Scholarlink

<p>In the task of infrared image denoishing, overlooking temperature as a physical prior makes it difficult to differentially process noise-sensitive regions, leading to residual thermal noise in high-temperature objects and over-smoothing of details in low-temperature backgrounds. To address this issue, this paper proposes a U-Net based, temperature-guided infrared image denoising network, termed TD-UNet. To enhance the network’s capability for differential modeling of noise and details in infrared images, a Multi-Scale Residual Block (MSRB) is introduced, which strengthens the modeling capacity for complex textures and global structures through densely connected dilated convolutions and a dual-attention mechanism. Furthermore, to enable the network to achieve physics-aware adaptive denoising by dynamically balancing noise suppression and detail preservation across different temperature regions, a Decoupled Temperature-Guided Fusion Block (DTGF) is constructed. Experimental results demonstrate that, compared to state-of-the-art methods, TD-UNet achieves an effective balance between noise suppression and detail preservation.</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

Temperature-Guided Multi-Scale Residual Denoising Network for Infrared Images

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Yuting Xu,
  • Haoran Li,
  • Yuyu Zhu,
  • Qiangming Cai,
  • Qiuyue Xu,
  • Yi Yu

摘要

In the task of infrared image denoishing, overlooking temperature as a physical prior makes it difficult to differentially process noise-sensitive regions, leading to residual thermal noise in high-temperature objects and over-smoothing of details in low-temperature backgrounds. To address this issue, this paper proposes a U-Net based, temperature-guided infrared image denoising network, termed TD-UNet. To enhance the network’s capability for differential modeling of noise and details in infrared images, a Multi-Scale Residual Block (MSRB) is introduced, which strengthens the modeling capacity for complex textures and global structures through densely connected dilated convolutions and a dual-attention mechanism. Furthermore, to enable the network to achieve physics-aware adaptive denoising by dynamically balancing noise suppression and detail preservation across different temperature regions, a Decoupled Temperature-Guided Fusion Block (DTGF) is constructed. Experimental results demonstrate that, compared to state-of-the-art methods, TD-UNet achieves an effective balance between noise suppression and detail preservation.