Scholarlink

The need for high-quality images has increased in many industries in the digital age, but taking sharp pictures in low light is still quite difficult. To boost denoising performance while keeping fine details and colour fidelity, this research investigates the constraints of conventional low-light enhancement methods and existing deep learning models. We suggest a unique deep neural network architecture with subspace attention modules (SSAs) for better image quality and nonlinear activation free (NAF) blocks to overcome these problems. Our method is quite effective, obtaining a PSNR of 24 dB and SSIM of 0.6 with only 2.5 GFLOPs, in contrast to models such as NAFNet (PSNR: 40 dB, SSIM: 0.83, FLOPs: 65 GFLOPs) and NBNet (PSNR: 38.5 dB, SSIM: 0.85, FLOPs: 22.5 GFLOPs). Because it maintains natural image grain while producing effective denoising, this lightweight design is ideal for real-world applications including surveillance, medical imaging, and smartphone photography. These results open the door for further advancements in the industry and demonstrate how deep learning may be used to overcome obstacles in low-light image improvement.

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

Low-Light Image Denoising Using Deep Neural Network Incorporating NAF Blocks and Subspace Attention Modules

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Huda Mariyam,
  • Sankalp Ichangi,
  • Vasuki Naik,
  • Nandan Chillal,
  • Shashank Hegde,
  • Uday Kulkarni

摘要

The need for high-quality images has increased in many industries in the digital age, but taking sharp pictures in low light is still quite difficult. To boost denoising performance while keeping fine details and colour fidelity, this research investigates the constraints of conventional low-light enhancement methods and existing deep learning models. We suggest a unique deep neural network architecture with subspace attention modules (SSAs) for better image quality and nonlinear activation free (NAF) blocks to overcome these problems. Our method is quite effective, obtaining a PSNR of 24 dB and SSIM of 0.6 with only 2.5 GFLOPs, in contrast to models such as NAFNet (PSNR: 40 dB, SSIM: 0.83, FLOPs: 65 GFLOPs) and NBNet (PSNR: 38.5 dB, SSIM: 0.85, FLOPs: 22.5 GFLOPs). Because it maintains natural image grain while producing effective denoising, this lightweight design is ideal for real-world applications including surveillance, medical imaging, and smartphone photography. These results open the door for further advancements in the industry and demonstrate how deep learning may be used to overcome obstacles in low-light image improvement.