Scholarlink

<p>In multispectral image compression, Convolutional Neural Network (CNN) effectively capture local features but are limited in modeling global context. Transformer architectures address this limitation by leveraging self-attention to model global dependencies, while the Mamba architecture–built upon state space model (SSM)–further enhances long-range sequence modeling. However, existing methods fail to effectively integrate the strengths of these approaches. To overcome this, we propose MSC-NET, a unified fusion network that combines Mamba-based sequence modeling with cross-channel aggregation Transformers. MSC-NET incorporates a Spectral Redundancy Elimination Attention (SREA) module to reduce spectral redundancy and an RMamba block to enhance global contextual representation. Additionally, a Cross-Channel Aggregation (CCA) module integrates the outputs of SREA and RMamba, improving the fusion of local features and inter-channel dependencies. Experimental results show that MSC-NET achieves superior rate-distortion performance, effectively balancing compression ratio and image quality, thereby validating the effectiveness of the proposed approach.</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

Mamba and cross-channel aggregation for efficient multispectral image compression

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月21日

查看全文
  • Jingang Wang,
  • Qizhi Fang,
  • Jiahui Liu,
  • Lili Zhang

摘要

In multispectral image compression, Convolutional Neural Network (CNN) effectively capture local features but are limited in modeling global context. Transformer architectures address this limitation by leveraging self-attention to model global dependencies, while the Mamba architecture–built upon state space model (SSM)–further enhances long-range sequence modeling. However, existing methods fail to effectively integrate the strengths of these approaches. To overcome this, we propose MSC-NET, a unified fusion network that combines Mamba-based sequence modeling with cross-channel aggregation Transformers. MSC-NET incorporates a Spectral Redundancy Elimination Attention (SREA) module to reduce spectral redundancy and an RMamba block to enhance global contextual representation. Additionally, a Cross-Channel Aggregation (CCA) module integrates the outputs of SREA and RMamba, improving the fusion of local features and inter-channel dependencies. Experimental results show that MSC-NET achieves superior rate-distortion performance, effectively balancing compression ratio and image quality, thereby validating the effectiveness of the proposed approach.