基于生成模型的视频图像重建方法综述  

Survey on video image reconstruction method based on generative model

作　　者：王延文 雷为民 张伟 孟欢 陈新怡 叶文慧 景庆阳 WANG Yanwen;LEI Weimin;ZHANG Wei;MENG Huan;CHEN Xinyi;YE Wenhui;JING Qingyang(School of Computer Science and Engineering,Northeastern University,Shenyang 110169,China)

机构地区：[1]东北大学计算机科学与工程学院,辽宁沈阳110169

出　　处：《通信学报》2022年第9期194-208,共15页Journal on Communications

基　　金：中央高校基本科研业务费专项资金资助项目(No.N2216010);国家重点研发计划基金资助项目(No.2018YFB1702000)。

摘　　要：基于像素相关性的传统视频压缩技术性能提升空间受限,语义压缩成为视频压缩编码的新方向,视频图像重建是语义压缩编码的关键环节。首先介绍了针对传统编码优化的视频图像重建方法,包括如何利用深度学习提升预测精度和利用超分辨率技术增强重建质量;其次讨论了基于变分自编码器、基于生成对抗网络、基于自回归模型和基于Transformer模型的视频图像重建方法,并根据图像的不同语义表征对模型进行分类,对比了各类方法的优缺点及其适用场景;最后总结了现有视频图像重建存在的问题,并进一步展望研究方向。Traditional video compression technology based on pixel correlation has limited performance improvement space,semantic compression has become the new direction of video compression coding,and video image reconstruction is the key link of semantic compression coding.First,the video image reconstruction methods for traditional coding optimization were introduced,including how to use deep learning to improve prediction accuracy and enhance reconstruction quality with super-resolution techniques.Second,the video image reconstruction methods based on variational auto-encoders,generative adversarial networks,autoregressive models and transformer models were discussed emphatically.Then,the models were classified according to different semantic representations of images.The advantages,disadvantages,and applicable scenarios of various methods were compared.Finally,the existing problems of video image reconstruction were summarized,and the further research directions were prospected.

关 键 词：视频压缩编码 图像重建 生成对抗网络 变分自编码器 Transformer模型 

分 类 号：TN91[电子电信—通信与信息系统]