Scholarlink

<p>Recently upgraded video capture devices have generated massive amounts of high-definition and ultrahigh-definition videos, leading to an escalating challenge in video storage and transmission. Correspondingly, the demand for new-generation video codecs has become increasingly urgent. This paper proposes a novel framework for the codec of fixed-scene videos. The framework decouples the background and foreground for separate processing and employs a generative adversarial network (GAN) model to reconstruct the foreground targets. The proposed method extracts features at the object level and utilize the disentangled features of these objects to achieve high-compression-ratio video encoding. The proposed decoder leverages conditional GANs to reconstruct the compressed features and integrate them with the background for video reconstruction. Experimental results show that the proposed framework and algorithm outperform other methods, particularly in terms of compression rates, while preserving high video reconstruction quality. Models and source code can be referenced at: <a href="https://github.com/icirelease/decoupled-video-compression-recon">https://github.com/icirelease/decoupled-video-compression-recon</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Disentanglement-based compression and generative reconstruction of fixed-scene videos

  • 期刊论文

  • 发表日期：

    2025年5月23日

查看全文
  • Hongbo Huang,
  • Longfei Xu,
  • Meng Li,
  • Linkai Huang,
  • Muye Sun,
  • Feng Gu,
  • Guoqing Hu,
  • Zhehai Zhou

摘要

Recently upgraded video capture devices have generated massive amounts of high-definition and ultrahigh-definition videos, leading to an escalating challenge in video storage and transmission. Correspondingly, the demand for new-generation video codecs has become increasingly urgent. This paper proposes a novel framework for the codec of fixed-scene videos. The framework decouples the background and foreground for separate processing and employs a generative adversarial network (GAN) model to reconstruct the foreground targets. The proposed method extracts features at the object level and utilize the disentangled features of these objects to achieve high-compression-ratio video encoding. The proposed decoder leverages conditional GANs to reconstruct the compressed features and integrate them with the background for video reconstruction. Experimental results show that the proposed framework and algorithm outperform other methods, particularly in terms of compression rates, while preserving high video reconstruction quality. Models and source code can be referenced at: https://github.com/icirelease/decoupled-video-compression-recon.