Scholarlink

<p>Low-light raw image enhancement is a critical yet challenging task in computer vision. Existing methods often rely on convolutional neural networks (CNNs) and Vision Transformer (ViT). CNN-based methods face limitations in capturing multi-scale features and global dependencies, while ViT suffer from high computational complexity. They both struggle to balance efficiency and performance. To address these issues, we propose a novel end-to-end Low Light <b>Raw</b> Image Enhancement framework which leverages the Receptance Weighted Key Value (<b>RWKV</b>) architecture, namely <b>RawRWKV</b>. With the linear attention mechanism of RWKV and the multi-scale U-Net framework, RawRWKV can achieve the goal of enhancing model performance while reducing model complexity, which balances model performance and efficiency. Experiments on the SID and MCR raw datasets demonstrate that RawRWKV achieves state-of-the-art (SOTA) performance, with significantly the highest PSNR (30.84 dB on SID, 33.12 dB on MCR) and SSIM (0.826 on SID, 0.926 on MCR) compared to CNN and Transformer baselines, while maintaining low parameters (4.00 M) and FLOPS (21.26 G). This work highlights the potential of RWKV in bridging the gap between performance and efficiency for low-light raw image enhancement tasks.</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

RawRWKV : An efficient raw image enhancement framework via RWKV architecture

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Xingbo Dong,
  • Yongzhou Liu,
  • Zhe Jin,
  • Wen Sha

摘要

Low-light raw image enhancement is a critical yet challenging task in computer vision. Existing methods often rely on convolutional neural networks (CNNs) and Vision Transformer (ViT). CNN-based methods face limitations in capturing multi-scale features and global dependencies, while ViT suffer from high computational complexity. They both struggle to balance efficiency and performance. To address these issues, we propose a novel end-to-end Low Light Raw Image Enhancement framework which leverages the Receptance Weighted Key Value (RWKV) architecture, namely RawRWKV. With the linear attention mechanism of RWKV and the multi-scale U-Net framework, RawRWKV can achieve the goal of enhancing model performance while reducing model complexity, which balances model performance and efficiency. Experiments on the SID and MCR raw datasets demonstrate that RawRWKV achieves state-of-the-art (SOTA) performance, with significantly the highest PSNR (30.84 dB on SID, 33.12 dB on MCR) and SSIM (0.826 on SID, 0.926 on MCR) compared to CNN and Transformer baselines, while maintaining low parameters (4.00 M) and FLOPS (21.26 G). This work highlights the potential of RWKV in bridging the gap between performance and efficiency for low-light raw image enhancement tasks.