Scholarlink

<p>Color is a crucial perceptual cue in the human visual system, and the same holds true for machine vision systems. However, existing image fusion algorithms typically inherit the color channels directly from the source visible light images. This often leads to severe color distortion and diminishes the saliency of perceived targets in the fused images. To overcome this limitation, we propose a novel <b>C</b>olor-information-driven dual-task parallel network designed for simultaneous visible (<b>R</b>GB) and thermal infrared (<b>T</b>) <b>F</b>usion and <b>S</b>aliency object detection, which we name CRTFS. Drawing on the science of color perception, we introduce a color loss in a coarse-to-fine manner, which enhances the color fidelity of the fusion network. Furthermore, although multitask learning often compromises the performance of individual tasks, we demonstrate the contrary, that is achieved by image fusion benefiting from the saliency and contour loss of the salient object detection task, while salient object detection benefiting from the color loss proposed for the image fusion branch. We show that the mutual re-enforcement of the tasks via the complementarity of their learning objectives, instigated through an innovative task-interaction mechanism, impacts positively on the performance of both. Our code is available at <a href="https://github.com/Yukarizz/CRTFS">https://github.com/Yukarizz/CRTFS</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

A Color Information Driven Collaborative Training of Dual Task Parallel Network for Visible and Thermal Infrared Image Fusion and Saliency Object Detection

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Zeyang Zhang,
  • Hui Li,
  • Tianyang Xu,
  • Xiaojun Wu,
  • Muhammad Awais,
  • Josef Kittler

摘要

Color is a crucial perceptual cue in the human visual system, and the same holds true for machine vision systems. However, existing image fusion algorithms typically inherit the color channels directly from the source visible light images. This often leads to severe color distortion and diminishes the saliency of perceived targets in the fused images. To overcome this limitation, we propose a novel Color-information-driven dual-task parallel network designed for simultaneous visible (RGB) and thermal infrared (T) Fusion and Saliency object detection, which we name CRTFS. Drawing on the science of color perception, we introduce a color loss in a coarse-to-fine manner, which enhances the color fidelity of the fusion network. Furthermore, although multitask learning often compromises the performance of individual tasks, we demonstrate the contrary, that is achieved by image fusion benefiting from the saliency and contour loss of the salient object detection task, while salient object detection benefiting from the color loss proposed for the image fusion branch. We show that the mutual re-enforcement of the tasks via the complementarity of their learning objectives, instigated through an innovative task-interaction mechanism, impacts positively on the performance of both. Our code is available at https://github.com/Yukarizz/CRTFS.