基于深度学习帧内编码快速划分算法  

Fast intra coding partition algorithm based on deep learning

作　　者：曾奥迪 杨静[1] ZENG Ao-di;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：《计算机工程与设计》2023年第4期1014-1020,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61902239)。

摘　　要：为降低H.266/VVC(versatile video coding)帧内编码复杂度,提出一种基于纹理分类的卷积注意力机制神经网络CBAM-CNN(convolutional block attention module-convolution neural network)模型的快速帧内编码单元划分方法,替代嵌套多类型四叉树QTMT(quadtree with nested multi-type tree)的遍历搜索。综合考虑每个编码单元CU(codinguint)纹理特征,建立基于阈值的纹理分类模型,判断64×64的编码单元是否划分;设计并训练一种CBAM-CNN网络模型,预测32×32,16×16CU的划分结果。仿真结果表明,所提算法使帧内编码时间平均减少43.217%,编码比特率仅增加0.894%,有效降低了帧内预测的编码复杂度。To reduce the intra coding complexity of H.266/VVC(versatile video coding),a fast intra coding unit partition method of CBAM-CNN(convolutional block attention module convolution neural network)model based on texture classification was proposed to replace the mandatory QTMT(quadtree with nested multi-type tree)search.Considering the relationship between CU depth,quantization parameters and texture complexity,a threshold-based texture classification model was established to determine whether 64×64 sized coding units needed to be divided.CBAM-CNN network model was designed and trained to predict 32×32,16×16 sized CTU division result.Simulation experimental results show that the average intra-frame coding time can be reduced by 43.217%,while BDBR only increases by 0.894%,which effectively reduces the coding complexity of intra prediction.

关 键 词：帧内预测 纹理复杂度 阈值分类 编码块划分 嵌套多类型四叉树划分 注意力机制 编码复杂度 

分 类 号：TN919.81[电子电信—通信与信息系统]