Scholarlink

<p>Thermal infrared image colorization aims to synthesize realistic visible-domain representations from single-band thermal inputs, a task inherently ill-posed due to the indirect relationship between thermal radiation and visual reflectance. To address this challenge, we propose a Top-Down Feature-Guided Generative Adversarial Network (TDFG-GAN) that introduces a hierarchical top-down generation strategy. Unlike conventional bottom-up frameworks that rely solely on local evidence, the proposed model first establishes a global semantic hypothesis and then refines fine details under this high-level guidance. The generator comprises multiple Top-Down Feature-Guided Modules (TDFGMs), each jointly performing global semantic reasoning through VMamba-based sequence modeling, structural perception via channel–spatial large-kernel convolution, and local consistency enforcement through feature affinity propagation. This integrated design enables the network to maintain semantic coherence while preserving geometric structure and texture continuity. Extensive experiments on the KAIST and IRVI datasets demonstrate that TDFG-GAN significantly outperforms existing state-of-the-art methods in both quantitative metrics and perceptual quality. The results confirm that top-down feature guidance effectively reduces ambiguity, stabilizes color mapping, and produces visually natural and semantically consistent colorization across diverse thermal scenes. Code is available at <a href="https://github.com/gy-dotit/TDFG-GAN">https://github.com/gy-dotit/TDFG-GAN</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

TDFG-GAN: Top-down-feature guided GAN for thermal infrared image colorization

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Yang Gao,
  • Qian Jiang,
  • Kai Yang,
  • Hongyue Huang,
  • Xuyang Wang,
  • Wei Zhou,
  • Xin Jin

摘要

Thermal infrared image colorization aims to synthesize realistic visible-domain representations from single-band thermal inputs, a task inherently ill-posed due to the indirect relationship between thermal radiation and visual reflectance. To address this challenge, we propose a Top-Down Feature-Guided Generative Adversarial Network (TDFG-GAN) that introduces a hierarchical top-down generation strategy. Unlike conventional bottom-up frameworks that rely solely on local evidence, the proposed model first establishes a global semantic hypothesis and then refines fine details under this high-level guidance. The generator comprises multiple Top-Down Feature-Guided Modules (TDFGMs), each jointly performing global semantic reasoning through VMamba-based sequence modeling, structural perception via channel–spatial large-kernel convolution, and local consistency enforcement through feature affinity propagation. This integrated design enables the network to maintain semantic coherence while preserving geometric structure and texture continuity. Extensive experiments on the KAIST and IRVI datasets demonstrate that TDFG-GAN significantly outperforms existing state-of-the-art methods in both quantitative metrics and perceptual quality. The results confirm that top-down feature guidance effectively reduces ambiguity, stabilizes color mapping, and produces visually natural and semantically consistent colorization across diverse thermal scenes. Code is available at https://github.com/gy-dotit/TDFG-GAN.