Scholarlink

<p>In low-light imaging scenarios, camera sensors capture RAW data that often suffer from significant noise and signal coupling, leading to blurred details and colour distortion in reconstructed images. Traditional image signal processing techniques, primarily designed for well-lit conditions, struggle to maintain image quality under such constraints. To address these challenges, we introduce a two-stage joint denoising and demosaicing framework that integrates frequency decoupling with bionic colour perception. The first stage employs a wavelet transform for frequency decomposition, coupled with a dual-branch structure to enhance low-frequency data and suppress high-frequency noise collaboratively. The second stage incorporates a colour cross-attention module inspired by human visual perception, ensuring local colour consistency and improving overall image fidelity. Our experiments demonstrate superior performance in noise removal and colour restoration, outperforming existing methods in both objective metrics (e.g., PSNR and SSIM) and subjective evaluations. This work presents a significant advancement in low-light RAW image enhancement, offering practical solutions for imaging systems operating under challenging lighting conditions. The relevant code is publicly available at <a href="https://github.com/JavaAiNiU/JDD_CCA/tree/main">https://github.com/JavaAiNiU/JDD_CCA/tree/main</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

Frequency-decoupled and bionic-inspired RAW image enhancement for low-light conditions

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Pengcheng Zhao,
  • Bingjie Han,
  • Zuojun Chen,
  • Meimei Zhang,
  • Pinle Qin,
  • Jianchao Zeng,
  • Ni Li,
  • Guangan Xie

摘要

In low-light imaging scenarios, camera sensors capture RAW data that often suffer from significant noise and signal coupling, leading to blurred details and colour distortion in reconstructed images. Traditional image signal processing techniques, primarily designed for well-lit conditions, struggle to maintain image quality under such constraints. To address these challenges, we introduce a two-stage joint denoising and demosaicing framework that integrates frequency decoupling with bionic colour perception. The first stage employs a wavelet transform for frequency decomposition, coupled with a dual-branch structure to enhance low-frequency data and suppress high-frequency noise collaboratively. The second stage incorporates a colour cross-attention module inspired by human visual perception, ensuring local colour consistency and improving overall image fidelity. Our experiments demonstrate superior performance in noise removal and colour restoration, outperforming existing methods in both objective metrics (e.g., PSNR and SSIM) and subjective evaluations. This work presents a significant advancement in low-light RAW image enhancement, offering practical solutions for imaging systems operating under challenging lighting conditions. The relevant code is publicly available at https://github.com/JavaAiNiU/JDD_CCA/tree/main.