77.84%,and the coding efficiency has been greatly improved.

关 键 词：视频图像编码 VP9 编码单元划分 深度学习 SAD值 

分 类 号：TN919.81[电子电信—通信与信息系统]