Scholarlink

<p>Recent advancements in deep learning for image compression have shown promise in reducing data redundancy and optimizing rate-distortion performance using CNNs and Transformer architectures. However, the challenge of balancing efficiency and performance persists. Drawing inspiration from the growing attention on state space models (SSMs) for their long-range modeling capabilities and efficiency, we propose a method that combines the strengths of CNNs and LocalMamba to balance local and global features. This leads to the development of an efficient parallel CNN-LocalMamba hybrid module that enhances the rate-distortion performance of learned image compression (LIC) under controllable complexity. The entropy model is critical for estimating the distribution of latent representations and improving rate-distortion performance. However, most existing methods capture conditional probabilities in only one dimension, leading to inaccurate estimates and residual redundancy. Moreover, current spatial context modules rely on computationally intensive quadratic complexity algorithms, limiting the efficiency of high-resolution image encoding. To address these issues, we propose a multi-context entropy model that effectively captures local, global spatial, and channel contexts, featuring linear complexity and suitability for efficient high-resolution image encoding. Experimental results show that our method outperforms recent studies on PSNR and MS-SSIM metrics across Kodak, Tecnick, and CLIC Pro datasets. Compared to VTM, our approach reduces BD-rate while maintaining computational efficiency. Our method demonstrates superior rate-distortion performance and exhibits greater potential and scalability, especially in high-resolution image processing.</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

Learned image compression with CNN-LocalMamba architectures and multi-context entropy model

  • 期刊论文

  • 发表日期：

    2025年7月31日

查看全文
  • Qian Zhang,
  • Huijin Li,
  • Mingwen Shao,
  • Hong Liang

摘要

Recent advancements in deep learning for image compression have shown promise in reducing data redundancy and optimizing rate-distortion performance using CNNs and Transformer architectures. However, the challenge of balancing efficiency and performance persists. Drawing inspiration from the growing attention on state space models (SSMs) for their long-range modeling capabilities and efficiency, we propose a method that combines the strengths of CNNs and LocalMamba to balance local and global features. This leads to the development of an efficient parallel CNN-LocalMamba hybrid module that enhances the rate-distortion performance of learned image compression (LIC) under controllable complexity. The entropy model is critical for estimating the distribution of latent representations and improving rate-distortion performance. However, most existing methods capture conditional probabilities in only one dimension, leading to inaccurate estimates and residual redundancy. Moreover, current spatial context modules rely on computationally intensive quadratic complexity algorithms, limiting the efficiency of high-resolution image encoding. To address these issues, we propose a multi-context entropy model that effectively captures local, global spatial, and channel contexts, featuring linear complexity and suitability for efficient high-resolution image encoding. Experimental results show that our method outperforms recent studies on PSNR and MS-SSIM metrics across Kodak, Tecnick, and CLIC Pro datasets. Compared to VTM, our approach reduces BD-rate while maintaining computational efficiency. Our method demonstrates superior rate-distortion performance and exhibits greater potential and scalability, especially in high-resolution image processing.