Scholarlink

<p>In the last few decades, video applications have advanced rapidly throughout the world. More than 80% of the internet traffic is caused by video data. The traffic percentage is expected to increase further in the future since video applications are becoming increasingly crucial. Therefore, video compression is essential to provide high-quality videos under specific bandwidth limits. Existing methods for video compression are based on blocks, which employ sequential processing on every block. Even though these models perform better in efficiently compressing video, the overall coding speed is high. The compression performance is also reduced by local optimization. In this paper, we present a novel unified video compression framework that combines differential operator-based region of interest detection (Sobel, Prewitt, Laplacian), with two original modules Cross-Channel Searching and Feature Grouping inside a hybrid temporal-channel attention pipeline. Unlike prior works that treat these components in isolation, our architecture fuses interpretable edge-aware operators with deep temporal and channel-wise context modeling under a unified, end-to-end rate-distortion optimization scheme, enabling explicit structure-aware compression with high efficiency. This framework demonstrates high efficiency in video compression in the UVG-RGB, HEVC_image, and UCF101 datasets by achieving the highest compression ratio that facilitates efficient storage and transmission of video content.</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

Differential Operator-Based ROI Detection and Hybrid Attention for High-Efficiency Video Compression

  • 期刊论文

  • 发表日期：

    2025年10月16日

查看全文
  • L. C. Manikandan,
  • Mong-Fong Horng,
  • Siva Shankar Subramanian,
  • Maithili Kamalakannan

摘要

In the last few decades, video applications have advanced rapidly throughout the world. More than 80% of the internet traffic is caused by video data. The traffic percentage is expected to increase further in the future since video applications are becoming increasingly crucial. Therefore, video compression is essential to provide high-quality videos under specific bandwidth limits. Existing methods for video compression are based on blocks, which employ sequential processing on every block. Even though these models perform better in efficiently compressing video, the overall coding speed is high. The compression performance is also reduced by local optimization. In this paper, we present a novel unified video compression framework that combines differential operator-based region of interest detection (Sobel, Prewitt, Laplacian), with two original modules Cross-Channel Searching and Feature Grouping inside a hybrid temporal-channel attention pipeline. Unlike prior works that treat these components in isolation, our architecture fuses interpretable edge-aware operators with deep temporal and channel-wise context modeling under a unified, end-to-end rate-distortion optimization scheme, enabling explicit structure-aware compression with high efficiency. This framework demonstrates high efficiency in video compression in the UVG-RGB, HEVC_image, and UCF101 datasets by achieving the highest compression ratio that facilitates efficient storage and transmission of video content.