Scholarlink

<p>Low-light image enhancement (LLIE) is crucial for vision tasks such as object detection and recognition. Traditional Retinex-based methods often fail to adequately suppress noise in the reflectance component, leading to noisy or over-enhanced images. To address this limitation, we propose DSRTNet, a model with two branches to optimize the reflectance component. The Multi-scale Feature Estimation module refines the U-Net architecture to capture multi-scale features. Meanwhile, the Feature Enhancement and Fusion module further processes and fuses these features to suppress global noise. The Image Decomposition-Denoising-Reconstruction (IDDR) module employs a Laplacian pyramid to decompose the input image, apply localized denoising, and reconstruct the enhanced output. For the illumination component, gamma correction adjusts brightness, and the adjusted result is combined with the denoised reflectance component to generate the final enhanced image. Evaluations on diverse low-light datasets and quantitative comparisons of PSNR, SSIM, and LPIPS metrics demonstrate the superiority of DSRTNet, achieving a 6.2% mean Average Precision (mAP) improvement in object detection tasks. Its novel architecture offers fresh insights for low-light LLIE.</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

Dual-stream noise suppression low-light image enhancement network based on Retinex theory

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Guang Han,
  • Binming Zeng,
  • Zifan Rui,
  • Yaolong Hu,
  • Xiaofan Yu

摘要

Low-light image enhancement (LLIE) is crucial for vision tasks such as object detection and recognition. Traditional Retinex-based methods often fail to adequately suppress noise in the reflectance component, leading to noisy or over-enhanced images. To address this limitation, we propose DSRTNet, a model with two branches to optimize the reflectance component. The Multi-scale Feature Estimation module refines the U-Net architecture to capture multi-scale features. Meanwhile, the Feature Enhancement and Fusion module further processes and fuses these features to suppress global noise. The Image Decomposition-Denoising-Reconstruction (IDDR) module employs a Laplacian pyramid to decompose the input image, apply localized denoising, and reconstruct the enhanced output. For the illumination component, gamma correction adjusts brightness, and the adjusted result is combined with the denoised reflectance component to generate the final enhanced image. Evaluations on diverse low-light datasets and quantitative comparisons of PSNR, SSIM, and LPIPS metrics demonstrate the superiority of DSRTNet, achieving a 6.2% mean Average Precision (mAP) improvement in object detection tasks. Its novel architecture offers fresh insights for low-light LLIE.