Scholarlink

<p>Recent advancements in computer vision necessitate efficient image compression methods that cater to both human and machine perception. Traditional learned image compression models optimized for human visual quality often fail to preserve semantic information crucial for machine vision tasks. This paper introduces a lightweight adapter-based tuning framework incorporating the ChebSpace Block (CSB), which leverages Chebyshev polynomial bases to modulate intermediate features adaptively. This approach enhances convolutional neural network responses to task-relevant features while suppressing spatial noise. Experimental results demonstrate that our method outperforms state-of-the-art Image Compression for Machine and Human Vision (ICMH) approaches across multiple tasks, including classification, object detection, and instance segmentation, with significantly reduced trainable parameters and computational overhead. The source code link is: <a href="https://github.com/DSFDSFSer2/CSB/tree/master">https://github.com/DSFDSFSer2/CSB/tree/master</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

Enhancing joint human–machine image compression via Chebyshev space modulation

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Zhicheng Ma,
  • Ping An,
  • Shipei Wang,
  • Chao Yang,
  • Xinpeng Huang

摘要

Recent advancements in computer vision necessitate efficient image compression methods that cater to both human and machine perception. Traditional learned image compression models optimized for human visual quality often fail to preserve semantic information crucial for machine vision tasks. This paper introduces a lightweight adapter-based tuning framework incorporating the ChebSpace Block (CSB), which leverages Chebyshev polynomial bases to modulate intermediate features adaptively. This approach enhances convolutional neural network responses to task-relevant features while suppressing spatial noise. Experimental results demonstrate that our method outperforms state-of-the-art Image Compression for Machine and Human Vision (ICMH) approaches across multiple tasks, including classification, object detection, and instance segmentation, with significantly reduced trainable parameters and computational overhead. The source code link is: https://github.com/DSFDSFSer2/CSB/tree/master