Scholarlink

<p>In complex environments or low-light conditions, human vision and traditional visible-light imaging techniques are often limited. To address these limitations, thermal imaging techniques are widely used for monitoring and visual enhancement, but thermal images often lack detailed semantic information and visual clarity. Current approaches primarily employ GANs or CNNs for thermal-visible translation, yet struggle with thermal-specific artifacts and identity preservation due to their inherent training instability. Our solution innovates through: (1) an encoder-decoder architecture with thermal-infrared feature perception and cross-modality feature injection, (2) hierarchical feature alignment for identity preservation, and (3) a balanced training strategy maintaining both pixel fidelity and biometric consistency. This establishes the first feature-enhanced encoder-decoder framework for reliable biometric-preserving thermal-to-visible translation. Extensive experiments demonstrate that our proposed LWIR2Vis-Net significantly outperforms state-of-the-art methods in both visual quality and identity preservation, while maintaining computational efficiency suitable for real-world applications.</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

LWIR2Vis-Net: feature injection-augmented encoder-decoder for efficient thermal-to-visible image translation

  • 期刊论文

  • 发表日期：

    2026年3月18日

查看全文
  • Hongzhen Quan,
  • Yuhang Dai,
  • Dan Bai,
  • Ye Lin

摘要

In complex environments or low-light conditions, human vision and traditional visible-light imaging techniques are often limited. To address these limitations, thermal imaging techniques are widely used for monitoring and visual enhancement, but thermal images often lack detailed semantic information and visual clarity. Current approaches primarily employ GANs or CNNs for thermal-visible translation, yet struggle with thermal-specific artifacts and identity preservation due to their inherent training instability. Our solution innovates through: (1) an encoder-decoder architecture with thermal-infrared feature perception and cross-modality feature injection, (2) hierarchical feature alignment for identity preservation, and (3) a balanced training strategy maintaining both pixel fidelity and biometric consistency. This establishes the first feature-enhanced encoder-decoder framework for reliable biometric-preserving thermal-to-visible translation. Extensive experiments demonstrate that our proposed LWIR2Vis-Net significantly outperforms state-of-the-art methods in both visual quality and identity preservation, while maintaining computational efficiency suitable for real-world applications.