基于增强可逆性插值滤波器设计的编码方法  

An enhanced invertibility-driven interpolation filter for vedio coding

作　　者：张秋阳 黄晓峰 殷海兵 ZHANG Qiuyang;HUANG Xiaofeng;YIN Haibing(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China)

机构地区：[1]杭州电子科技大学通信工程学院,浙江杭州310018

出　　处：《杭州电子科技大学学报（自然科学版）》2022年第2期14-20,55,共8页Journal of Hangzhou Dianzi University：Natural Sciences

基　　金：国家自然科学基金资助项目(61901150)。

摘　　要：可逆性分像素插值滤波器可以解决插值没有真实样本的难点,但是,存在传统卷积核形状固定、正则项损失函数冲突等不足。为此,提出一种增强的可逆性插值滤波器设计优化方案。首先,引入可变形卷积层,改变卷积核的形状和不同位置的像素参与卷积的权重,增大了感受野,提高了网络的适应性;然后,在正则项设计中,引入生成对抗网络,提升了网络的收敛能力;最后,使用基于运动模糊方法生成的训练样本来替代原本的基于离散余弦变换生成的样本,达到更逼近真实运动的效果。实验结果表明,和H.265相比,改进方案的BD-rate指标提升了2.56%。The Invertibility-driven fractional interpolation filter can solve the problem that the fractional interpolation task does not have ground truths,but it still has some drawbacks such as inefficient convolution layer and conflicting loss functions.In this paper,we propose an enhanced InvIF design to overcome those drawbacks.Firstly,we introduce the deformable convolution layer which has the ability to change the shape of the kernel and the weight of each pixel.It can enlarge the reception field and enhance the network s flexibility.Secondly,we bring the generative adversarial networks into the training scheme to solve the contradictory loss function.Finally,we utilize the motion blur operation to generate the regularization samples in order to simulate the real world motion.Experiments show that our method reach 2.56%BD-rate improvement compared to H.265.

关 键 词：视频编码 帧间预测 分像素插值 可变形卷积神经网络 

分 类 号：TN919.81[电子电信—通信与信息系统]