基于细节还原卷积神经网络的压缩视频质量增强技术研究
DOI:
作者:
作者单位:

南京信息工程大学 计算机学院

作者简介:

通讯作者:

中图分类号:

基金项目:

国家自然科学基金项目(面上项目,重点项目,重大项目)


Study on Quality Enhancement via Detail Recovery Convolutional Neural Network For Compressed Video
Author:
Affiliation:

1.School of Computer Science,Nanjing University of Information Science and Technology;2.School of Computer Science, Nanjing University of Information Science and Technology

Fund Project:

The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    视频编码技术有效地解决了原始视频数据量过大的问题。然而,压缩效率的提升带来了视频质量下降。为了提高压缩视频的视觉质量,本文提出了一种基于细节还原卷积神经网络(Detail Recovery Convolutional Neural Network, DRCNN)的视频质量增强方法,该方法由一个主去噪分支和一个细节补偿分支组成。为了有效地提取和消除压缩失真,本文提出了一个主去噪分支,在此分支中提出了一个多尺度失真特征提取块(Multi-Scale Distortion Feature Extraction Block, MDFEB),使其更加关注压缩视频中的失真区域,并提高所提出的DRCNN的失真特征学习能力。此外,为了丰富压缩视频中的细节,本文提出了细节补偿分支。细节补偿分支首先采用预训练的ResNet-50组成的内容特征提取器,提供丰富的内容特征,如突出的物体、形状、细节等;然后通过设计的细节响应块(Detail Response Block, DRB)从内容特征中有效地提取细节特征。大量的实验结果表明,与最先进的方法相比,所提出的DRCNN实现了最佳的压缩视频质量增强性能。

    Abstract:

    Video coding techniques effectively address the data volume problem of raw video, however, the achieved compression efficiency comes at the cost of video quality degradation. To improve the visual quality of compressed video, a Detail Recovery Convolutional Neural Network (DRCNN)-based video quality enhancement method is proposed in this paper, which consists of a main denoising branch and a detail compensation branch. To effectively extract and eliminate the compression distortions, a main denoising branch is proposed, in which a Multi-Scale Distortion Feature Extraction Block (MDFEBs) is proposed to pay more attention to the distortion areas in the compressed video, and improve the distortion feature learning ability of the proposed DRCNN. Furthermore, to enrich the details in the compressed video, a detail compensation branch is proposed, which firstly adopts a content feature extractor composed of a pre-trained ResNet-50 to provide abundant content features, such as salient objects, shapes, details and so on, and then a Detail Response Block (DRB) is designed to efficiently extract the detailed features from the content features. Extensive experimental results show that the proposed DRCNN achieves the best compressed video quality enhancement performance as compared with the state-of-the-art methods.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2022-03-21
  • 最后修改日期:2022-03-27
  • 录用日期:2022-03-28
  • 在线发布日期:
  • 出版日期:

地址:江苏南京,宁六路219号,南京信息工程大学    邮编:210044

联系电话:025-58731025    E-mail:nxdxb@nuist.edu.cn

南京信息工程大学学报 ® 2022 版权所有  技术支持:北京勤云科技发展有限公司