Scholarlink

<p>Low-light image enhancement is crucial for improving visibility and detail in dark images, which is essential for various computer vision tasks. Traditional methods often introduce artifacts, while deep learning approaches may lack illumination-aware linguistic guidance. This paper introduces NaLSuper, a natural language supervision-driven framework that leverages illumination-descriptive prompts for enhanced brightness and detail recovery. By integrating a Textual Guidance Conditioning Mechanism and an Information Fusion Attention module, NaLSuper achieves deep, multi-scale text-image interactions that continuously condition the enhancement process with illumination-related linguistic priors. Experiments on LOL, SID, and SDSD datasets demonstrate state-of-the-art performance. On the LOL dataset, NaLSuper surpasses previous methods by 0.36dB in Peak Signal-to-Noise Ratio and 0.013 in Structural Similarity Index Measure, balancing objective performance with subjective naturalness. Our approach effectively enhances low-light images, providing a robust solution for real-world applications. Code is available at <a href="https://github.com/jiahuitang1/NAL">https://github.com/jiahuitang1/NAL</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

Enhancing low-light images via text-guided multimodal learning

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Jiahui Tang,
  • Kaihua Zhou,
  • Zhijian Luo,
  • Yueen Hou

摘要

Low-light image enhancement is crucial for improving visibility and detail in dark images, which is essential for various computer vision tasks. Traditional methods often introduce artifacts, while deep learning approaches may lack illumination-aware linguistic guidance. This paper introduces NaLSuper, a natural language supervision-driven framework that leverages illumination-descriptive prompts for enhanced brightness and detail recovery. By integrating a Textual Guidance Conditioning Mechanism and an Information Fusion Attention module, NaLSuper achieves deep, multi-scale text-image interactions that continuously condition the enhancement process with illumination-related linguistic priors. Experiments on LOL, SID, and SDSD datasets demonstrate state-of-the-art performance. On the LOL dataset, NaLSuper surpasses previous methods by 0.36dB in Peak Signal-to-Noise Ratio and 0.013 in Structural Similarity Index Measure, balancing objective performance with subjective naturalness. Our approach effectively enhances low-light images, providing a robust solution for real-world applications. Code is available at https://github.com/jiahuitang1/NAL.