Scholarlink

<p>With the rapid development of the Internet of Video Things and the explosive growth of video data, the resulting pressure on storage and transmission has made video compression increasingly important. Conventional codecs lack sufficient content adaptivity, while recent learned compression methods often incur high deployment cost. To address this challenge, this paper proposes a novel visual large model-enhanced soft actor-critic framework (VLM-SAC) for content-aware adaptive video compression. The proposed VLM-SAC formulates video compression as a quality-constrained bandwidth minimization problem and exploits the semantic capability derived from the VLM to guide lightweight state representation learning for efficient compression decision-making. Based on this content-aware state, a soft actor-critic policy is trained to jointly optimize bitrate reduction and quality preservation. Experiments show that VLM-SAC achieves competitive bandwidth efficiency and compression performance, while maintaining substantially lower deployment overhead and better practical deployability than heavy learned codecs.</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

VLM-SAC: a visual large model-enhanced DRL approach for bandwidth-efficient video compression in multi-cloud environment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Qian Wei,
  • Enfang Cui,
  • Zhiyuan Liang,
  • Yuting Wu,
  • Gang Lu,
  • Chao Zheng

摘要

With the rapid development of the Internet of Video Things and the explosive growth of video data, the resulting pressure on storage and transmission has made video compression increasingly important. Conventional codecs lack sufficient content adaptivity, while recent learned compression methods often incur high deployment cost. To address this challenge, this paper proposes a novel visual large model-enhanced soft actor-critic framework (VLM-SAC) for content-aware adaptive video compression. The proposed VLM-SAC formulates video compression as a quality-constrained bandwidth minimization problem and exploits the semantic capability derived from the VLM to guide lightweight state representation learning for efficient compression decision-making. Based on this content-aware state, a soft actor-critic policy is trained to jointly optimize bitrate reduction and quality preservation. Experiments show that VLM-SAC achieves competitive bandwidth efficiency and compression performance, while maintaining substantially lower deployment overhead and better practical deployability than heavy learned codecs.