Scholarlink

<p>Underwater foreground object segmentation is crucial for marine exploration. This paper proposes an unsupervised segmentation method for unlabeled underwater images based on visual saliency. By integrating a self-supervised vision transformer with the Jenks natural breaks algorithm (JNBA), we introduce an adaptive region-weighted (ARW) clustering framework to address uneven foreground distributions. Our method achieves significant improvements in segmentation accuracy, with IoU enhancements of 9.9%, 5.7%, and 10.8% on the UFO-120 dataset compared to methods without ARW. Furthermore, our approach demonstrates superior performance on the SUIM, USOD10K, and UFO-120 datasets, outperforming state-of-the-art methods by 6.3%, 0.6%, and 1.7%, respectively. The source code for our method is publicly available at: <a href="https://github.com/LLL-YUE/ARW-JNBA.">https://github.com/LLL-YUE/ARW-JNBA.</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

Adaptive region-weighted clustering with Jenks algorithm for underwater object segmentation

  • 期刊论文

  • 发表日期：

    2026年1月30日

查看全文
  • Yue Liu,
  • Yun Xu,
  • Junpeng Shang,
  • Dongfang Ma

摘要

Underwater foreground object segmentation is crucial for marine exploration. This paper proposes an unsupervised segmentation method for unlabeled underwater images based on visual saliency. By integrating a self-supervised vision transformer with the Jenks natural breaks algorithm (JNBA), we introduce an adaptive region-weighted (ARW) clustering framework to address uneven foreground distributions. Our method achieves significant improvements in segmentation accuracy, with IoU enhancements of 9.9%, 5.7%, and 10.8% on the UFO-120 dataset compared to methods without ARW. Furthermore, our approach demonstrates superior performance on the SUIM, USOD10K, and UFO-120 datasets, outperforming state-of-the-art methods by 6.3%, 0.6%, and 1.7%, respectively. The source code for our method is publicly available at: https://github.com/LLL-YUE/ARW-JNBA.