Scholarlink

<p>Passive non-line-of-sight (NLOS) imaging, which relies solely on ambient light sources and reflective surfaces, faces significant challenges due to limited information, leading to poor imaging quality. Recent advancements in diffusion models have demonstrated remarkable capabilities in handling complex imaging tasks, particularly under low-light and incomplete data conditions. Motivated by this, we introduce a diffusion-based model, MSPDiff, for NLOS imaging, leveraging polarization cues and long-wavelength infrared (LWIR) images. Polarization information provides additional cues about light sources and surface characteristics, enhancing the reconstruction process. MSPDiff employs a progressive training schedule, transitioning from lower to higher resolutions of polarized LWIR images, to guide the iterative gradient descent. Extensive experiments on a large-scale passive polarized NLOS dataset demonstrate the superiority of our approach, achieving a PSNR of 25.78 and SSIM of 0.92, outperforming state-of-the-art methods. Our work highlights the potential of diffusion models, guided by physical priors, in advancing NLOS imaging technology. The code is available at <a href="https://github.com/simatanke/MSPDiff">https://github.com/simatanke/MSPDiff</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

Enhanced passive non-line-of-sight imaging via multi-scale polarization-guided diffusion model

  • 期刊论文

  • 发表日期：

    2025年7月5日

查看全文
  • Shaohui Jin,
  • Peng Zheng,
  • Guangpeng Li,
  • Huimin Wang,
  • Manman Zhang,
  • Wenhao Zhang,
  • Hao Liu

摘要

Passive non-line-of-sight (NLOS) imaging, which relies solely on ambient light sources and reflective surfaces, faces significant challenges due to limited information, leading to poor imaging quality. Recent advancements in diffusion models have demonstrated remarkable capabilities in handling complex imaging tasks, particularly under low-light and incomplete data conditions. Motivated by this, we introduce a diffusion-based model, MSPDiff, for NLOS imaging, leveraging polarization cues and long-wavelength infrared (LWIR) images. Polarization information provides additional cues about light sources and surface characteristics, enhancing the reconstruction process. MSPDiff employs a progressive training schedule, transitioning from lower to higher resolutions of polarized LWIR images, to guide the iterative gradient descent. Extensive experiments on a large-scale passive polarized NLOS dataset demonstrate the superiority of our approach, achieving a PSNR of 25.78 and SSIM of 0.92, outperforming state-of-the-art methods. Our work highlights the potential of diffusion models, guided by physical priors, in advancing NLOS imaging technology. The code is available at https://github.com/simatanke/MSPDiff.