采用双网络结构的压缩视频超分辨率重建  

Super-resolution Reconstruction of Compressed Video Using Dual Network Architecture

作　　者：周航 何小海[1,2] 王正勇 熊淑华[1] Karn Pradeep[1] ZHOU Hang;HE Xiaohai;WANG Zhengyong;XIONG Shuhua;Karn Pradeep(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China;Key Laboratory of Wireless Power Transmission of Ministry of Education,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,成都610065 [2]无线能量传输教育部重点实验室,成都610065

出　　处：《电讯技术》2020年第1期81-86,共6页Telecommunication Engineering

基　　金：国家自然科学基金资助项目(61871279,61471248);成都市产业集群协同创新项目(2016-XT00-00015-GX)

摘　　要：在实际应用中,为了节省带宽和方便存储,图像和视频通常被下采样和压缩,而降质的图像与视频无法满足人们的实际需求。针对这一问题,采用了一种双网络结构的超分辨率重建方法,首先建立下采视频与压缩后的低分辨率视频的映射关系,然后建立质量增强的压缩视频与原始视频的映射关系,最终在输出端可以得到质量提升的视频帧。在网络中,采用密集残差块来提取压缩视频中丰富的局部分层特征,并结合全局残差学习恢复视频中的高频信息。在压缩环节,采用高性能视频编码来验证所提算法的有效性。实验结果表明,相比于主流的视频编码标准和先进的超分辨率重建算法,所提方法能有效提升编码视频的率失真性能。In practical applications,in order to save bandwidth and facilitate storage,images and videos are usually downsampled and compressed,while low-quality and low-resolution images and videos can not meet the actual needs of people.In order to solve this problem,a super-resolution reconstruction method based on dual-network structure is adopted.Firstly,the mapping relationship between downloaded video and compressed low-resolution video is established,and then the mapping relationship between compressed video with enhanced quality and original video is established.Finally,the video frame with improved quality can be obtained at the output end.In the network,dense residual blocks are used to extract rich local hierarchical features in compressed video,and global residual learning is used to recover high-frequency information in video.In the compression stage,high efficiency video coding(HEVC)is used to verify the effectiveness of the proposed algorithm.The experimental results show that the proposed method can effectively improve the rate-distortion performance of coded video compared with the mainstream video coding standards and advanced super-resolution reconstruction algorithms.

关 键 词：高性能视频编码 视频压缩 超分辨率重建 卷积神经网络 

分 类 号：TN919.81[电子电信—通信与信息系统]