基于深度学习的快速QTMT划分  

Fast QTMT partition decision based on deep learning

作　　者：彭双 王晓东[1] 彭宗举[1,2] 陈芬 PENG Shuang;WANG Xiaodong;PENG Zongju;CHEN Fen(Faculty of Information Science and Engineering,Ningbo University,Ningbo 315211,China;Faculty of Electrical and Electronics,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211 [2]重庆理工大学电气与电子学院,重庆400054

出　　处：《电信科学》2021年第4期73-81,共9页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.61771269,No.61620106012);浙江省自然科学基金资助项目(No.LY20F010005);宁波市自然科学基金资助项目(No.2019A610107);重庆理工大学科研启动基金资助项目(No.2020ZDZ029,No.2020ZDZ030)。

摘　　要：与之前的编码标准相比,多功能视频编码(versatile video coding,VVC)进一步提高了压缩效率。嵌套多类树的四叉树(quadtree with nested multi-type tree,QTMT)结构是提高编码增益的关键之一,同时极大地增加了编码复杂度。为降低VVC编码复杂度,提出了一种基于深度学习的快速QTMT划分方法。首先,提出了注意力−非对称卷积结构来预测划分模式的概率。然后,基于阈值提出了快速划分模式决策。最后,提出了编码性能与时间的代价函数来求解最优阈值,提出了阈值决策方法。实验表明,算法在不同档次下的时间节省分别为48.62%、52.93%、62.01%,BDBR分别为1.05%、1.33%、2.38%。结果表明,算法的时间节省和率失真性能优于其他快速算法。Compared with the predecessor standards,versatile video coding(VVC)significantly improves compression efficiency by a quadtree with nested multi-type tree(QTMT)structure but at the expense of extremely high coding complexity.To reduce the coding complexity of VVC,a fast QTMT partition method was proposed based on deep learning.Firstly,an attention-asymmetric convolutional neural network was proposed to predict the probability of partition modes.Then,the fast decision of partition modes based on the threshold was proposed.Finally,the cost of coding performance and time was proposed to obtain the optimal threshold,and the threshold decision method was proposed.Experimental results at different levels show that the proposed method achieves an average time saving of 48.62%/52.93%/62.01%with the negligible BDBR of 1.05%/1.33%/2.38%.Such results demonstrate that the proposed method significantly outperforms other state-of-the-art methods.

关 键 词：VVC QTMT 快速划分决策 深度学习 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]