Scholarlink

<p>In real-world scenarios, insufficient exposure in lighting conditions may adversely affect the results of computer vision tasks. Low-light image enhancement techniques aim to improve the visibility of images captured in low-light environments. However, the enhanced images often exhibit issues such as uneven brightness and color deviation. To better recover low-light images, we propose an effective low-light image enhancement model called the Illumination-Guided Dual-Branch Transformer (IGDBT). IGDBT consists of two main branches: a local curve estimation branch and a global image correction branch. To address the problem of uneven illumination in the restored images, the local curve estimation branch utilizes a deformed U-Shape network architecture composed of an Illumination-guided Attention Block (IAB) to generate parameter mappings for image adjustment. These parameter mappings are then substituted into our designed cubic pixel-wise curve formula to iteratively optimize the image. To tackle the color deviation issues that arise during image restoration, we add a global image correction branch in the model. In this branch, a color matrix and gamma parameters are generated through a cross-attention module, allowing for global color and gamma correction of the locally adjusted images to achieve optimal enhancement results. We conducted extensive evaluations of IGDBT on LOL-v1, LOL-v2-real_captured, LOL-v2-synthetic, and the SMID dataset. Extensive quantitative and qualitative experiments demonstrate that IGDBT outperforms existing methods on public datasets, validating its superiority and practicality in low-light image enhancement.</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

IGDBT: Illumination-guide dual branch transformer for low-light image enhancement

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Yunuo Xie,
  • Zhijia Zhang,
  • Xin Huang,
  • Xingqi Na

摘要

In real-world scenarios, insufficient exposure in lighting conditions may adversely affect the results of computer vision tasks. Low-light image enhancement techniques aim to improve the visibility of images captured in low-light environments. However, the enhanced images often exhibit issues such as uneven brightness and color deviation. To better recover low-light images, we propose an effective low-light image enhancement model called the Illumination-Guided Dual-Branch Transformer (IGDBT). IGDBT consists of two main branches: a local curve estimation branch and a global image correction branch. To address the problem of uneven illumination in the restored images, the local curve estimation branch utilizes a deformed U-Shape network architecture composed of an Illumination-guided Attention Block (IAB) to generate parameter mappings for image adjustment. These parameter mappings are then substituted into our designed cubic pixel-wise curve formula to iteratively optimize the image. To tackle the color deviation issues that arise during image restoration, we add a global image correction branch in the model. In this branch, a color matrix and gamma parameters are generated through a cross-attention module, allowing for global color and gamma correction of the locally adjusted images to achieve optimal enhancement results. We conducted extensive evaluations of IGDBT on LOL-v1, LOL-v2-real_captured, LOL-v2-synthetic, and the SMID dataset. Extensive quantitative and qualitative experiments demonstrate that IGDBT outperforms existing methods on public datasets, validating its superiority and practicality in low-light image enhancement.